Now’s the time to get good at it.

The whole exome should compound harder than a 648-gene panel. The market isn't paying for it yet.

Share

In essence, Caris and Tempus AI do the same thing: take a tumour sample, sequence its DNA and RNA, subtract the background noise, match the profile against a large outcomes-linked database, run AI to pick the therapy or trial. It is the identical Ontology I described in my latest Hims piece: drug plus biomarkers implies updated biomarkers, which approximates a clinical delta. Both feed every result back as proprietary data that improves the next prediction.

Caris and Tempus AI differ fundamentally in terms of how much of your genome they read.

Caris reads the whole exome, every protein-coding gene, hypothesis-free. Tempus reads a curated 648-gene panel. On RNA they converge: both run whole-transcriptome sequencing, because fusions and splicing cannot be captured by a fixed DNA list. So the disagreement is specifically about DNA, bringing forth the following tradeoff: depth per patient versus breadth across patients and modalities.

Caris has less scale. Tempus has more. On raw cumulative volume Tempus is ahead, not behind: ~1.5M patients with sequenced data versus Caris's ~1.07M profiled cases. On revenue Tempus is roughly 1.6x larger, $348M versus $216M in Q1 2026. The one axis where Caris leads is slope, 79% growth against 36%.

But Caris is already printing cash. This is the real separation. Positive adjusted EBITDA and positive free cash flow, multiple quarters running, while Tempus is still climbing to breakeven. The smaller company is self-funding; the larger one is not yet.

Caris does not win on scale on any axis except depth-per-patient. It wins on cash discipline and on the uniformity of its corpus. Everything that follows is a bet on whether those two edges out-compound Tempus's larger, broader, faster-growing book.

Full exome gives Caris pricing power

ASPs (average selling prices) are rising. Tissue ASP moved from ~$3,256 in Q2 2025 to ~$4,089 in Q3 to ~$4,300 with true-ups, on its way to a guided ~$4,200 for 2026. A large slice of 2025 was reimbursement catch-up: Q3 carried a $37.9M true-up (underlying gross margin 61% versus 68% reported), Q4 carried $33.6M of prior-year collections. Adjust for those and the base tissue ASP sits around $3,500 to $3,900. Still best-in-industry, and still climbing.

Management is explicit about the cause. The CFO, unprompted, credits the pricing edge to "the decision David Halbert made to go the whole exome and whole transcriptome." The President takes the direct shot at the panel model: the new AI signatures exist because "small panels of hundreds of genes are not sufficient to offer these types of insights."

Full exome at this scale is unique to Caris, and the decision dates to 2018. Halbert: "In 2018, I made the decision to move Caris to Whole Exome and Whole Transcriptome sequencing." Every profile since has been generated that way, which is why the corpus is uniformly comprehensive rather than a patchwork of panel versions. Tempus offers a whole-exome assay (xE) but reserves it for research; its clinical default stays the 648-gene panel. So the comprehensive-at-clinical-scale position is genuinely Caris's alone.

This is a Singularity Scaler. Comprehensive sequencing means the back catalogue is re-mineable: when a new biomarker is discovered, you already measured it, so you re-value 1M+ existing cases at zero re-collection cost. A panel cannot do that; to capture tomorrow's target it must redesign and re-sequence. And because the data is comprehensive, the value extracted per unit of proprietary data rises faster as AI scales, there is simply more substrate for a better model to find. In digital systems a marginal advantage compounds into an exponential one. That delta, small now, is where the long-run separation from Tempus could come from.

However, the pricing power transmits through FDA approval, PLA codes and payer contracting, not depth alone. A whole-exome assay without those would not command this ASP. The good news for the bull case is that this makes the moat harder to copy, not easier, and PAMA pins the Medicare rate with no expected downward adjustment through 2029.

One core engine, many solutions, each feeding back

The platform is a flywheel. At the centre is the core engine: whole-exome and whole-transcriptome profiling on tissue and blood, which both is the original product (therapy selection) and generates the 1.07M-case database that powers everything else.

Around that engine, Caris launches new solutions that reuse the same data and the same channel: ChromoSeq for blood cancers, MI Clarity for breast-recurrence risk, Caris Detect for early detection, MRD for monitoring. Each is powered by the engine, and each new case feeds its data back in. Critically, the channel barely grows to support them, management ran roughly doubling volume on basically flat sales headcount for years, and new modalities ride the same oncologist relationships.

Subscribe now

The 2024 inflection

Revenue and operating cash flow both bend sharply upward between 2024 and 2025. Revenue: $258M, $306M, $412M, then $812M. Operating cash flow: roughly -$245M, -$276M, -$245M, then +$83M, and +$147M on an LTM basis. A company that burned a quarter-billion a year for three years flipped to cash generation.

The proximate drivers are in the calls: MI Cancer Seek's FDA approval (January 2025) and the $8,455 CMS rate that followed retroactively to it, payer contracting past 225M covered lives, PLA codes, and lab efficiency. Turnaround, the time from sample received to result, was 8 days for tissue and 7 for blood by Q3 2025, quick for an assay that reads the whole exome and whole transcriptome. Caris never disclosed the earlier figure, so read that as a level, not a measured improvement.

But the deeper driver is that AI has become materially better at pulling signal out of biological data. The clearest signal is Caris Detect, the whole-genome early-detection test, built on the same data foundation. On the interim ACHIEVE-1 readout it posted an AUC of 0.90. AUC, the area under the ROC curve, is the chance the test scores a random cancer patient above a random healthy one, where 1.0 is perfect and 0.5 is a coin flip, so 0.90 is strong separation.

However, note that those interim figures softened on validation. Stage I/II sensitivity fell from 63.1% on the interim cohort to 60.3% once the blinded holdout was scored, the normal give-back when a model meets data it has not seen, and Caris has not restated the AUC on that blinded set, so quote 0.90 as interim. Caris also shipped five new AI-derived signatures on its tumour board and argued its profiling is "becoming more proprietary, not commoditised."

These are all visible manifestations of the same shift: models can now turn a comprehensive biological corpus into differentiated products. Reimbursement provided the mechanism to monetise that value, but AI is what made the corpus significantly more valuable in the first place.

This is the prize Hims, Tempus and Caris are chasing. Eventually everyone has a Biology Ontology running in the background 24/7. Hims builds it via D2C, with a relentless focus on patient outcomes. Caris and Tempus build it on traditional healthcare rails, and the only real difference between them is full exome versus not.

Caris and Tempus don’t talk about patient outcomes

Caris and Tempus AI earnings calls consist primarily of analysts and management trading reimbursement minutiae for hours, ASPs, true-ups, PAMA cycles, covered lives, gross margin to the basis point. What you do not find, in any meaningful sense, is a conversation about actually keeping patients well.

That is not an accident. It is the same core issue I raised with Tempus: incentives, inherited from the rails.

Caris and Tempus draw their data through the traditional system, and that system is paid when you are sick. They earn per test, per diagnosis, and exactly $0 if you never get sick. Hims runs on the opposite reflex, longevity as a service, where the company makes more money the longer you stay well. Same Ontology, opposite reflexes. One is paid when the system acts on illness; the other when the customer avoids it.

As the Ontology becomes maximally predictive, it stops describing disease and starts forecasting it years out. At that point prevention becomes the cheaper path, and systems eventually follow the cheaper path, though the reimbursement machinery resists, so this is a destination, not the present.

From there the tree splits, exactly as it did for Tempus. In the first branch, Caris builds its own D2C infrastructure and pivots toward prevention-as-a-service, at which point it stops being adjacent to Hims and becomes a competing Ontology fighting for the same patient. In the second, Caris does not build D2C, and becomes the depth engine distributed through D2C channels like Hims, the comprehensive-genome layer that breadth players license and pour through pipes they already own.

Caris's whole-exome corpus is arguably the better engine for that role precisely because it is comprehensive and re-mineable.

The bottom line is the one from the original Hims deep dive: the final application of a Biology Ontology is preventing sickness, and the monopsony economics of traditional health insurance are not built to reward that. Caris has built, on current evidence, the deepest per-patient corpus in oncology and the only self-funding model among the comparables. What it has not built, and on these calls does not yet seem oriented toward, is a rail whose incentives point at the thing the Ontology is ultimately for.

Pharma investments as an additional signal

There is one number that cuts through everything else, and it runs against the depth thesis. A molecular ontology is only worth what someone pays to use it, and the buyer who matters right now is pharma. On that test, the deeper corpus is not the more valuable one. Tempus did roughly $316M in data and applications revenue in 2025, about a quarter of its $1.26B, growing 31% with the data-licensing line up 38%, against ~126% net revenue retention and a contracted book above $1.1B.

Caris, by contrast, guides its entire pharma and research line to $75–85M for 2026, under 8% of revenue, after that line actually shrank in 2025. Tempus's data business alone is roughly four times Caris's whole non-clinical revenue, and management at Caris won't even split out a data-licensing number until pharma crosses 10–20% of mix, which it is nowhere near.

Caris reads ~23,000 genes per patient; Tempus reads 648. Caris has the deeper data by construction. Yet the most sophisticated buyers in the world, the pharma R&D teams who could build this themselves, are paying Tempus multiples more for its maps. Either depth is not what pharma is buying, breadth across patients and modalities, the multimodal record wired to outcomes, is, or Tempus is simply years ahead at packaging and selling what it has. Both readings point the same way commercially.

And it may reframe Tempus as the more interesting bet right at an inflection: a contracted, retention-heavy data business already monetising the corpus, trading around 4x sales, just as generative AI starts spinning that corpus into predictive diagnostic products. The corpus depth argument still favours Caris on a ten-year horizon. The corpus value realised today favours Tempus.

Whichever way you read it now, Caris versus Tempus is a great experiment to see whether depth or breadth makes the faster-improving Ontology. One of them is wrong, and the data will say which.

⚡ If you enjoyed the post, please feel free to share with friends, drop a like and leave me a comment.

You can also reach me at:

Twitter: @alc2022

LinkedIn: antoniolinaresc

Disclosure

These are opinions only of the individual author. The contents of this piece do not contain investment and/or medical advice and the information provided is for educational purposes only and no discussions constitute an offer to sell or the solicitation of an offer to buy any securities of any company or any drug or medical treatment. All content is purely subjective and you should do your own due diligence.

Antonio Linares makes no representation, warranty or undertaking, express or implied, as to the accuracy, reliability, completeness or reasonableness of the information contained in the piece. Any assumptions, opinions and estimates expressed in the piece constitute judgments of the author as of the date thereof and are subject to change without notice. Any projections contained in the Information are based on a number of assumptions as to market conditions and there can be no guarantee that any projected outcomes will be achieved. Antonio Linares does not accept any liability for any direct, consequential or other loss arising from reliance on the contents of this presentation. Antonio Linares is not acting as your financial, legal, accounting, tax, medical or other adviser or in any fiduciary capacity.

My view remains that Biology will deliver most of the returns in the stock market in the coming few decades.

The bull case for a therapy that targets a common pathway across nervous system injuries.

Share

The body is good at healing skin, bone and liver. It is famously bad at healing nerves. Sever a peripheral nerve and recovery is slow and incomplete. Damage the spinal cord or the brain and you typically get none at all. The deficit is permanent, the patient is often young, and the cost compounds across an entire lifetime of care.

The global nerve repair and regeneration market sits around $11.5B in 2025 and is forecast near $32B by 2034. But roughly two thirds of that is neurostimulation and neuromodulation devices. That is the tell. The current market is almost entirely about managing the deficit, because there has never been a drug that reverses it. No approved therapy regenerates a damaged central nerve. The market understates the opportunity precisely because the product category does not yet exist.

And this is not one market. Everywhere the nervous system is damaged, the body lays down the same scar and engages the same brake. That turns the addressable surface into a stack of indications rather than a single line.

Ozempic for the nervous system

The brake is not specific to one injury. It is a general feature of how the mature nervous system responds to damage. Release it in the cord and you treat paralysis. Release it after a stroke and you treat the brain. Release it in a peripheral nerve and you treat the limb. Release it slowly, repeatedly, in an ageing nervous system, and you are preventing all of the above.

GLP-1 drugs did not stay in diabetes. They became a metabolic platform because they hit a node that touches appetite, weight, heart, kidney and liver at once. NervGen's peptide hits a node that touches every place where nerves fail to repair. If it works, this is closer to an Ozempic for the nervous system than to a single-injury drug. That is the long shape of the bet.

NervGen on a first principles basis

The therapeutic target sits inside the neuron, on the inner face of the membrane. Most drugs cannot reach it: antibodies are too large, most peptides get stuck on the outside. NervGen solved this by borrowing a sequence from HIV. The virus evolved a short protein segment, called TAT, whose entire function is to ferry cargo across the cell membrane. Bolt that segment onto your peptide and the whole construct walks straight into the cell. Part of this drug is a piece of a virus repurposed as a delivery van.

After injury, scars fill with molecules called CSPGs (Chondroitin Sulfate Proteoglycan). A CSPG is a ligand: a molecule whose job is to bind something and deliver a message. What it binds is a receptor on the axon called PTPσ. When CSPG docks onto PTPσ, the message delivered is unambiguous: stop growing. The axon obeys, freezing in the scar and recovery fails.

NervGen's sequence is shaped like the internal part of PTPσ that the receptor uses to pass that message inward. Once inside the cell, the peptide jams that relay. The receptor still sits in the membrane. CSPG still docks onto it. But the message no longer goes through. The stop sign is deleted, and the axon's own growth programme, which was never lost, simply switches back on.

Biology runs on ligands and receptors. A signalling molecule binds a sensor and triggers a response. Hormones, neurotransmitters, immune signals, growth factors: all of it is ligand meeting receptor. Once you see biology this way, drugs stop being mysterious. It can be a fake ligand, a blocked receptor, or, as here, a jammed relay sitting behind the receptor. NervGen picked the third option, which is the hardest and the most elegant.

NVG-291 drug seems to trigger four distinct repair programmes, which are downstream of the above proteomic even. The human data is early and the numbers are small, but they point in the same direction the animal models did.

Subscribe now

Understanding the nervous system and longevity implications

To see why a brake-release could matter beyond injury, you have to see what the nervous system actually is. Build it up in layers.

1. Cell types. A handful of specialists. The neuron carries the signal. The oligodendrocyte (central) and Schwann cell (peripheral) wrap the neuron in myelin, the insulation that makes signals fast. The astrocyte supports the tissue and, after injury, builds the scar. The microglia is the resident immune cell.

2. The functional unit. These cells assemble into a circuit: an axon carrying current, insulated by myelin, connected at synapses, held up by glial support. This is the irreducible unit of nervous function. Every movement, sensation and thought you have ever had is circuits firing.

3. The system. Circuits bundle into tissue: the brain and spinal cord at the centre, the nerves running to every muscle and organ at the periphery.

Injury, ageing and disease all damage the same unit: the circuit. And they all trigger the same defensive response: the scar, the CSPG, the brake. That is the structural coincidence the whole platform thesis rests on.

The four programmes NervGen's peptide unlocks, sprouting, regeneration, remyelination and calmer inflammation, are exactly the four things that decline in a nervous system as it ages. The ageing brain accumulates inhibitory matrix, loses myelin, and drifts into chronic low-grade inflammation. In slow motion, that is the same picture as injury.

So picture the peptide not as a one-time repair but as a periodic intervention. Each cycle nudges the balance: a little more repair, a little less decay. Net repair greater than net decay, applied repeatedly, is the definition of rejuvenation. The injury market is the proof of concept, while the maintenance market is the prize.

About the moat

Everything above rests on a single peptide. The entire operation is one amino acid sequence doing two jobs. That is the source of the asymmetry and the source of the risk, and you do not get one without the other.

A 35-amino acid peptide is not a difficult object to synthesise. The defensibility is not the molecule, it is the intellectual property around using this specific sequence to hit this specific target, licensed from Case Western. A patent on a single sequence is powerful while it holds and a cliff when it expires. With one product and one patent family, NervGen is a concentrated bet on a single legal and biological fact staying true.

Deleting the PTPσ brake is not the only road to regeneration. Reason it out from first principles, axons fail to regrow for two independent reasons, and you need a lever for each:

1. The environment says stop (extrinsic brakes). The scar is full of inhibitors. CSPG hitting PTPσ is one. Myelin debris carries others, Nogo, MAG, OMgp, that hit a different receptor (NgR). These largely funnel into one downstream switch, RhoA/ROCK, which collapses the growth cone. NervGen sits on the CSPG/PTPσ branch. But you can theoretically attack that same branch other ways: digest the CSPG sugar directly (chondroitinase), block the myelin inhibitors (anti-Nogo antibodies), or jam the shared downstream switch (ROCK inhibitors). NervGen owns one lever, on one branch, of one of the two problems.

2. The neuron has lost the will to grow (intrinsic capacity). Mature neurons switch off their own growth machinery. Re-activating it, by removing PTEN to wake the mTOR growth pathway, or removing SOCS3, produces some of the most dramatic regeneration ever recorded in a lab. These are completely different levers, and mostly not yet druggable in humans.

Releasing extrinsic brakes alone produces only limited regeneration. The largest effects come from waking the neuron's intrinsic growth, or from combining levers. So NervGen's edge is not that its lever is the only one, or even the strongest. Its edge is that its lever is the most druggable: a systemically injected peptide that works in chronic injury, years after the event, where nearly every rival is a mouse genetic trick, an enzyme with delivery problems, or an antibody that already stalled in the clinic.

NervGen does not have the only path to regeneration. It has a uniquely deliverable path, protected by a patent, on a node that recurs across many indications. If the Phase 3 confirms the signal, that combination, deliverable plus protected plus recurring, is exactly what may turn this one-drug company into a platform. Apart from getting better at Biology, that’s a great reason to continue watching NervGen.

I keep saying peptides are the tip of the iceberg. Here is a clear instance. A 35-letter string, half of it borrowed from a virus, walks into a cell and deletes a biological stop sign that evolution spent millions of years installing. The body then repairs itself. That is not medicine as we inherited it but biology being programmed. Peptides are the first and crudest interface to that programming layer. The epigenome and the genome are next. NervGen is a single, legible proof that the layer is real and that it is editable.

The body is built in layers: metabolism, immunity, the nervous system, and the rest. Each layer is a system, and every system has a control surface. My working assumption is that each foundational layer has one core proteomic node: a single protein-level lever where the biology is most druggable and the therapeutic yield relative to toxicity is highest. Most of biology is noise around these nodes and the returns concentrate on them.

And when you find a node and learn to address it precisely, the economics stop being normal. GLP-1 is the proof. A peptide hit the core node of metabolism, and the result was not merely a good drug. It became the largest drug category in medicine and created on the order of a trillion dollars of market value. That is the template. I expect a trillion-dollar peptide to emerge for each foundational layer, one after another, as we learn to address each node with the precision GLP-1 brought to metabolism.

We’ve only just started building the map:

• The immune system. ImmunityBio is tentatively staking this claim. The node is IL-15, the cytokine that orchestrates the immune response, addressed through Anktiva. If it holds, that is the immune layer’s GLP-1 candidate: earlier and less certain, but on the board.

• The nervous system. NervGen is the interesting name. The node is the CSPG/PTPσ brake on regeneration, addressed through NVG-291. Earlier still, and far less proven than the other two, but on exactly the same map.

Each layer’s flagship climbs the same curve, from early to proven, and the value compounds toward trillion-dollar scale as the node de-risks. Metabolism is already at the top. The immune layer is on its way. The nervous system is at the start. Every other layer is still sitting at the origin, waiting for its node to be found.

NervGen, then, is not only a spinal cord bet. It is an early position on the nervous system's square of a map that I believe will define most of the returns in the market over the next few decades.

To get better at Biology, you can read my previous biotech deep dives:

Investment Ideas by Antonio

Immunity Bio: The Company That Could Save the World.

This content is not medical or financial advice and should not be interpreted as such. This content is for educational and entertainment purposes only…

Read more

8 months ago · 25 likes · 2 comments · Antonio Linares

Investment Ideas by Antonio

Nautilus Biotechnology: A Digital Twin of Your Proteome.

Iterative proteoform mapping is on course to become the epicenter of capitalism and human life, like (Palantir’s) digital twins were in 2022. The human body is just an electromagnetic LEGO puzzle, otherwise known as the proteome. Iterative mapping will soon allow us to decipher it and by simply airdropping our proteome's digital twin into AI, we will be…

Read more

3 months ago · 20 likes · 4 comments · Antonio Linares

Investment Ideas by Antonio

MiNK Therapeutics: One Cell. Every Disease. The $50M Company Hiding in Plain Sight.

Biology is code…

Read more

22 days ago · 33 likes · 3 comments · Antonio Linares

Investment Ideas by Antonio

Recursion Pharmaceuticals: The Drug Factory That Designs Itself

This deep dive follows the deployment of the Biology module of my course, in which I teach a first principles framework that enables you to understand biology-based value creation processes…

Read more

16 days ago · 17 likes · 2 comments · Antonio Linares

Investment Ideas by Antonio

Tempus AI: Palantir With a Lab Coat.

I bought Palantir at $7. Tempus is fundamentally the same trade, in biology: a data-and-AI moat the market keeps pricing like a lab…

Read more

10 days ago · 37 likes · 1 comment · Antonio Linares

⚡ If you enjoyed the post, please feel free to share with friends, drop a like and leave me a comment.

You can also reach me at:

Twitter: @alc2022

LinkedIn: antoniolinaresc

Disclosure

These are opinions only of the individual author. The contents of this piece do not contain investment and/or medical advice and the information provided is for educational purposes only and no discussions constitute an offer to sell or the solicitation of an offer to buy any securities of any company or any drug or medical treatment. All content is purely subjective and you should do your own due diligence.

Antonio Linares makes no representation, warranty or undertaking, express or implied, as to the accuracy, reliability, completeness or reasonableness of the information contained in the piece. Any assumptions, opinions and estimates expressed in the piece constitute judgments of the author as of the date thereof and are subject to change without notice. Any projections contained in the Information are based on a number of assumptions as to market conditions and there can be no guarantee that any projected outcomes will be achieved. Antonio Linares does not accept any liability for any direct, consequential or other loss arising from reliance on the contents of this presentation. Antonio Linares is not acting as your financial, legal, accounting, tax, medical or other adviser or in any fiduciary capacity.

Hims and Tempus are building the same machine from opposite ends of medicine: one from your phone, one from the hospital. The only thing keeping them apart is incentives.

Share

Strip away the surface and Hims and Tempus are doing the same thing: building an Ontology of human biology. Not a drug, not a test, not a subscription, but a continuously improving model of the world they operate in, assembled from proprietary data no one else holds.

Hims sources its Ontology from the D2C channel. Every subscriber is a node, and as Hims expands Labs, it harvests more biomarkers about each one over time. The longer you stay, the denser your longitudinal record becomes. That density is the raw material.

What the network is actually learning is a function. It observes what input, a drug, a peptide, a protocol, modifies which biomarkers, in which direction, and how that change maps to a clinical outcome. On a first principles basis, the entire thesis compresses into one line:

drug + biomarkers ⟹ updated biomarkers ≈ clinical delta

Run that loop across millions of subscribers and you no longer have a pharmacy. You have a model that knows, before you act, what a given intervention will do to a body like yours.

Tempus is building the identical Ontology from the opposite end of the system. Its data does not come from a subscriber typing into an app. It comes from the clinical core: scans, biopsies, genomic assays across DNA and RNA, pathology, treatment records, response and adverse event data, all connected.

The consequence is depth. Hospitals currently go deeper than any D2C channel can, so Tempus inherits that depth by default. Its richest seam is oncology, over 4.5 million sequenced samples and 400,000 plus deeply multimodal records, where every layer of a real patient's cancer trajectory is wired to every other.

And depth is what lets the model generalise. As I covered in the deep dive, once the Ontology is dense enough, AI begins predicting things that were previously impossible to compute. The immune profiling score (IPS) is the proof of concept. It took a team of computational scientists years to build, it goes underneath the standard biomarkers, and it predicts who actually benefits from immunotherapy. IPS is not the achievement. It is the first output of a machine that will print the next thousand insights like it.

Here is where the framework I have been building for years applies directly. Both companies run the same loop,

drug + biomarkers ⟹ updated biomarkers ≈ clinical delta

and both feed every result back as proprietary data that sharpens the next prediction. As the Ontology approaches exhaustiveness and AI scaling laws keep doubling the model beneath it, the system crosses a threshold. It stops describing what happened and starts being biologically predictive.

It becomes a machine that prints insight at the marginal cost of inference, a Singularity Scaler, where the thing produced decouples from the thing spent and the gap widens as compute scales.

The endpoint is identical for both: an Ontology that tells you what to do before you get sick. Prevention computed in advance, rather than disease treated in arrears.

But they arrive carrying opposite shapes. Tempus has depth, Hims has breadth. Tempus knows end customers the atomic level through the hospital. Hims knows a its end customers continuously through the phone. And because each is strong exactly where the other is weak, both have a mechanical incentive to move toward the other:

1. Tempus wants the D2C channel, wearables, direct to patient pipes. Not just more data, but a delivery path that bypasses its dependency on the hospital to action anything.

2. Hims wants deeper data, hospital grade depth it can pour through the D2C infrastructure it already owns, to deliver sharper insights to the subscriber it already has.

This is Ontology Velocity pulling two companies toward the same point from opposite directions.

Subscribe now

If you accept that both are building predictive Ontologies of human biology, then at the level that matters they are the same company. What separates them today is not technology, nor even the data type. It is incentives, and incentives are inherited from the rails each one runs on.

Tempus draws its data through the hospital, and the hospital does not make money if you don't show up. So even though Tempus itself profits by delivering more and better insight at a lower cost, those insights get actioned through an infrastructure whose economics improve the sicker you are. The Ontology points toward prevention. The system it plugs into is paid for disease.

Hims runs on the opposite incentive. Its model is longevity as a service, LaaS, and under LaaS the company makes more money the longer you stay well. The subscription is the alignment. Keep the customer healthy and optimising, and lifetime value compounds.

Same Ontology, opposite reflexes. One is paid when the system acts on illness. The other is paid when the customer avoids it.

Whatever happens next, the base case holds: Tempus is converting the hospital into a set of generative AI algorithms. The ECG read, the pathology call, the immunotherapy match all become model outputs rather than human ones. And that conversion is precisely the thing that can, over time, change the incentives above. Once the model can tell you years in advance who has the heart attack, prevention becomes the cheaper path, and systems eventually follow the cheaper path, though the reimbursement machinery that resists it bends slowly, so this is a destination, not the present.

From there the tree splits.

In the first branch, Tempus builds its own D2C infrastructure, wearables, direct to patient delivery, and pivots toward LaaS itself. In that world it stops being adjacent to Hims and becomes a competing Ontology: same shape, same incentive, fighting for the same patient.

In the second, Tempus does not build D2C. Then it becomes the engine distributed through D2C channels like Hims, the depth layer that companies with breadth license and pour through their own pipes. Lefkofsky has already told us which way the wind blows: customers increasingly don’t just want the data, they want to build their models on the platform.

Either branch resolves the same way at the consumer level, and this is what anchors the whole thesis: customers do not want to go to hospitals. D2C channels will keep climbing the acuity ladder, from primary care toward urgent care, and the onset of robotic surgeons collapses much of what currently forces a hospital visit at all. The depth that only the hospital can capture today migrates, over time, to the channel the patient actually prefers.

⚡ If you enjoyed the post, please feel free to share with friends, drop a like and leave me a comment.

You can also reach me at:

Twitter: @alc2022

LinkedIn: antoniolinaresc

Disclosure

These are opinions only of the individual author. The contents of this piece do not contain investment and/or medical advice and the information provided is for educational purposes only and no discussions constitute an offer to sell or the solicitation of an offer to buy any securities of any company or any drug or medical treatment. All content is purely subjective and you should do your own due diligence.

Antonio Linares makes no representation, warranty or undertaking, express or implied, as to the accuracy, reliability, completeness or reasonableness of the information contained in the piece. Any assumptions, opinions and estimates expressed in the piece constitute judgments of the author as of the date thereof and are subject to change without notice. Any projections contained in the Information are based on a number of assumptions as to market conditions and there can be no guarantee that any projected outcomes will be achieved. Antonio Linares does not accept any liability for any direct, consequential or other loss arising from reliance on the contents of this presentation. Antonio Linares is not acting as your financial, legal, accounting, tax, medical or other adviser or in any fiduciary capacity.

It built the data pipes into two thirds of American medicine. Now it is turning that data into an AI that gets better at keeping you alive every time compute scales.

Share

Ten years ago Eric Lefkofsky asked a single question: could you use artificial intelligence to unlock precision medicine. To answer it you need two things, and almost nobody has both.

1. You need vast amounts of proprietary data to train the models.

2. And you need a distribution system to put what the models learn back into the hands of the physicians who treat patients.

Tempus has built both, and it spent a decade doing it whilst everyone watched the wrong number. The same playbook Palantir ran on government data, Tempus is running on the hospital

This is a Singularity Scaler. The thing it produces decouples from the thing it spends, and the gap widens as AI improves.

Tempus generates proprietary data correlating biomarkers to outcomes, and it generates that data as a free byproduct of clinical work it is already paid to do. Each test feeds the dataset. The dataset trains the models. The models make the tests smarter, which wins more physicians, which runs more tests, which feeds the dataset.

Run that loop for ten years across two thirds of American medicine and you do not get a diagnostics company. You get the substrate for an ambient, superhuman AI doctor that sits in the background of the clinic and improves the more the world's compute improves, at no additional cost to the loop.

The data is the moat and it’s already built

Start with the magnitude, because the magnitude is the whole argument. Tempus sits on over 500 petabytes of multimodal health data. Five years ago that number was 50. It has gone up tenfold whilst the cost of holding it collapsed, which is the shape you want. That data spans more than 45 million patients and contains over 9 million medical images, one of the largest digitised pathology and radiology sets known to exist anywhere.

Beneath it sit over 4.5 million sequenced samples, and at the very bottom of the funnel, over 400,000 deeply multimodal records: patients for whom Tempus holds DNA, RNA, clinical history, imaging, treatment, response and adverse-event data, all connected. That bottom layer is the rarest asset in biology. It is the totality of what you need to ask why a real patient lived or died, and answer it.

Then there is the compute. The foundation model alone sits on a cluster of 1,008 H200s, run for roughly 90 days at full capacity to pretrain. Behind it Tempus runs a roughly equal cluster of GB200s, each chip around four times the H200, plus more compute beyond the two foundation-model clusters. By management's own account, Tempus holds compute capacity comparable to the entire pharmaceutical industry combined. You do not have to take that on faith. An analyst on the call put the field in numbers: Eli Lilly around 1,000 GPUs in the lead, Recursion around 500, Amgen and BioNTech lower still.

Subscribe now

As the data compounds, Tempus turns into a model that can generatively ingest any biomarker and tell you what it means for the patient in front of you. Not "this patient has EGFR," which is targeted medicine and which any sequencer can give you, but "this patient has EGFR, and based on millions of comparable trajectories, they sit in the quarter that barely responds, not the quarter that stays on the drug for five years." That second sentence is precision medicine, and it is worth orders of magnitude more, because it changes what the doctor does next.

That is where the operating leverage of the next decade comes from. A useful clinical insight like the immune profile score took a team of computational scientists the better part of three years to build and validate. The next phase replaces that team with the model.

Instead of one new insight every six to twelve months, generate one a week. Each one generated, validated and dropped into the report at the marginal cost of inference. The leverage going forward is a function of the model's ability to spin up bespoke algorithms on the fly, in a way no competitor can match, because no competitor has the proprietary data that makes the algorithms work. And this is the move that turns a great company into a category: Tempus becomes the platform everyone else builds their own models on.

Increasingly, people don't just want our data, they also want models. Almost every conversation we have now is a blend of license some data and use our capabilities to build models, build those models on the Lens platform.

-Tempus co-founder and CEO, Erik Lefkofsky during 2026 Investor Day

Tempus has a stranglehold on American clinical care

Tempus has the deepest distribution in American medicine. It is connected to more than 5,000 institutions out of roughly 8,000 hospitals in the country, and reaches around two thirds of all academic medical centres. The vast majority of American oncologists order its tests, and there are only about 14,000 oncologists in the United States, of whom only half order comprehensive genomic profiling today. The room left to grow is inside a market Tempus already dominates.

The moat is mechanical, and it is the part investors consistently underprice. Proprietary data is a function of superior tests, and superior tests are a function of the pipes that connect Tempus to the institution: the BAAs, the legal agreements, the IT integrations, the EHR connections that let data flow out and insights flow back in.

Laying those pipes takes years. It is slow, unglamorous work, and it is exactly why a better-funded competitor cannot simply appear. What makes it a platform rather than a project is that Tempus has proven it can reuse the pipes. It earned its position in oncology first, then ran the same play into hereditary testing, rare disease, cardiology and radiology.

The tests themselves evolved the same way, by addition. Tempus started in solid tumour profiling with its xT assay, now FDA approved across its full DNA portfolio. It earned the right to move into liquid biopsy, the xF assays that read tumour DNA shed into the blood for patients without tissue. It added RNA, the xR assay, which reads a layer of biology DNA alone misses. It acquired Ambry to take the gold standard in hereditary testing, and folded in Paige to bring digital pathology models into the stack.

And it does not run these tests in isolation. It connects every test to every other test, and all of them to outcome data. So the platform a physician orders from does not stay the same. It gets smarter over time with every test. That is Ontology Velocity: the rate at which the system expands what it is able to know. The clearest way to see the value is to look at what the data surfaces that nothing else can.

A DNA test alone is a single lens, and a single lens misses things by construction. Add RNA and you start seeing the fusions and splice variants that DNA cannot show you, the ones that turn a patient with “nothing to target” into a patient with an approved therapy. Add a liquid biopsy and you catch the tumour DNA shed into the blood that a tissue sample never detected.

Each modality is not redundant with the last. It is a new window onto the same patient, and what it surfaces is a patient who was previously invisible to treatment. The numbers below are not marginal improvements to a test. They are people who would otherwise have been told there was nothing more to do, now matched to a therapy, because Tempus read a layer that everyone else left on the table.

To illustrate further, consider IPS (immune profiling score): an algorith developed manually by the Tempus data science team. IPS predicts which patients will actually benefit from immunotherapy. The standard biomarkers, tumour mutational burden and PD-L1, are right often enough to use and wrong often enough to hurt people.

IPS goes underneath them. Now hold the most important fact about it in your head: IPS tooks years to build. The entire thesis of the next decade is that the foundation model produces insights like IPS generatively, at a cadence humans cannot match. IPS is not the achievement. It is the proof of concept for the machine that makes the next thousand.

Heart disease is the number one killer in the country. One of the most common diagnostic insights in all of medicine is the ECG, run a few hundred million times a year. The read technology in most machines is thirty, forty, fifty years old, built before AI existed, and it is effectively wrong about 3% of the time. It tells people they are fine when they are not, when they are in fact likely to have a heart attack or stroke within the year.

Three percent of a few hundred million is millions of people, every year. Tempus trained on millions of ECGs connected to what actually happened, and built algorithms that catch what the old machines miss. Same system, different organ.

You can tell the platform works, because pharma pays for it

The cleanest proof that the data is valuable is that the most sophisticated buyers in the world keep paying more for it. Tempus has signed over 2 billion dollars in data licensing deals. It works with 19 of the 20 largest pharmaceutical companies in America, and over 250 biotechs. The customer base has gone from 35 companies in 2020 to around 240 today. And the concentration is collapsing in the healthiest possible direction: in 2020, 85% of the data business came from the top five clients; today that figure is 59% and falling.

Net revenue retention is 126%. A business that started as a handful of whale contracts is turning into a broad, diversified, compounding base.

The reason the money flows is that Tempus answers a question worth more than any other in drug development: will this drug work in this patient, and if not, why not. The legacy data vendors describe what happened, like playing back the news. Tempus addresses why a patient is not responding to standard of care. That why is the difference between a 1 billion dollar failed trial and a 10 billion dollar franchise.

A Phase III go or no-go decision carries 200 to 500 million dollars of risk. Against numbers like that, a 20 million dollar data licence that derisks the decision is not a cost, it is the cheapest insurance a pharma executive will ever buy. So they spend more, and the revenue line just goes up.

The market does not understand what it is looking at

Tempus trades at roughly four times sales. For a company growing the top line around 25%, with a data business growing faster, generating real operating leverage and approaching positive free cash flow, that is not a valuation. It is a structural misunderstanding. Technology investors who think in AI do not understand diagnostics. Diagnostics investors steeped in next-generation sequencing do not understand data and AI. So the stock sits in the gap between two audiences, that I am proud to bridge.

The ecosystem is very hard to displace. It is wired into the plumbing of American healthcare in a way that took a decade and billions to build, and the rebuild cost for a competitor is the same decade nobody wants to spend. And it is even harder to beat Tempus at the layer above, the apps and algorithms, because generative AI has quietly changed what an app is worth. The value of an algorithm is now a function of the proprietary data it was trained on. Tempus owns the data. So the better the AI gets, the more valuable Tempus's apps become relative to everyone else's, automatically.

That is what gives me clear visibility into sustained operating leverage going forward: not a forecast, a structural property of the loop.

A price-to-sales of four, a moat that compounds rather than erodes, and a genuine Singularity Scaler shape make this one of the more compelling names in my bio/acc watchlist. I’m pending a deeper read on the reimbursement path for algorithmic diagnostics, which is cracking but not yet open.

Where Tempus sits in the bio/acc stack

Bio/acc is one thesis expressed at four layers of a single value chain. Biology is becoming code, each company owns proprietary data at a different layer of the stack, and each becomes a Singularity Scaler on the data it owns. Measure, build, deliver, distribute. Tempus is the delivery layer, and it is another instance of the market mispricing the thesis.

Tempus operates inside the traditional, insurance-reimbursed healthcare system, and the incentives there are not aligned with truly optimising health over a lifetime. That system is built to code, bill and reimburse discrete episodes of disease, not to keep a person well for forty years, and a company that lives on its rails inherits its reflexes. That puts Tempus, for now, at odds with the layer below it: Hims is built around the patient optimising their own health continuously, direct and outside the reimbursement maze, whereas Tempus is paid when the system acts.

Long term, it is possible that Tempus is precisely what reverts those incentives, because an AI that can tell you years in advance who will have the heart attack makes prevention the cheaper path, and the system eventually follows the cheaper path. But that is a destination, not the present. What is not in tension, and what holds regardless of how the incentive question resolves, is the data.

Because of its stronghold at the clinical level, the underlying model captures intelligence about real patients and real outcomes that no one operating outside the hospital can see. The incentives may be imperfect today but the data is valuable anyway. Unless D2C fundamentally disrupts hospital care, Tempus’ free cash flow per share is likely to rise fast in the coming five years.

Lefkofsky's roughly $1.7 billion stake, more than half his net worth, is genuine alignment, and he funds the long bets from the core rather than diluting. Dual-class voting does mean more rests on his judgement than on the board's, and Groupon is worth remembering for context: public near $13 billion, around $600 million today, with some $398 million paid to him in the rounds before that IPO. None of it disqualifying, just a reason to weigh the narration carefully.

⚡ If you enjoyed the post, please feel free to share with friends, drop a like and leave me a comment.

You can also reach me at:

Twitter: @alc2022

LinkedIn: antoniolinaresc

Disclosure

These are opinions only of the individual author. The contents of this piece do not contain investment and/or medical advice and the information provided is for educational purposes only and no discussions constitute an offer to sell or the solicitation of an offer to buy any securities of any company or any drug or medical treatment. All content is purely subjective and you should do your own due diligence.

Antonio Linares makes no representation, warranty or undertaking, express or implied, as to the accuracy, reliability, completeness or reasonableness of the information contained in the piece. Any assumptions, opinions and estimates expressed in the piece constitute judgments of the author as of the date thereof and are subject to change without notice. Any projections contained in the Information are based on a number of assumptions as to market conditions and there can be no guarantee that any projected outcomes will be achieved. Antonio Linares does not accept any liability for any direct, consequential or other loss arising from reliance on the contents of this presentation. Antonio Linares is not acting as your financial, legal, accounting, tax, medical or other adviser or in any fiduciary capacity.

It sits alongside my other biology deep dives, each one an application of the same lens to a different layer of the stack: Immunity Bio, building a platform to rejuvenate the immune system at marginal cost; Nautilus Biotechnology, building the digital twin of your proteome; and MiNK Therapeutics, turning a forgotten branch of the immune system into an off the shelf cell therapy.

Read together, they map the same thesis across the field: biology is becoming code, and the companies learning to read and write it fastest are the ones that compound.

They claim to have built a superior drug discovery engine. I think they are right. Here is the evidence and why Recursion becomes a Singularity Scaler if the evidence holds.

Share

Every biotech has a pipeline. What Recursion claims to have is something one level up: an engine that produces pipelines, an AI native system that reads biology, designs the molecule, and carries it into the clinic faster and cheaper than the industry, and does so again and again. They call it the Recursion OS, and the framing they want you to hold is that the molecules are merely outputs while the OS itself is the asset.

If that claim is true, then you are not valuing seven early stage programmes at all. You are valuing a learning machine, and learning machines do not grow linearly. Similar to Abcellera, but my impression is Recursion has a far more nuanced understanding of biology.

Which is why the only question that matters is whether the engine is real, rather than whether any single drug works. I believe it is, and the case for it breaks cleanly into four pieces of evidence, followed by what those four pieces add up to once you put them side by side.

1.0 A differentiated Pipeline

A me-too pipeline is really a pipeline of coincidences, where everyone arrives at the same crowded targets because everyone is reading the same literature. Recursion's pipeline does not look quite like that. Management sorts it into three buckets:

1. The first bucket is novel biological insight, meaning targets the platform surfaced that the field had not prioritised.

2. The second is emerging biology, validated enough to be interesting but never properly drugged.

3. The third is validated biology with unmet need, where the target is well known and the job is simply a better molecule.

Most of the Recursion’s seven programmes sit in buckets two and three. CDK7, LSD1, MALT1, MEK in general are all known targets, and that is not where the differentiation lives. The differentiation is the small number of programmes in bucket one, and chiefly one of them, RBM39, which is why that single asset ends up carrying so much of the argument in Section IV.

2.0 They make fewer molecules

To move a single molecule toward the clinic, the industry synthesises somewhere between 2,500 and 5,000 candidates, whereas Recursion synthesises about 330, and it does so in roughly 17 months against an industry average above 40. Ninety per cent fewer compounds at twice the speed is not a rounding error. It is a different cost structure for the same act of discovery.

The mechanism is simple to state: they predict more, so they make less. The biology models and chemistry models simulate which molecules are worth synthesising, so the wet lab only builds the ones the system already believes in. Failure happens in silico, where it is cheap, instead of in glassware, where it is slow and expensive. Recursion truly simulates first and then downloads into physical space.

I went looking for movement between the Q3 2025 call and Q1 2026, and the interesting finding is that the headline number barely moved. It sat at roughly 330 compounds and roughly 17 months across all three quarters.

That sounds like nothing, but it is closer to the opposite of nothing. A one off efficiency win is luck, whereas a number that holds steady across quarters, while new programmes are added underneath it, is a process. What changed quarter to quarter was not the average but the pile of per programme receipts sitting beneath it: PI3Kα at 242 compounds in 10 months, RBM39 at around 200 in 18, LSD1 at just over 400 in 20. The stability is the signal, because it tells you they are reporting a repeatable rate rather than a single highlight.

3.0 Some of these drugs do things old methods could not

Two things are happening in this pipeline that conventional discovery struggles to do:

1. One is reaching a target that has no handle to grab.

2. The other, quieter and in my view more underrated, is the chemistry engine tuning a molecule's biological properties just enough to remove the toxicity that killed the entire class before it.

RBM39 is a splicing factor, and in Recursion's own framing it plays a central role in splicing fidelity. It has no convenient pocket to plug, which is why classic inhibitors cannot touch it, so they did not try to inhibit it. They degraded it instead, using a molecular glue that recruits the cell's own disposal machinery and shreds the protein entirely. The point is selectivity by context: tumours already straining their splicing die from the disruption, while cells with margin to spare survive it.

Subscribe now

Splicing is the edit step between a gene and the protein it codes for. A freshly transcribed RNA arrives full of stretches that have to be cut out and stretches that have to be kept, and the splicing machinery decides where to cut and what to stitch back together.

That choice is anything but trivial, because by selecting which segments survive, splicing determines which version of a protein a single gene actually produces. RBM39 is one of the proteins that keeps those cuts accurate, so degrading it makes the machinery misfire across thousands of transcripts at once, and the tumours already running their splicing hardest are the ones that cannot survive the errors.

That is the modality story, but the platform story is the better one. RBM39 emerged from genome scale phenomic mapping as a functional analog of CDK12, a relationship the company says was not obvious from sequence homology or pathway analysis. CDK12 is a wanted target that has resisted drugging because it is too similar to CDK13, and hitting both is toxic, yet Recursion's degrader mimics the loss of CDK12 without touching CDK13.

In other words, the engine found a back door into a target the field had been stuck on for years.

Then it translated. The first clinical data came in Q1 2026 across sixteen patients, with, per the company, no dose-limiting toxicities to date and target engagement confirmed. It is early, small, and mid escalation, but the undruggable target is now a tolerated drug in humans, which makes it the single most important fact in the file.

Most of the pipeline is not about finding new targets. It is about rescuing good targets that failed on tolerability. This is where the chemistry AI earns its place, and it is the part the market underweights.

The pattern repeats across the portfolio. A target everyone agrees is valuable has one toxicity that caps the dose below the level where it works, and Recursion uses generative chemistry to redesign the molecule's properties, its half life, its permeability, its selectivity, just enough to widen the therapeutic window. The target stays the same while the molecule changes, and the window itself becomes the product.

4.0 Half a billion dollars says the platform works

Here is the piece of evidence that does not depend on any drug succeeding. Recursion has taken in over $500 million in cash from pharma partners, among them Sanofi, Roche and Genentech, and they are not buying finished medicines. They are buying maps, the proprietary atlases of biology the engine generates.

Consider what that means. The most sophisticated R&D buyers in the world, companies that could in principle build this themselves, are writing cheques for the engine's raw output. Roche paid a milestone for a whole genome map of microglia, the immune cell of the brain, something many people did not believe could be built at all, and Sanofi has accepted five lead packages on hard targets.

This is the clearest validation in the file precisely because it is market priced. It does not prove any Recursion drug will be approved, but it does prove the underlying platform produces something valuable enough that pharma pays cash for it today. And that correlates directly with the differentiated nature of the molecules, because the same maps Roche buys are the maps that surface targets like RBM39.

If pharma pays for the maps, the maps are real. If the maps are real, the molecules built on them are not coincidences.

5.0 A closed loop on physics, chemistry and biology

And this is where the framework I have been building for years applies directly. A business whose output decouples from its input, and that grows exponentially more powerful at near zero marginal cost as AI compounds, is what I call a Singularity Scaler. Recursion has the shape of one, because every map they sell funds the next map, and every experiment, success or failure, feeds back as proprietary data that makes the next model better.

The engine pays for its own improvement.

A closed loop that learns the rules of biology and turns them into better medicine per dollar. The wet lab generates data, the dry lab builds models, the models predict, the lab tests the predictions, and the results, right or wrong, retrain the models. Round and round, getting tighter with each turn.

The reason this matters more than any single model is that drug discovery is a chain of decisions rather than one decision, so improving a single step is not enough. The value comes from compounding better decisions across the whole chain, from target to molecule to patient, and that is the thing an end to end loop does that a point solution cannot.

This is precisely the thing I named Ontology Velocity. Every company has an ontology, a working model of the world it operates in, built from the data it collects. The faster that model improves, the faster the company delivers more value per dollar. In an AI native world, the company with the fastest improving ontology does not just win. It opens a gap that compounds until it becomes uncrossable.

Recursion's ontology is biology itself. Their loop is a machine for improving that ontology faster than anyone else can. That is the whole bet, drawn as a flywheel.

Operationally, the loop is now spawning its own workforce. Recursion is building internal scientific agents: AI systems that sit inside the process, read the data, propose hypotheses, choose experiments, and capture the reasoning trail so the next agent is smarter. They do not replace the scientists. They remove the toil so the scientists do more of the hard part.

This is the economies of scale answer for a TechBio company. You do not scale by hiring linearly, you scale by letting the agents absorb the repetitive cognition, and that is what bends the cost curve down while the output curve bends up. It is the Singularity Scaler mechanism again, this time expressed in headcount.

6.0 A printer for medicine

Follow the loop forward and you arrive somewhere strange. Today the engine designs molecules one programme at a time, but the thing actually being built is not a molecule designer. It is a model of biology.

If the ontology keeps improving, the unit of output stops being a drug and starts being a capability. Give the system a target, a patient and a context, and it designs the molecule, first for a class of patients, then eventually for the patient sitting in front of you. A system that does not search for medicine so much as compute it.

That endpoint is years away and may never fully arrive, since biology is nowhere near finished being mapped, and the company will be the first to tell you only about ten per cent of it is understood. But the direction is the point, because a model that designs and prints personalised medicine on demand is, by construction, almost impossible to displace. You would not be competing with a drug, you would be competing with an accumulated understanding of biology that took a decade and fifty petabytes to kickstart.

7.0 Seven molecules, on first principles

Biology is code and the primary question is how deep into that code you have to reach. The shallowest intervention works at the signalling layer, the dynamic messages a cell is sending right now. Reach deeper and you arrive at the epigenome, the instructions that decide which genes get read at all. Deeper still sits the genome itself, the source code. And when none of that can be repaired in place, the last resort is to replace the hardware altogether with cell therapy.

That is my ladder for the whole field: peptides, then the epigenome, then the genome, and cell therapy if all of that fails. Signal, then reading, then source, then replacement. The deeper the rung you are forced to touch, the harder the delivery becomes and the more radical the intervention, which is why a company would always rather solve a disease high on the ladder than low on it.

Here is where Recursion's seven molecules actually sit.

The map says something the company never quite states outright. Every molecule it is running today sits on the top two rungs, which makes it a small molecule house operating at the signalling and epigenome layers. It does not edit genes and it does not build cells, and that is not a criticism so much as the actual shape of the thing, because the deeper rungs are where the maps could one day lead rather than where the drugs are now. With the ladder in place, here is each molecule on first principles, and how each one actually works.

The signalling layer

REC-4881 (MEK1/2, FAP). FAP patients inherit a broken APC gene, the brake on Wnt signalling, and without the brake the colon carpets itself in precancerous polyps. You cannot easily restore a deleted gene, so REC-4881 clamps a node downstream, MEK, the dedicated gateway to ERK, and starves the polyps of the growth signal they lean on.

The edge over standard of care is notable, because there is no standard of care: surgery is the option. The Phase 2 showed a 43% median polyp reduction with durability off treatment, and the platform edge is that phenomics surfaced, without a prior hypothesis, that MEK inhibition would attenuate the disease at all.

REC-3565 (MALT1, B-cell malignancies). MALT1 sits in the survival signal that B-cell lymphomas run on, so it is a known and validated target, and the reason it stalled was tolerability rather than biology. The platform edge here is pure chemistry: a molecule designed away from the UGT1A1 liability, so that it can be combined with BTK inhibitors, which is where the real value in this setting sits.

REC-7735 (PI3Kα H1047R, solid tumours). PI3Kα is one of the most important oncogenic targets there is, yet existing drugs cause hyperglycemia in 65 to 85% of patients because they hit the healthy wild type enzyme as well as the mutant. Recursion used molecular dynamics to find a novel pocket and designed a molecule with over 100x selectivity for the mutant form, so the edge is the same powerful target without the metabolic toxicity that limits how long patients can stay on it.

REC-102 (ENPP1, hypophosphatasia). This is the odd one out, and an insightful first principles move. Patients cannot clear pyrophosphate, which is the molecule that blocks bone mineralisation. This means their bones get weaker over time.

Since you cannot easily fix the broken clearing enzyme, Recursion instead inhibits ENPP1, the enzyme that produces the pyrophosphate in the first place. It cuts the supply rather than trying to restore the disposal, and the edge is convenience: an oral molecule against today's burdensome injected enzyme replacement.

The expression layer

REC-617 (CDK7, solid tumours). CDK7 is the dual master switch: one arm licenses the cell cycle, while the other fires transcription of the super-enhancer genes that oncogenes depend on, so blocking it hits the tumour's growth engine from both sides at once.

The whole class has been capped by a narrow therapeutic window, and the platform edge is chemistry tuning permeability and efflux to widen that window, alongside using patient data to steer the drug toward ovarian cancer rather than the crowded breast setting.

REC-1245 (RBM39, solid tumours and lymphoma). This is the one covered at length in Section IV: a splicing factor with no pocket, reached by degradation rather than inhibition, and found via the CDK12 analogy. It is the expression layer in its purest form, because it does not change which genes exist but corrupts which version of the protein gets made, and the tumours addicted to high splicing throughput die first. It is the lead novel-target receipt, and its mechanism is the one already drawn in Exhibit 3.

REC-4539 (LSD1, small cell lung cancer and AML). This is the epigenome tuning. LSD1 is a histone demethylase, an eraser that strips activating marks to keep cancer cells locked in an immature, proliferating state, so inhibiting it pushes the cell to differentiate instead.

The biology is known and was killed repeatedly by thrombocytopenia, and the platform edge is a brand new chemical scaffold designed for reversibility and a shorter half life to dial out that platelet toxicity, plus brain penetrance for the patients who develop brain metastases.

8.0 A new CEO

In the third quarter of 2025, the founder stepped back. Chris Gibson, who built Recursion from zero and arguably created the TechBio category, moved to Chairman, and Najat Khan, who had joined in 2024 to run R&D and Commercial, took over as CEO effective the first of January 2026.

This matters because the two profiles are genuinely different, and the company needs the second one now. Gibson is the visionary founder, unapologetic, the person who convinces the world a new category exists, whereas Khan is the operator. She came from running R&D strategy and data science at J&J, and at Recursion she led the Exscientia combination, the largest TechBio deal of 2024 and the thing that handed the company its chemistry engine in the first place. The vision is built; the job now is translation, discipline and proof, and that is precisely her skill set.

Khan has a base salary of $680,000. On top of that, equity: an initial grant of $500,000 in restricted stock and $500,000 in options, then a much larger grant with an aggregate value of at least $7.5 million, split evenly between restricted stock and options, all of it vesting over four years.

So roughly 97% of her pay is equity and her cash salary is only about 3% of the total, which makes the alignment real and long dated. The options only pay if the stock rises, tying her directly to shareholder upside.

A large slice is time vesting restricted stock, which pays on tenure as much as on performance, and her direct ownership is still modest because she is a hired operator rather than a founder with a decade of stock. The net read is that she is well aligned on the upside and locked in for four years, though without true founder skin in the game, which is the correct profile for an execution chapter.

Her stated philosophy is also exactly the one I would want: build a proprietary data moat that translates into better medicines. That is Ontology Velocity in plain language. She is not promising a blockbuster, she is promising to make the model of biology improve faster than anyone else's and to spend every dollar as if she were an investor, and that is the right idea.

Since she took over:

1. Molecules are moving faster into humans.

2. Financial discipline is up.

Molecules are moving faster into humans. RBM39 went from preclinical to dosed with clean early safety, LSD1 dosed its first patient, and every clinical stage programme now has a readout inside the next twelve to eighteen months. On cost, the discipline is visible: operating expense down by roughly a third year on year, full year guidance held under $390 million, and the runway pushed out to early 2028 even as the cash balance fell, because partner milestones are now funding the burn.

That last point is the whole operating thesis in one line. She is spending the engine’s own revenue to buy the time the clinical readouts need, so that if the readouts confirm, the re-rate follows, and if they do not, the discipline means she gets more shots before the money runs out.

9.0 Conclusion

The four pieces of evidence hold up. The pipeline contains genuine novelty, the discovery is an order of magnitude leaner than the industry and the leanness is stable, which tells you it is a process rather than a fluke. Some of the molecules do things old methods could not, both by reaching the undruggable and by tuning away the toxicity that buried whole classes. And pharma is has paid half a billion dollars in cash for the engine's raw output, which is the one validation that does not need a single drug to succeed.

Put together, that is the shape of a Singularity Scaler: a loop that learns biology faster than anyone, funds its own improvement, and bends its cost curve down while the value curve bends up. The operator now in the seat understands exactly this, and her incentives are locked to it for four years.

I don’t have a position and I will be watching Recursion closely going forward.

⚡ If you enjoyed the post, please feel free to share with friends, drop a like and leave me a comment.

You can also reach me at:

Twitter: @alc2022

LinkedIn: antoniolinaresc

Disclosure

These are opinions only of the individual author. The contents of this piece do not contain investment and/or medical advice and the information provided is for educational purposes only and no discussions constitute an offer to sell or the solicitation of an offer to buy any securities of any company or any drug or medical treatment. All content is purely subjective and you should do your own due diligence.

Antonio Linares makes no representation, warranty or undertaking, express or implied, as to the accuracy, reliability, completeness or reasonableness of the information contained in the piece. Any assumptions, opinions and estimates expressed in the piece constitute judgments of the author as of the date thereof and are subject to change without notice. Any projections contained in the Information are based on a number of assumptions as to market conditions and there can be no guarantee that any projected outcomes will be achieved. Antonio Linares does not accept any liability for any direct, consequential or other loss arising from reliance on the contents of this presentation. Antonio Linares is not acting as your financial, legal, accounting, tax, medical or other adviser or in any fiduciary capacity.

Peptides are the tip of the spear: they’re a short cut in a deeper value chain.

The body runs on biological code. You can intervene at different depths. Genome editing rewrites the source code. Epigenetic tuning changes which parts of that code get read. Peptides act further downstream, ready made signals that tell cells what to do.

Cell therapy gets the same end job done from the other direction, by inserting entirely new cells into the body. MiNK’s iNKTs are a powerful example of how that can save and improve lives.

Enjoy.

Share

How a single immune cell kills cancer, heals lungs, and reboots a collapsed immune system, and why one tiny company is the only one left building it.

iNKTs (invariant natural killer T cells) are a high performance branch of the immune system that immunotherapy has historically left for dead. The only public company in the US moving iNKT towards the clinic is MiNK Therapeutics. iNKTs outperform natural killer and T cell based therapies in the following way:

1. They detect the bad guys through lipid sensing rather than the MHC pathway tumors learn to switch off, which means it’s far harder for tumors, viruses and bacteria to trick them into oblivion.

2. They reboot the immune system. A single activated iNKT licenses dendritic cells, wakes up natural killer cells, and primes conventional T cells, igniting three arms of immunity at once from one signal.

3. They natively traffic to and persist in tissues where other cell therapies fail. For example liver, lung, and the dense stroma of solid tumors (the hard, walled-off terrain conventional T cells can’t reach).

4. They present a low toxicity profile, because they can both kill the bad guys and modulate the immune system towards an anti-inflammatory state.

5. They are thus, on a first principles basis, a breakthrough in immunotherapy.

The primary reason iNKTs have been left behind in immunology is they weren't easy to manufacture. Low, donor dependent yields gated the field for two decades. IL-15 superagonist biology cracked that open, and MiNK has built its own proprietary process to make and store iNKTs at scale. MiNK then takes it a step further: it pairs its iNKTs with ImmunityBio's IL-15 superagonist Anktiva in the clinic, stacking two clinically validated immune mechanisms into one combination with early results in critically ill patients.

0.0 iNKTs in the immune family tree

The immune system splits into two branches that evolved for different jobs. The innate branch is fast, blunt, and pre-programmed: it reacts to broad danger signals within minutes, kills indiscriminately, and never learns. The adaptive branch is slow, precise, and trainable: it takes days to recognise a specific threat, but once it does, it remembers forever. Almost every cell is committed to one branch or the other. iNKTs are a rare exception that combine a T cell receptor (adaptive hardware) but it fires instantly like an innate cell.

They sit at the junction between innate and adaptive immune cells.

1.0 iNKTs see what others can’t

The usefulness comes from one structural fact: they recognise threats by reading lipids presented on a molecule called CD1d, rather than reading protein fragments (peptides) presented on MHC (Major Histocompatibility Complex), the display system cells use to show what is inside them to immune cells.

Conventional T cells need MHC to work, and the favourite escape trick of tumours and chronic viruses is to switch MHC off so they become invisible. CD1d does not switch off the same way, so iNKTs keep seeing the threat after conventional immunity has gone blind. And because CD1d is the same in everyone, an iNKT from one donor works in any patient. No matching, no rejection, no graft versus host. That is the entire basis for off the shelf cell therapy with genuine scalability.

Sometimes a viral peptide closely resembles one of the body's own peptides. T cells trained to attack the virus then cross react against healthy tissue that carries the lookalike peptide, and the immune system starts attacking the body itself. This is known as molecular mimicry, and it has been a particularly relevant phenomenon in the world since 2020.

iNKTs sidestep this entirely. Molecular mimicry is a quirk of peptide recognition and iNKTs do not read peptides. Reading lipids on a separate channel, they are not subject to the cross reactivity that turns the adaptive system against its own host.

2.0 iNKTs reboot the immune system

But seeing the threat is only half of it. When an iNKT fires, it does not just kill. It releases a burst of signalling molecules, chiefly a cytokine called IFN gamma, that wake up the rest of the immune system in sequence. Other immune cells will do this to various degrees, but iNKTs matter most in the patients other immunotherapies cannot help: the ones whose innate or adaptive branches have already collapsed. The iNKT fires from a third axis that is still working, so it can restart the cells that have gone dark.

The chain runs like this. The iNKT licenses dendritic cells, the scouts that carry threat information back to the wider system. Those activated dendritic cells then prime conventional T cells, handing the slow, precise, trainable branch a target it can learn and remember. At the same time, the iNKT transactivates NK cells, switching on the innate killers that mop up any tumour cell which tried to hide by shutting off its MHC. One cell pulls three triggers, and each trigger recruits a different arm of immunity.

The reason this matters is that it is self amplifying. A conventional T cell kills what it is pointed at and stops there. An iNKT starts a cascade: it activates cells that activate further cells, and the response builds rather than fades. It does not just add another killer to the fight, it restarts the coordination between the innate and adaptive branches that chronic disease and cancer spend years dismantling. That is the literal sense in which it is a reboot. The system was not destroyed, it was switched off, and the iNKT turns it back on.

3.0 iNKTs reach the places other therapies can't

Solid tumours are clumps of cancer cells with a fortress around them. A layer of support cells lays down a dense, fibrous matrix around the tumour, walling it off from the bloodstream and from the immune system. This is the stroma, and it is the single biggest reason cell therapy has worked in blood cancers but stalled in solid ones. A conventional T cell can be engineered to recognise the tumour perfectly and still never reach it, because it cannot get through the wall. The drug is correct and the delivery fails.

iNKTs are different because they were built for this terrain. The body's own iNKTs spend their lives patrolling exactly these tissues: the liver, the lungs, and the dense fibrous environments where the stroma forms. That is their native beat. So when an iNKT is infused, it does not have to be coaxed into the hard places, it traffics there by default.

At the molecular level, this comes down to the receptors they carry. In the Biology module of my course I have a class just for ligands and receptors, because they are the grammar of the entire body.

Almost nothing in biology happens by a cell deciding to act. It happens because a signalling molecule, a ligand, fits into a receptor on the cell surface like a key into a lock, and that fit triggers everything downstream. Where a cell goes, what it kills, when it calms down, whether it lives or dies: all of it is receptors reading ligands. Understand the lock and key and you stop memorising biology and start predicting it.

Trafficking (where a cell goes to) is a pristine example of that.

Subscribe now

Tissues like the liver and lung constantly release a homing signal, a chemokine, into their surroundings. iNKTs carry the matching receptor, so they follow the signal straight to the source. In the liver the signal is a molecule called CXCL16 and the receptor is CXCR6, and the pairing is so central that iNKTs literally crawl along the walls of liver blood vessels reading it, stopping the instant they detect a threat. Strip that receptor away and iNKTs vanish from the liver almost entirely. The lung uses a different but identical idea, its own signal drawing in iNKTs that carry its own matching receptor.

The key point is that this receptor does not just attract iNKTs, it retains them. Most cell therapies wash out of tissue quickly. iNKTs are held in place by the same signal that drew them in, which is why they persist where they are needed rather than drifting back into circulation. They arrive at the fortress natively, and they stay.

The table might suggest each receptor is locked to one organ, but one of them breaks that pattern in a way that matters enormously for cancer. CXCR3 does not home to a place, it homes to a state. Its signals, CXCL9, CXCL10 and CXCL11, are switched on by interferon, the alarm the body releases wherever it is under attack. So CXCR3 does not ask which organ it is in, it asks where the fighting is. Inflamed tissue lights up with these signals, and so do tumours, almost regardless of where they sit in the body. An iNKT carrying CXCR3 is therefore not a liver cell or a lung cell. It is a cell that follows danger to its source.

Hypothetically, this is the property that lets a single off the shelf iNKT pursue a tumour in the lung, the gut, the pancreas or the breast, because it is not navigating by anatomy, it is navigating by the distress signal that almost all solid tumours emit.

CXCR6, by contrast, stays narrow. It is the liver homing receptor, and it shows up against tumours only because tumours happen to switch on its specific ligand, CXCL16. The same key cut for the liver turns out to fit a few tumour locks too. Between them the two receptors sketch the opportunity: one that follows danger anywhere, and one that was built for a single organ but gets opportunistically reused. The first is the more powerful idea, and it is the reason an organ agnostic iNKT against solid tumours is not a fantasy but a mechanistic possibility.

So far this has been mechanism: what the receptors are, where they point, and why the biology predicts an iNKT should be able to reach trouble almost anywhere in the body. The fair question is whether any of it actually shows up in patients. It does, at least as a signal. What follows is a first principles reading of MiNK’s own reported results, not a medical claim or an endorsement of the therapy.

The point is narrow but striking: across very different diseases, the same unmodified off the shelf cell, agenT-797 (a native iNKT with no engineering), appears to do what the biology predicts. In cancer it skews pro inflammatory and cytotoxic, turning cold tumours hot. In damaged lungs it does the opposite, calming runaway inflammation while still clearing pathogens.

One cell, opposite outputs, no re engineering between settings.

Whether these early signals hold up under randomised conditions is the open question, and the randomised lung injury readout due in the second half of 2026 is the first real test. But as a demonstration that the dual mode iNKT mechanism behaves in real patients the way first principles say it should, the pattern is unusually coherent.

4.0 iNKTs both kill bad guys and create conditions for tissue repair

The fourth advantage is the strangest one, and it follows directly from the reboot. Most therapies do one thing. A chemotherapy kills. A steroid suppresses. An iNKT does the appropriate thing for the situation it finds itself in, and the situation can be opposite from one disease to the next.

This dual capacity, cytotoxic when the problem is a hidden enemy and restorative when the problem is friendly fire, is why a single off the shelf cell can credibly be pointed at diseases as different as refractory gastric cancer and acute respiratory failure. One mechanism, two opposite jobs, selected by the tissue rather than by the engineer.

In a tumour, the job is to kill. The iNKT reads the cancerous tissue, fires in cytotoxic mode, and drives the Th1 interferon gamma response that turns a cold tumour hot. This is the mode behind MiNK’s oncology signals: in heavily pretreated, checkpoint refractory patients whose expected survival was measured in months, the company reports median survival beyond 23 months and several patients alive years after a single course, with complete remissions in cancers like metastatic testicular disease.

The cell is doing what a killer should do.

In a damaged lung, the job is the reverse. Here the threat is not a tumour hiding from the immune system, it is an immune system that has gone into overdrive and is destroying the tissue it is supposed to protect. The same unmodified iNKT, agenT-797, reads this environment and shifts into a restorative mode, releasing anti inflammatory signals like IL-4 and IL-13 rather than the cytotoxic burst it produces in cancer. It dampens the runaway inflammation instead of adding to it. This is not the cell being reprogrammed between diseases. It is the same cell reading two different rooms and responding correctly to each.

Crucially, calming the fire is only part of it. The iNKT also clears the pathogens that damaged tissue cannot fight off on its own. In MiNK's severe lung injury work, the cells cleared fungal and multidrug resistant gram negative infections and reduced the secondary infections that are often the actual cause of death in the ICU. This matters because the usual tradeoff in critical care is brutal: suppress the inflammation and you cripple the patient's ability to fight infection, but leave the inflammation and the tissue keeps burning.

The iNKT does not face that tradeoff, because it can dampen harmful inflammation while preserving, even enhancing, pathogen clearance. That is the combination conventional anti inflammatory approaches, including mesenchymal stromal cell therapies, have repeatedly failed to deliver.

The iNKT does not rebuild tissue the way a stem cell would. What it does is remove the two things preventing the tissue from healing itself: the runaway inflammation and the unchecked infection. Once both are gone, the body's own repair machinery can resume. In MiNK's ARDS patients this showed up as restored endothelial function and improved oxygenation, with patients coming off the most extreme form of life support. The iNKT did not regenerate the lung. It created the conditions under which the lung could recover.

5.0 The pipeline and H2 2026 catalyst

MiNK's pipeline is best understood as one native cell plus two engineered upgrades, all built on the same chassis. The lead asset, agenT-797, is the unmodified iNKT we have been describing this whole time. It is the furthest along, and it is being pointed at strikingly different diseases precisely because the cell reads its environment rather than being engineered for one target. It is in clinical trials across solid tumours, graft versus host disease, and critical pulmonary failure, all with the same off the shelf product.

The near term catalyst sits here. MiNK has moved agenT-797 into a randomized Phase 2 trial in acute lung injury and ARDS, designed on a seamless Phase 2/3 track, with regulatory interactions underway and preliminary data expected in the second half of 2026. That readout is the single most important date on the calendar for this company, because it is the first time the platform faces a randomised, controlled test rather than the single arm and emergency use settings that produced the early signals.

The two engineered programs add targeting to the native cell without touching its core biology. MiNK-215 bolts a FAP-CAR onto the iNKT to attack the cancer associated fibroblasts that build the stromal wall around solid tumours, the fortress described earlier. The PRAME-TCR program adds a peptide specific receptor aimed at pediatric cancers, advanced through the C-Further consortium on non-dilutive funding. Both are earlier than agenT-797, but both inherit the trafficking, the low toxicity, and the reboot for free, which is why MiNK can describe them as one-edit cells rather than the three-to-five-edit cells that conventional allogeneic CAR-T platforms require.

5.1 MiNK-215: the FAP-CAR

Start from the problem. A solid tumour is not just cancer cells, it is a fortress, and the wall is built by cancer associated fibroblasts that lay down dense fibrous tissue, secrete signals that suppress the immune system, and physically block immune cells from reaching the tumour inside. Most cell therapies aim at the tumour cells and never get through the wall.

MiNK-215 aims at the wall. The single added part is a CAR, a synthetic receptor, tuned to FAP, a protein found almost exclusively on those fibroblasts and almost nowhere on healthy adult tissue. That is the entire engineered change: one gene encoding one receptor that says “kill the cells building the wall.” Everything else is the native iNKT it was built from. So the cell still traffics to the tumour on its own, still senses lipids, still reboots the surrounding immune response, and now additionally tears down the stroma that was keeping every other therapy out.

The gain of function is access. You are not making a better killer, you are making a cell that can reach the fight, demolish the fortifications, and let the immune cascade it already triggers finish the job from the inside.

5.2 PRAME-TCR: the engineered T cell receptor

Here the problem is precision in a setting where precision matters most. PRAME is a tumour antigen, a protein that appears on cancer cells but is largely absent from healthy tissue, and it shows up across several pediatric cancers.

The native iNKT recognises threats through lipids, which is broad and powerful but not aimed at any one specific protein. The added part is an engineered T cell receptor that recognises a PRAME peptide, giving the cell a second, sharper way to see: keep the native lipid sensing, and bolt on a sniper sight for one defined cancer marker. The gain of function is targeted specificity layered onto innate breadth. The cell still does everything a native iNKT does, the trafficking, the low toxicity, the immune reboot, and now also locks onto PRAME bearing tumour cells directly.

The reason this matters especially for children is practical as much as biological: an off the shelf cell needs no lengthy individualised manufacturing and no harsh pre conditioning, so a child with an aggressive cancer can be treated fast, with a cell that combines a precise target and a favourable safety profile, exactly where time and tolerability are most precious.

5.3 iNKT is a platform in the making

Step back and the individual programs matter less than the pattern they form. A drug treats one disease. A platform is one underlying technology that keeps producing new products, each cheaper and faster than the last because it reuses the same core. A system that solves a growing volume of acute customer pains, in a way that’s increasingly harder to replicate while producing more and more cash (if it works).

iNKT behaves like a platform because the native cell already does the hard part. The trafficking, the low toxicity, the reboot, the ability to kill in one setting and calm in another all come built in, so a new program does not start from scratch. In cancer it is a cytotoxic agent, in lung injury a restorative one, in transplant a tolerising one, and where a sharper aim is needed you add a single module, a FAP-CAR for stroma, a PRAME-TCR for pediatric tumours, rather than rebuilding the cell.

One chassis, many diseases, low marginal cost for each new shot on goal. That is the difference between owning a product and owning a platform, and it is why the addressable opportunity is not one indication but a widening set of them. The qualifier is "in the making." None of this is proven until a randomized trial says so, and the first real test arrives with the lung injury data in the second half of 2026. But mechanically, the architecture of a platform is already in place.

6.0 Moat

The most striking thing about MiNK is not any single result. It is that almost nobody else is here. For an asset class with this much theoretical promise, the competitive field is nearly empty, and the reasons it is empty are also the reasons it is hard to enter.

The first barrier is manufacturing, and it is the one that gated the entire field

for two decades. iNKTs are rare cells, and historically the yield from a single donor was low and unreliable, which made an off the shelf product economically impossible. The unlock was IL-15 superagonist biology, which finally let iNKTs be expanded at scale, and MiNK built its own proprietary process around it. The company now reports getting roughly a billion cells per donor with stability measured in years, which is what turns a laboratory curiosity into a stockable, deployable drug. That process maturation took years and is not trivially copied.

The second barrier is that the rest of the field walked away or fell over. The companies that chased NKT based therapies have largely exited. Athenex's Kuur programs hit an FDA clinical hold after a patient death and wound down, which chilled capital across the whole space. And the one venture that looked like a serious rival, Appia Bio, partnered with Gilead's Kite on CAR-iNKTs in a deal once valued at up to $875 million, never made it. The Kite partnership expired in 2024 and Appia shut down in 2025, running out of cash just short of filing its first clinical application. In the public US market, MiNK is now effectively alone in moving native allogeneic iNKTs toward the clinic, not because the idea lacked backing, but because the execution is genuinely hard and the better funded attempts did not survive it.

The third barrier is the most durable. MiNK's real moat is not just that it makes iNKTs, it is the demonstrated context dependent biology: the same unmodified product behaving as a cytotoxic agent in cancer and a restorative one in lung injury, validated in patients. That is a property of the cell that a competitor cannot simply engineer around, and it sits on top of a decade of accumulated manufacturing and clinical know how. Capital chased CAR-T and CAR-NK for ten years because they fit the dominant mental model; iNKTs were dismissed as too rare, too hard to make, too unclear commercially.

MiNK spent that decade quietly solving exactly those problems, which is why the moat is structural rather than rhetorical. The barrier to entry is not a patent. It is the years of hard, failure prone work that everyone else either skipped or did not survive.

7.0 Financials

This is essentially a pre-revenue company with no debt, but lean operating expenses as they move the platform forward. Big financial and dilutive risk, but the platform shows great potential, with substantial upside if the H2 2026 readout is good. With a market cap of just ~$50M, this is a potentially vast asymmetry in the making.

⚡ If you enjoyed the post, please feel free to share with friends, drop a like and leave me a comment.

You can also reach me at:

Twitter: @alc2022

LinkedIn: antoniolinaresc

Disclosure

These are opinions only of the individual author. The contents of this piece do not contain investment and/or medical advice and the information provided is for educational purposes only and no discussions constitute an offer to sell or the solicitation of an offer to buy any securities of any company or any drug or medical treatment. All content is purely subjective and you should do your own due diligence.

Antonio Linares makes no representation, warranty or undertaking, express or implied, as to the accuracy, reliability, completeness or reasonableness of the information contained in the piece. Any assumptions, opinions and estimates expressed in the piece constitute judgments of the author as of the date thereof and are subject to change without notice. Any projections contained in the Information are based on a number of assumptions as to market conditions and there can be no guarantee that any projected outcomes will be achieved. Antonio Linares does not accept any liability for any direct, consequential or other loss arising from reliance on the contents of this presentation. Antonio Linares is not acting as your financial, legal, accounting, tax, medical or other adviser or in any fiduciary capacity.

Tesla is at an inflection point and is now positioned to print way more cash.

Share

Manufacturing efficiency has compounded hard since 2022. The chart below shows OCF ÷ CapEx. It’s cyclical because Tesla periodically cuts prices to drive volume, but the peaks are what matter. Those peaks have held in the 2.5–2.8x range even with rates ~4 to 5% higher than the 2022 era. Normalise for that rate drag and the post 2022 cash generation never actually slowed. It accelerated in relative terms.

Q1 2026 confirms the inflection: EMEA deliveries +150% QoQ, Berlin at a record 61k units, auto margin ex credits sequentially from 17.9% to 19.2%, highest Q1 backlog in two years. CFO Vaibhav Taneja shared how this was primarily driven by Tesla delivering more value per dollar to end customers:

Whilst the recent increase in gas prices has had a positive impact on the order rate, this improvement started before the uptrend in gas prices.

This is due to the work done by the Tesla team in bringing more compelling and affordable vehicles to market. 10 years back, when we launched Model 3 in the U.S. with a promise of $35,000 starting price, which if you adjust today for inflation translates to about $48,000 in today’s dollar terms.

The starting price of Model 3 today is way less than that while the product is way more compelling from where it started.

Tesla is a machine that prints autos, batteries, and soon robots. Manufacturing efficiency is the precursor of a value:price ratio no one can match at scale, which drives cash production. The $25B+ 2026 CapEx ramp will pull free cash flow negative for the year, but will also add new levers to monetise the underlying efficiency gains: Robotaxi at scale, Optimus production, and vertical integration into silicon via Terafab. Each of these will emerge as a multiplier of cash from operations, which in a CapEx cycle is a good proxy for free cash flow.

LTM operating cash flow has gone from $9B to $16B in five years, through the rate hiking cycle and through Tesla’s most aggressive price cuts. That is the baseline the new levers compound on.

Quick math on one of those levers. Robotaxi at 1M autonomous vehicles, 30k miles per vehicle per year, $0.40 contribution margin per mile is $12B in annual high margin recurring revenue, that’s highly accretive to the cash flow profile. Dropped on top of a $16B base that is still growing. And that is one lever of three. Optimus and Terafab sit behind it. This is what makes the next 2 to 5 years asymmetric. The cash engine that funds the CapEx is bigger than it has ever been and still scaling.

Further, cash from operations and CapEx pointing in opposite directions in Q1 2026 suggest the revenue may be even more accretive than expected. Both going up at the same time. Tesla is investing more aggressively than ever and still generating more cash than ever. The 2026 ramp accelerates the CapEx line further. The OCF line keeps compounding behind it, on a base that is structurally more efficient than the last cycle.

⚡ If you enjoyed the post, please feel free to share with friends, drop a like and leave me a comment.

You can also reach me at:

Twitter: @alc2022

LinkedIn: antoniolinaresc

Disclosure

These are opinions only of the individual author. The contents of this piece do not contain investment and/or medical advice and the information provided is for educational purposes only and no discussions constitute an offer to sell or the solicitation of an offer to buy any securities of any company or any drug or medical treatment. All content is purely subjective and you should do your own due diligence.

Antonio Linares makes no representation, warranty or undertaking, express or implied, as to the accuracy, reliability, completeness or reasonableness of the information contained in the piece. Any assumptions, opinions and estimates expressed in the piece constitute judgments of the author as of the date thereof and are subject to change without notice. Any projections contained in the Information are based on a number of assumptions as to market conditions and there can be no guarantee that any projected outcomes will be achieved. Antonio Linares does not accept any liability for any direct, consequential or other loss arising from reliance on the contents of this presentation. Antonio Linares is not acting as your financial, legal, accounting, tax, medical or other adviser or in any fiduciary capacity.

Tokens are profitable. The AI ramp just started.

Share

The key takeaway from this earnings report is that tokens are profitable, according to Jensen. This rewrites the entire AI thesis, which for three years has been softened by profitability concerns. The world is therefore now positioned to evolve from 1 billion human users to many billions of agentic users on top, which will drive staggering demand for additional AI compute. There’s no limit to the number of agents that can be deployed profitably and the total value each can produce. This is a historic value unlock, with tremendous implications for the entire AI value chain.

Tokens are now profitable, so model makers are in a race to produce more. In the AI era, compute capacity is revenue and profits.

Jensen Huang, NVIDIA Q1 FY27 Earnings Call

Subscribe now

The second takeaway is Physical AI is taking off, with Nvidia’s physical AI revenue crossing $9 billion over the trailing twelve months. This is implicit in the above model, in that many agents will interface with the world via physical devices. This is important for the AMD thesis, because physical AI demands inference at the edge, and AMD's ability to slap an FPGA onto any device delivers exactly that: hyper efficient local inference, optimised at the circuit level for the exact workload it runs. Jensen confirmed the structural logic directly:

There’s no choice but to put the computer where the context is, where the action is. You can’t put that in the cloud. It has to respond reliably, quickly every single time. Can’t imagine a chip plant, a chip fab being connected to a cloud service provider, doesn’t make any sense.

Jensen Huang, NVIDIA Q1 FY27 Earnings Call

This is what I believe sets AMD’s free cash flow per share growth on an Nvidia-like trajectory.

⚡ If you enjoyed the post, please feel free to share with friends, drop a like and leave me a comment.

You can also reach me at:

Twitter: @alc2022

LinkedIn: antoniolinaresc

Disclosure

These are opinions only of the individual author. The contents of this piece do not contain investment advice and the information provided is for educational purposes only and no discussions constitute an offer to sell or the solicitation of an offer to buy any securities of any company. All content is purely subjective and you should do your own due diligence.

Antonio Linares makes no representation, warranty or undertaking, express or implied, as to the accuracy, reliability, completeness or reasonableness of the information contained in the piece. Any assumptions, opinions and estimates expressed in the piece constitute judgments of the author as of the date thereof and are subject to change without notice. Any projections contained in the Information are based on a number of assumptions as to market conditions and there can be no guarantee that any projected outcomes will be achieved. Antonio Linares does not accept any liability for any direct, consequential or other loss arising from reliance on the contents of this presentation. Antonio Linares is not acting as your financial, legal, accounting, tax or other adviser or in any fiduciary capacity.

Palantir is at an Nvidia-2016 inflection point

Palantir becomes a $20T company.

Exponentials make little sense at the beginning, when progress is scarcely visible. They make even less sense later on, when progress is staggering. Palantir is where Nvidia was in 2016:

1. A (digital twin) monopoly.

2. Only starting to scale.

3. 0.01% TAM penetration.

4. Accelerates faster as it gets bigger, because it’s a Singularity Scaler.

AI capabilities keep emerging with incremental parameters, making Palantir's platform more valuable in tandem and at a marginal cost. Per Shyam, Mythos can now spot and fix vulnerabilities across the value chain: the "Sputnik" moment in the AI race. As value per token explodes, Palantir's free cash flow per share goes vertical.

Q1 2026 free cash flow is larger than Q1 2025 total revenue.

At the current rate, quarterly free cash flow per share will soon exceed total 2026 revenue. In the exponential realm, selling a winner early is a generational mistake.

Palantir solves a growing volume of acute customer pains in a way that’s increasingly harder to imitate. Every step in productisation and verticalisation widens the moat. The dynamic is fractal: every new vertical paves the way for N additional verticals, exponentiating the rate at which Palantir delivers specialised digital twins.

This is my next 100-bagger, joining AMD.

⚡ If you enjoyed the post, please feel free to share with friends, drop a like and leave me a comment.

You can also reach me at:

Twitter: @alc2022

LinkedIn: antoniolinaresc

Disclosure

These are opinions only of the individual author. The contents of this piece do not contain investment advice and the information provided is for educational purposes only and no discussions constitute an offer to sell or the solicitation of an offer to buy any securities of any company. All content is purely subjective and you should do your own due diligence.

Antonio Linares makes no representation, warranty or undertaking, express or implied, as to the accuracy, reliability, completeness or reasonableness of the information contained in the piece. Any assumptions, opinions and estimates expressed in the piece constitute judgments of the author as of the date thereof and are subject to change without notice. Any projections contained in the Information are based on a number of assumptions as to market conditions and there can be no guarantee that any projected outcomes will be achieved. Antonio Linares does not accept any liability for any direct, consequential or other loss arising from reliance on the contents of this presentation. Antonio Linares is not acting as your financial, legal, accounting, tax or other adviser or in any fiduciary capacity.

Whoever owns biomarkers owns healthcare.

Share

A company can outgrow its prior financial mould without outgrowing its value creation process. Hims has done this in Q1 2026 by becoming a two side operating system. The company’s previous financial model was total pills sold and now it’s total tokens delivered, as Hims becomes the intelligence layer that connects supply and demand in the healthcare system. To maximise free cash flow per share, Hims has to prioritise proprietary data over short term profits in order to deliver more value per token than generic AI models.

Claude is pretty good at delivering healthcare insights if you throw your biomarkers at it. The long term business opportunity is building an infrastructure that harvests biomarkers at a level of scale and quality that others cannot. At the network level, this is what allows you to train whatever frontier model to generate superior insights and deliver more value per token. More patients and more biomarkers than anyone means your proprietary AI becomes biologically more predictive. Long term, this is how you build the healthcare AI that front runs every other AI.

This puts you in a dominant distribution position because customers don’t care about second best, especially when it’s a matter of life or death. This is where the ability to charge extra per token stems from and thus, the new financial model is:

1. Number of tokens.

2. Multiplied by net margin per token.

This will serve as the best proxy for Hims’s free cash flow per share over the next decade, rising exponentially as both demand and supply lean ever deeper into the intelligence layer. The moat is the increasing difficulty of replicating the physical infrastructure required to bootstrap the biomarker Ontology, a barrier that compounds every quarter. And it compounds further still because the signals extracted from the physical layer feed directly back into making that same infrastructure bigger, better and harder to replicate.

Over time this flywheel will spin faster and faster. The more biomarkers the network produces, the faster the physical infrastructure evolves, printing even more and better biomarkers in return. This is not a quarterly play. It is a multi-trillion dollar, multi-decade wealth creation event.

Ontology Velocity is the single most important metric for Hims.

Investors have spent a great deal of time and energy focused on short term GLP-1 dynamics. As a subset of peptides, ex-legal dynamics which I dare not model, we will tend to see this part of the operation as a structurally deflationary one. As bio/acc goes mainstream, this will tend to be true for any healthcare offering. However, the pricing power of superior tokens will rise over time. Value is accruing to the network.

⚡ If you enjoyed the post, please feel free to share with friends, drop a like and leave me a comment.

You can also reach me at:

Twitter: @alc2022

LinkedIn: antoniolinaresc

Disclosure

These are opinions only of the individual author. The contents of this piece do not contain investment advice and the information provided is for educational purposes only and no discussions constitute an offer to sell or the solicitation of an offer to buy any securities of any company. All content is purely subjective and you should do your own due diligence.

Antonio Linares makes no representation, warranty or undertaking, express or implied, as to the accuracy, reliability, completeness or reasonableness of the information contained in the piece. Any assumptions, opinions and estimates expressed in the piece constitute judgments of the author as of the date thereof and are subject to change without notice. Any projections contained in the Information are based on a number of assumptions as to market conditions and there can be no guarantee that any projected outcomes will be achieved. Antonio Linares does not accept any liability for any direct, consequential or other loss arising from reliance on the contents of this presentation. Antonio Linares is not acting as your financial, legal, accounting, tax or other adviser or in any fiduciary capacity.

AMD is my first 100x investment. This is a break down of why I believe it will soon become a 1000X investment and more.

Share

AMD's free cash flow per share is going an order of magnitude or two higher over the next 5-10 years. The platform dominates new AI workloads as they emerge, without reinventing its technology to serve completely different businesses. As scaling laws persist, exponential demand for unforeseen workloads keeps accelerating. We don't know what AI will need next but AMD can deliver a total cost of ownership advantage at scale. This will increase the top line speedily and bring much of that down to the bottom line.

The best proxy for free cash flow per share is earning power, which broadly correlates to the strength and defensibility of the core technology multiplied by the number of vertical applications. AMD mastered chiplets between 2014 and 2021 and built a CPU business on top of that foundation. Free cash flow per share followed fast, because Intel could neither replicate the chiplet platform nor apply it successfully to CPUs. Chiplets won because they delivered higher manufacturing yield upfront and, over time, superior performance at a lower price.

The next chapter has been AMD applying chiplets to grow its AI GPU business, which Zuck and co started leaning into last year. That was one of the strongest signals that the new vertical was gaining traction, and one I highlighted incessantly. But the core thesis always remained that AMD is neither of its verticals. It is a platform with the growing ability to mix and match compute engines at will. A personalised compute printer of sorts, set to become exponentially more valuable as AI generates exponential demand for novel workloads in ways nobody can predict. FPGAs, for instance, are positioned to emerge next as a powerful new vertical instance.

CPUs are hot now because they turned out to be essential for AI workloads. That wasn’t predicted here, but the versatile nature of the platform was. Whatever new AI workload emerges, AMD can be expected to enter the scene with an increasingly dominant position. And most importantly for the long term thesis, we are at the very start of this process. CPU demand today is perhaps the first instance of AI creating demand for a workload nobody foresaw. But as we move towards Agentic AI, Physical AI, and other formats we cannot yet name, the volume of specialised workloads will rise exponentially.

This is where 99% of the upside in the AMD thesis resides.

Regardless, customers care about one thing: more intelligence per dollar. This requires optimising the entire compute value chain end to end in order to minimise latency. Some models will require a lot of compute and power to train and will be trained in extremely large datacenters. But the end customer may be far from that large cluster and would experience lower latency pulling inference workloads from a medium datacenter. In that case, it makes sense to train the model in the large datacenter first and then store it in the medium one, where it might be fine tuned over time.

In another example, the customer might be sitting at the Edge using a smartphone. The model in the above example that’s sitting in a medium datacenter might take too long to arrive to his location and therefore, latency would be lower if the model were stored in a small datacenter close to him. This would deliver more value per $ to this customer. Alternatively, the best option might be to store the model locally in his smartphone and have it run inferences locally. That would require a personalised compute engine, capable of storing a large model (lots of memory) and running inferences, on a strict power regimen.

Subscribe now

The potential scenarios are endless. This is fundamentally an optimisation problem where the goal is to minimise latency while remaining constrained by thermodynamic requirements and physical proximity. A datacenter in everyone's house would minimise latency perfectly, but it would not be economically viable. A massive datacenter running on everyone's phone would solve the proximity problem, but it would also set the phone on fire. The optimal distribution of compute across the value chain is a function of both constraints simultaneously.

Each case requires a specific solution, and that solution can only be delivered via personalised compute infrastructure at scale.

The nature of this problem is fractal. Personalised vertical solutions will tend to verticalize further in pursuit of delivering more value per dollar. CPUs, themselves a vertical instance of the platform, already break down into three further verticals:

1. general purpose CPUs that can do anything;

2. head nodes, the CPUs required to prepare data before it enters an AI accelerator;

3. and agentic CPUs, which power agents after they receive intelligence tokens from AI accelerators.

AMD’s chiplet platform allows them to print any variant or sub-variety of these CPUs at marginal cost. No one can meaningfully compete with that.

General purpose CPUs need flexibility, because they must be ready to perform any task. Head nodes require higher memory bandwidth and I/O capabilities, since large volumes of data pass through them before reaching AI accelerators. Agentic CPUs need to minimise serial latency while maximising memory, because agents perform one task after another and what they do next depends entirely on what they just did and what is happening in the environment. Each specification demands a specialised architecture, and the average customer will likely need all of them at scale.

Who is to say we will not see specialised Agent types emerge that require their own architectural configurations to operate profitably? The odds are extremely high for all three CPU types. The verticalisation has no logical end.

The financials now are appealing because several of these verticals are starting to work in parallel. Over time, the volume and number of these verticals will tend to increase faster as AI scales, with AMD being able to provide at a marginal cost. In this sense, the rise in free cash flow per share below is appealing, but is likely only an appetiser. Over the coming 5-10 years, I believe this metric will rise vertically by several orders of magnitude. This is why I choose to continue holding my AMD shares, despite having obtained a return of 100X since inception.

⚡ If you enjoyed the post, please feel free to share with friends, drop a like and leave me a comment.

You can also reach me at:

Twitter: @alc2022

LinkedIn: antoniolinaresc

Disclosure

These are opinions only of the individual author. The contents of this piece do not contain investment advice and the information provided is for educational purposes only and no discussions constitute an offer to sell or the solicitation of an offer to buy any securities of any company. All content is purely subjective and you should do your own due diligence.

Antonio Linares makes no representation, warranty or undertaking, express or implied, as to the accuracy, reliability, completeness or reasonableness of the information contained in the piece. Any assumptions, opinions and estimates expressed in the piece constitute judgments of the author as of the date thereof and are subject to change without notice. Any projections contained in the Information are based on a number of assumptions as to market conditions and there can be no guarantee that any projected outcomes will be achieved. Antonio Linares does not accept any liability for any direct, consequential or other loss arising from reliance on the contents of this presentation. Antonio Linares is not acting as your financial, legal, accounting, tax or other adviser or in any fiduciary capacity.

The narrative is entirely dislocated from Duolingo’s fundamentals and the upside is exponential.

Share

Duolingo has a 100% track record of improving any metric it decides to focus on. When high-level numbers dip, most investors extrapolate the decline indefinitely — but when a company's organisational properties remain intact, as is the case with Duolingo, the financials rarely stay down for long. And the stock follows rather quickly, as we saw with Meta and Amazon in 2022, or more recently AMD in 2025.

What's peculiar here is that metrics aren't actually pointing downwards, despite the broader sentiment suggesting otherwise. MAUs grew from 133.1M to 137.8M quarter over quarter, and DAUs from 52.7M to 56.5M. The year-over-year numbers aren't bombastic, but the narrative clearly doesn't reflect the fundamentals. Duolingo doesn't appear to be dying at all and if anything, seems to be picking up steam again.

Further, engagement is going up in a straight line ever since LLMs entered the scene, as depicted by the DAU/MAU ratio. Q1 2026 saw a notable uptick too.

As DAU/MAU goes up, it means that more monthly users are deciding to engage daily with Duolingo. Engagement is a good proxy for teaching quality, because users spend more time on Duolingo as they get more value from it. In turn, the rate at which teaching quality improves is a direct function of the shipping rate, which seems to be going up quite a bit and perhaps explains the DAU and MAU upticks:

We now offer courses up to professional proficiency, which is B2 on the CEFR scale across all our 9 most learned languages and we got there fast.

In Q1 alone, we published 20,500 course units. To put that in context, that’s more than 10x what we were shipping per quarter just 2 years ago.

AI has fundamentally changed what’s possible for us, and I believe we’re just scratching the surface. The product is better than it has ever been, and I couldn’t be more excited about what’s ahead.

-Luis von Ahn, Duolingo CEO during the Q1 2026 earnings report.

Average words spoken per Video Call is a proxy for Duolingo's progress toward becoming a superhuman AI tutor. At the limit, this metric could theoretically reach pathological levels, not unlike what Meta has achieved with Reels, where the AI is not only an effective teacher but knows exactly what to say to keep you on the call. Words spoken per user are up over 2x year over year. That's a better indicator of where things are headed than any headline and in line with the above observations.

The above is why the DAU growth deceleration is cyclical rather than structural. It takes time for deeper metrics to show up in YoY DAU growth, but Duolingo's renewed focus on user growth is working, as evidenced by the QoQ uptick in DAUs and the DAU/MAU ratio. For the historians among us, Duolingo experienced something similar in 2021 and DAU growth came in at 60%+ a year later. Conditions then are not the same as now, but it illustrates Duolingo's ability to focus on user growth cyclically and, over the long term, grow the platform in absolute terms.

Subscribe now

Fundamentally, AI continues to accelerate Duolingo’s value creation process. Generic AI models can teach, but Duolingo has a lot of behavioural data that allows it to understand how to motivate people through a learning experience. Using that data to fine tune a generic model produces better results than the generic model alone, which is why DAUs are not plummeting to zero and DAU/MAU continues going up in a straight line. The volume and quality of content on Duolingo is going up exponentially, to the point that hyper-personalised content is starting to make sense for them, according to Luis during the Q1 2026 earnings call:

And we’re also likely going to move to something where we have all this content, this is awesome, but we may even start generating content just for you.

Based on everything that we know about you, we may just be able to generate content just for you, maybe not the immediate next exercise, but like two exercises from now based on everything.

As you may have realised, this is a pristine example of the Personalised Token mental model. Using tokens is not an option, but a token is much more valuable to someone when it’s personalised and for that, you need a lot of data about that person. Duolingo is effectively an Education Ontology, which enables them to deliver more value per token than anyone else for their specific goal of improving learning outcomes. Duolingo’s shipping rate is an order of magnitude higher than two years ago and it won’t take long to increase several by orders.

In the AI space, competition happens at the infinitesimal edge. This is what the market has been trying to understand throughout 2026.

If your Ontology has a little more proprietary data, which enables you to train an AI that's marginally better than the next one, that marginal improvement compounds into an exponential advantage in user experience, which drives more users, which generates more proprietary data. The flywheel doesn't just spin faster. It becomes harder to stop.

So long as the model fine tuned with proprietary data offers even marginally more value per dollar to end customers, the company will stay ahead of the curve. This is because second best is meaningless to consumers, as we saw with Spotify in 2022, when the market believed Amazon and Apple were going to kill it. In the digital space, a marginal advantage translates into an exponential outperformance over the long term. AI is only enhancing this dynamic.

Free cash flow per share (blue line, below) is broadly tracking the recent DAU and DAU/MAU upticks. Duolingo is currently in investment mode, having deferred near-term financial guidance in favour of long-term value creation. What's interesting is that even in this context, Duolingo remains an extraordinary cash flow compounder. If we take free cash flow per share as the best proxy for intrinsic value, which it is, the latter is up 86% while the stock is down by roughly the same amount.

As a result, as Duolingo’s fresh round of enhanced experimentation translates into reaccelerated DAU growth, it seems that much of that will be highly accretive to the bottom line.

While Palantir is the first instance of a Singularity Scaler, Duolingo is the first instance of a Singularity Asymmetry. Duolingo’s intrinsic value is increasing in tandem with AI scaling laws, which are turbocharging its value creation processes. The market, however, has decided the opposite is true, with the stock decaying exponentially. The delta between these two realities is a Singularity Asymmetry: two exponential curves converging to yield a quadratic wealth creation event.

Until next time!

⚡ If you enjoyed the post, please feel free to share with friends, drop a like and leave me a comment.

You can also reach me at:

Twitter: @alc2022

LinkedIn: antoniolinaresc

Disclosure

These are opinions only of the individual author. The contents of this piece do not contain investment advice and the information provided is for educational purposes only and no discussions constitute an offer to sell or the solicitation of an offer to buy any securities of any company. All content is purely subjective and you should do your own due diligence.

Antonio Linares makes no representation, warranty or undertaking, express or implied, as to the accuracy, reliability, completeness or reasonableness of the information contained in the piece. Any assumptions, opinions and estimates expressed in the piece constitute judgments of the author as of the date thereof and are subject to change without notice. Any projections contained in the Information are based on a number of assumptions as to market conditions and there can be no guarantee that any projected outcomes will be achieved. Antonio Linares does not accept any liability for any direct, consequential or other loss arising from reliance on the contents of this presentation. Antonio Linares is not acting as your financial, legal, accounting, tax or other adviser or in any fiduciary capacity.

Spotify’s enterprise value is a direct function of its ability to personalise, which is about to go through a discontinuous leap.

Share

Spotify compounds free cash flow per share by personalising media at marginal cost. On a first principles basis, it’s a media engine that better predicts what you want to consume and does so more cheaply over time. The free cash flow per share explosion has been primarily the result of expanding this capability from one vertical (music) to a growing range of them, multiplying user lifetime value with marginal incremental investment, enabled by personalisation built on statistical signals.

Spotify is now entering a new era in which personalisation goes vertical, powered by AI and semantic feedback, and with it, free cash flow per share. Stock prices track free cash flow per share over the long term. Having entered at $97.50, I remain positioned for a historic return.

DJ has nearly 100M users and SongDNA has 52M users four weeks after launch. Apart from being engaging features, these are primarily sources of semantic proprietary data. Users tell Spotify what they want in natural language, which enables personalisation orders of magnitude more precise than anything seen to date. The leap is discontinuous, rendering statistical signals practically medieval overnight.

The financials we see, summarised by free cash flow per share above, are effectively obsolete. Not only has the stock price deviated from intrinsic value, but it is also completely oblivious to what is coming.

The stock went down 17% yesterday because OpEx is expected to trend up as Spotify continues to invest in building and marketing AI features such as the above. However, the graph below illustrate’s management’s ability to balance profitability and investments: Alex Norstrom, current co-CEO, is the guy who drafted Spotify’s plan to become a profitable business some years ago. Spotify continues to execute the Costco Algorithm, applied to media, as Alex explained during the Q1 2026 earnings call:

[…] we optimise for the long term, and we talk about optimising for lifetime value.

How do you bridge these two things with an ever-increasing catalog of content and lifetime value?

You know, it turns out that the number one reason for why people actually engage more with Spotify is personalisation.

You know, how we track that is if AI increases engagement for us, it generally means that it increases personalisation for us, right? Increased personalisation engagement, to Gustav’s points, are going to lead to.

They are gonna be the best proxies for the increase in retention that we’re gonna see over time with these investments.

If that happens, then we know that that will eventually translate to a longer lifetime value, which in turn translates to more enterprise value. That’s how we think about the investments.

The market does this often, but naturally long term investors shouldn’t be disappointed when a management team focuses on long term value creation.

Subscribe now

The second driver of enterprise value is the rate of incremental personalisation: as Spotify ships faster, it creates value faster. I often get asked, why use the term Ontology? Because you have to label and understand concepts clearly.

The market has broadly believed for the past year that software is dead, which is largely true, but that doesn’t mean “software companies” are dead. Some will die, but others have already evolved into something much bigger: Ontologies. Proprietary data sets are the moat of the 21st century and if you can translate that into an AI that’s marginally better than the next one, you get more proprietary data which keeps you ahead.

If your AI is 0.0001% better and your Ontology spins just a tiny bit faster, you take the whole market because consumers don’t care about second best. Spotify is a Taste Ontology and no matter how good generic AI models get, they won’t capture taste as well so long as Spotify continues increasing engagement and it’s ability to extract semantic signals. Co-CEO Gustav Soderstrom shed light on this during the Q1 2026 earnings call:

In terms of type of traffic, it depends on the feature, but what Alex mentioned up front is we have, for the first time in Spotify history, this ability for users to actually tell us in plain English, or actually whatever language they want, what they want.

[…]

I talked last time about the large personalisation model, which is a model that we’re training from based on open source models, but it’s trained on our proprietary data. This is not something that we rent from someone. This is something we’re building in-house. You know, the casual name for the large personalization model is a taste model. Why is that important?

It is because it turns out that taste is actually not a fact. It is an opinion, it differs between people, between markets, between use cases and activities.

Spotify’s Ontology Velocity is going up, according to Gustav:

We are spending more compute per employee. That is because we’re seeing tremendous return in terms of productivity. We talked about accelerating our ability to ship products already during the late fall. That has only accelerated since then. We’re simply doing much more. We’re getting very good return on that investment.

[…]

We think this opportunity is as big or possibly bigger [than the iPhone]. We’re taking that opportunity. We are very diligent and very disciplined about those investments.

Until next time!

⚡ If you enjoyed the post, please feel free to share with friends, drop a like and leave me a comment.

You can also reach me at:

Twitter: @alc2022

LinkedIn: antoniolinaresc

Disclosure

These are opinions only of the individual author. The contents of this piece do not contain investment advice and the information provided is for educational purposes only and no discussions constitute an offer to sell or the solicitation of an offer to buy any securities of any company. All content is purely subjective and you should do your own due diligence.

Antonio Linares makes no representation, warranty or undertaking, express or implied, as to the accuracy, reliability, completeness or reasonableness of the information contained in the piece. Any assumptions, opinions and estimates expressed in the piece constitute judgments of the author as of the date thereof and are subject to change without notice. Any projections contained in the Information are based on a number of assumptions as to market conditions and there can be no guarantee that any projected outcomes will be achieved. Antonio Linares does not accept any liability for any direct, consequential or other loss arising from reliance on the contents of this presentation. Antonio Linares is not acting as your financial, legal, accounting, tax or other adviser or in any fiduciary capacity.

And Nautilus Biotechnology, with a $380M market cap, is right at the center of it.

Share

Nautilus is pre-revenue, but three things are true:

1. No one is even close to their iterative mapping platform.

2. They have a clear track record of scaling platform capabilities while keeping costs in check.

3. And while it’s too early to evaluate Nautilus based on financials, the recent milestones suggest the odds of them succeeding eventually are material.

Iterative mapping consists in cycling fluorescently tagged probes (amino acid chains) across immobilised proteins and registering which ones bind. Every time a probe binds, a fluorescent light emission gets registered by a computer vision model. The aggregate of all light emissions is used to create a computational model of a protein's shape, which is just the inverse of all the probes that bound.

Iterative mapping thus solves the epistemic ceiling of mass spectrometry. The latter cannot possibly capture the state of a protein because it fragments it in the process of reading it, destroying the very modifications it is trying to detect. Iterative mapping, by contrast, reads the protein intact and in place, cycling through probes repeatedly until a complete picture emerges, giving it, in principle, infinite resolution over the proteoform state.

In practical terms, this means that iterative mapping is the path to digitising the proteome. Many proteoforms can emerge from a single protein: the former are merely slight variations of the latter, characterised by adding a glycan or a phosphate, for example. The diagram below shows how a single gene produces multiple base proteins, each of which can then be modified at specific sites via different methods.

For example, mass spectrometry cannot tell you if you are looking at a protein with two glycans at positions A and B or two proteins, one with glycan at A and another at B. However, being able to map out that difference is essential to truly understand the proteome and understand illness deeply enough to stop it at the root level. Every physiological process in the body emerges from a proteoform chain and by extension, so does all illness. Cells read genetic code and print a base protein, which then is converted into any number of proteoforms via post translational modification (PTM).

Illness is simply a coding bug at any stage in the diagram below, plus the fact that many proteoforms can emerge from a single protein and that proteins in general combine with other proteins to create entirely new functional structures.

To understand the power of a body-wide digital twin of the above information, consider Alzheimer’s and Parkinson’s, which manifest as proteoform misfolding diseases. In Alzheimer’s, amyloid beta plaques and hyperphosphorylated tau tangles destroy synaptic architecture. In Parkinson’s, alpha synuclein misfolds into Lewy bodies and kills dopaminergic neurons. At a deeper level, both seem to emerge from dysfunction at the same proteomic level.

Both diseases share a common upstream failure: the mitophagy axis. PINK1, the kinase that flags damaged mitochondria for disposal, and OPTN, the receptor that executes the clearance, are themselves proteoform dependent. When either arrives in a dysfunctional proteoform state, damaged mitochondria accumulate, oxidative stress floods the cell, and the misfolding cascade begins.

The plaques and tangles are not the disease. They are the wreckage left behind by a garbage disposal system that broke first. This remains a hypothesis, but the broader point stands: we will soon be able to map any illness at the proteomic level, and then deploy algorithms like AlphaFold and LigandForge to design targeted molecular interventions that halt it before it ever manifests clinically.

Subscribe now

In Q4 2025 Nautilus management announced the launch of the early access program for tau assays — ahead of schedule. Collaborators began submitting samples and receiving data, with initial response described as encouraging. The program is not designed to drive near term revenue but to support publications, grant applications, and workflow validation, which is exactly how transformative life sciences platforms have historically entered the market. The clinical signal is early, but the commercial motion has started.

This comes after announcing in Q4 2024 that they would shift resources towards reconfiguring their platform:

By making the decision to pursue modifications to our assay configuration, surface chemistry and related platform elements at this time, we believe that we are positioning Nautilus to ultimately make the maximum possible impact on the marketplace and on biological science.

This elongated development timeframe necessitated that we reevaluate our operating plan and organizational structure to ensure that we are in the best position to execute against both our broadscale and targeted proteoform goals.

Nautilus CEO and co-founder, Sujal Patel, during the Q4 2024 earnings call.

The tau assay is the beachhead, not the ceiling.

Nautilus is now moving into Parkinson’s, beginning with an 18-month collaboration funded by the Michael J. Fox Foundation to develop an alpha-synuclein proteoform quantification assay. This matters because alpha-synuclein is precisely the protein whose dysfunctional proteoforms drive the misfolding cascade described above, making this not just a commercial expansion but a direct scientific validation of the thesis.

We see a familiar pattern emerge: Nautilus is leveraging its core platform to deploy a growing volume of verticals at a marginal cost. As they launch and mature new verticals, the time to value will shrink, yielding an operational flywheel similar to Palantir’s: more productisation yielding incremental network effects. Nautilus Chief Scientific Officer, Parag Mallick, commented on this during the Q4 2025 earnings call:

While much of our current momentum is in neuro degeneration, it’s important to emphasise that Iterative Mapping is not limited to neuroscience.

We see meaningful long-term potential across oncology, immunology, cardiology and beyond.

We are currently evaluating multiple oncology-focused candidate proteins with the goal of having an oncology-focused proteoform assay enter early access in the second half of 2026.

The long term thesis rests on the innate superiority of the core platform, of which we have no exhaustive confirmation. However, dynamic range leaves a clue: the industry’s average is 1 order of magnitude, while Nautilus’ is over 3 orders of magnitude. That means the platform can detect proteins present at vanishingly low concentrations alongside those present at high abundance in the same sample.

This is the difference between seeing the full proteomic picture and seeing only the loudest signals. Most disease-relevant proteoforms, including the specific truncations and modifications that precede clinical symptoms, exist at low abundance. A platform that cannot see them cannot find the bug before it crashes the system: let alone intervene early, which is precisely the point if the long term business is prevention and optimisation rather than damage control.

What makes the platform hard to replicate (and increasingly so as new verticals are deployed, because every new protein class magnifies the error) is that binding kinetics change when fluorescent tags enter the picture. Tags are bulky molecules, and bulk interferes with precision binding. Attaching one to a probe risks altering the very interaction you are trying to measure, introducing noise that is hardest to separate from genuine signal at exactly the low abundance range where disease-relevant proteoforms live.

My read on the competitive environment is largely intuitive. It is too early to be certain, but this feels like one of those rare situations where there is simply no one else in the room.

Despite being pre-revenue, Nautilus presents a reassuring financial profile. Cash has declined from its 2021 peak of $345.7M to $103.4M at the end of 2025, which is a natural drawdown for a company in heavy development mode. But the burn rate has improved meaningfully, with operating losses narrowing from $81.5M in 2024 to $66.8M in 2025. Debt has remained flat at $30M for three consecutive years. The company is not in a race against a collapsing runway. It has time to execute.

I believe iterative mapping is on course to scale population-wide. While it currently requires a lab setting, the near term vision is a monthly proteomic snapshot processed immediately by an AI model alongside all other longitudinal biomarkers. Over the longer term, as the technology miniaturises, the destination is a real time digital twin of your proteome with AI running permanently in the background, optimising your protocol to maximise health outcomes before any disease has a chance to declare itself.

We are curing everything.

Until next time!

⚡ If you enjoyed the post, please feel free to share with friends, drop a like and leave me a comment.

You can also reach me at:

Twitter: @alc2022

LinkedIn: antoniolinaresc

Disclosure

These are opinions only of the individual author. The contents of this piece do not contain investment advice and the information provided is for educational purposes only and no discussions constitute an offer to sell or the solicitation of an offer to buy any securities of any company. All content is purely subjective and you should do your own due diligence.

Antonio Linares makes no representation, warranty or undertaking, express or implied, as to the accuracy, reliability, completeness or reasonableness of the information contained in the piece. Any assumptions, opinions and estimates expressed in the piece constitute judgments of the author as of the date thereof and are subject to change without notice. Any projections contained in the Information are based on a number of assumptions as to market conditions and there can be no guarantee that any projected outcomes will be achieved. Antonio Linares does not accept any liability for any direct, consequential or other loss arising from reliance on the contents of this presentation. Antonio Linares is not acting as your financial, legal, accounting, tax or other adviser or in any fiduciary capacity.

Biology is next.

I’ve spent years studying it and it’s going to drive most of the returns over the next next 20 years. The recent peptide craze is just the surface of a much deeper and powerful force.

Here’s everything you need to know.

The body is just a power plant that runs a giant LEGO puzzle. Mitochondria produce energy, which cells use to print proteins.

Shape determines function. Each protein binds onto another protein with an inverse shape, like LEGO.

The human proteome is just proteins interacting with each other, driven by predictable and weak electromagnetic attractions.

It’s just code, as I will explain.

And we can now do what we want with it.

Cells read DNA and produce amino acids. Cells then tie amino acids together to produce peptides.

Peptides then coil up and fold (driven by electromagnetism) to create proteins.

Cells read genetic code and print proteomic code.

Every physiological process in the body emerges from this axis.

For example, below is a peptide. It’s just 7 peptides tied together.

Each peptide can be represented by a letter.

The atoms at the edges carry a charge: they can be negative, positive or neutral.

This makes peptides coil up and form proteins.

We can now predict which amino acid sequences will produce specific proteins and determine their final structures, thanks to Alpha Fold.

For example, below is a collagen protein.

Collagen can also be represented as:

GPPGPSGPPGPPGPPGPPGFPGARGPSGPQGFQGPPGEPGEPGASGPMGPRGPPGPPGKNGDDGEAGKPGRPGERGPPGPQGAR

Each letter represents an aminoacid:

Then cells can turn proteins into proteoforms - slight variations of the original protein, with slightly different functions.

One common method is phosphorylation, which just adds one phosphorus atom and four oxygen atoms at different parts of a protein.

Remember, shape determines function.

Proteoforms are just (more) code.

Below you can see collagen’s phosphorylation map, by position and kinase type.

A kinase is an enzyme that adds phosphate groups to proteins, switching them on or off.

There’s different kinase types.

It’s just code.

Everything in the body emerges from the central proteomic axis, which we can now update with the concept of proteoforms.

Cells have many different ways of creating proteoforms.

You don’t need to know about all of them.

You just need to know that they produce code, because AI will understand it for us.

To illustrate, consider Glycosylation.

It’s just about adding glycans (sugars) at different parts of a protein.

It just changes the shape of a protein slightly, modifying its function in the LEGO puzzle.

This is the method cells use to share information about their own state with other cells in the body.

Cells leverage glycosylation to say “I’m well fed” or “I’m not well fed.”

“I have a pathogen invading me.”

It’s just code.

It follows that all illness is rooted at this axis.

It manifests in this axis.

It can be stopped before it even manifests.

The proteome is complex, but also simple in a way.

It’s just three dimensional structures interacting with each other and we’re about to map this out exhaustively thanks to a technology called Iterative Mapping.

We’re all getting a digital twin of our proteome soon.

This is about to change medicine forever by creating a new key:value pair.

Stimuli: δProteome.

If we know exactly what’s happening in our proteome at all times, we will know:

1. Exactly what effect a molecule (or any other therapeutic vector) has on our body, at the proteomic level.

2. How to stop illness at the root before it manifests.

3. At essentially 0 cost, by using the key:value pair to train a superhuman AI doctor.

We don’t need to understand illness any more.

We just need proteome digital twins and to dump them into an AI model. As happened with the dog that had cancer, the AI will see things.

The AI will just see proteomic code bugs and figure out how to fix them.

This is the true power of peptides: they are proteomic code scripts that we can use to modify any part of our proteome with 0 side effects.

Many conditions that we consider incurable today will be cured with a single prompt.

Personalised peptides (amino acids chains) will fix them with surgical precision, just like we fix code on a screen today. And personalised peptides were only just made possible yesterday by Ligand Forge.

For example, consider Parkinson’s and ALS.

Two mysterious illnesses, that in fact seem to be the same proteomic coding error at different levels of the value chain.

HHV-6 (Herpes virus variant 6) viral proteins corrupt the OPTN–TBK1 mitophagy checkpoint, leaving damaged mitochondria untagged and invisible to autophagy, degrading nervous system tissue over time.

Just a hypothesis, yes.

But it doesn’t matter, AI will know. And specifically, the network will know.

Once anyone has a digital twin of their proteome, this is how fix any illness:

1. Cluster proteome digital twins by symptoms.

2. Throw them into an AI.

3. AI maps symptomatic picture (that we summarise via a name, such as Alzheimer’s) to proteomic dysfunction.

4. AI identifies key parts of proteome to engage with to stop and revert dysfunction.

5. LigandForge computes personalised peptide for each patient.

6. Proteome end state is assessed.

7. AI learns.

If we do this at the network level, the AI has more and more data about how to fix specific proteomic coding errors.

This drives better customer outcomes, which brings more customers, which generates more data.

This culminates in more value delivered per token: less hallucinations, no side effects, optimal outcomes, enhanced longevity per dollar spent.

This is the most valuable flywheel of our time.

Peptides are just the tip of the iceberg.

This space is going from 0 to a billion fast, because 3 things are happening:

1. The cost of diagnostics is trending towards zero.

2. The cost of peptides is trending towards zero (as evidenced with GLP-1 prices dropping 80% in the US YoY).

3. AI capabilities are rising vertically.

This process cannot be stopped.

We don’t understand illness, but we will suddenly understand it entirely once we start dumping proteome digital twins into AI models.

The proteome is code and we’re learning to write it.

Block has gone from broken to seemingly compounding, which merits an ongoing review of my assessment of its management back in 2023.

Share

Ironically, Block is now testament to my view that organisational traits are among the best predictors of financial performance. A few years ago it was a pain to read Block's earnings call transcripts. Today, the company exudes focus and clarity and the ultimate sign of this shift is the rapidly rising free cash flow per share, that you can see depicted below (orange line). This has been brought about by four main factors, which have compounded since early 2024:

1. Unifying Block’s organisational structure (from two companies to a single one) and increasing talent density.

2. Leverage AI capabilities to ship faster, which according to Jack went through an inflection point in December 2024.

3. Streamlined pricing, together with a more focused field sales force.

4. New buyer features driving non-linear engagement increments across the network.

The company is now smaller, faster and much more focused on delivering more value to end customers per dollar spent. If you read the last few earnings call, you will see this too.

We’re also inherently increasing talent density across the org, including the development org. And obviously, we’re getting all the benefits of the AI tools and the improvements in the foundational models, and that’s flowing through to everything that we’re doing, particularly on the software development side.

-Owen Jennings, Block’s Business Lead, during the Q4 2025 earnings call.

Subscribe now

As you can see above, the market isn’t quite buying this “shift” and I don’t have a great deal of conviction either. However, Block’s fundamentals today are noteworthy:

1. They’re sitting on a Payments/Lending Ontology which is fairly impossible to replicate at this stage: 59M MAUs on the Cash App side and a GPV of $67.2B in Q4 2025.

2. This affords them plenty of proprietary data, which they can use to resell intelligence to both sides of the marketplace: a CFO for buyers and a COO for merchants, as Jack put it.

3. As is the case with Shopify and Spotify, which are also leveraging agentic tech correctly, this positions Block to bring a whole new level of personalisation to customers.

And I do believe that a lot of the expectation going forward is not just that we deliver intelligence to our customers through Managerbot and Moneybot, for instance, but the very end of that is that they can build their own features and build their own visualisations on top of our capabilities through our interfaces.

We have incredible distribution.

We have incredible understanding in real time of real transactional data on both sides of the counter.

-Jack Dorsey during Block’s Q4 2025 earnings call.

Putting the loan origination risk aside, there’s plenty of continuity to Block’s current trajectory per the aforementioned non-linear engagement increases, being unlocked by incremental features on the buyer side. For example, management shared in the Q3 2025 earnings call that Borrow (cash advance feature within Cash App) actives had 3X higher inflows and 2X higher Cash App Card spend and 3X higher retention rate than non-Borrow actives. In Q4 2025 they shared that banking (Cash App's FDIC-insured deposit account with a debit card.) actives had 10X higher engagement than only peer-to-peer actives.

As previously mentioned, AI is clearly enabling Block to figure out what customers want and ship faster, which makes me think we’ll continue to see non-linear earning power improvements. Indeed, Block as a lot of competition, like perhaps Shopify, but it appears that its biggest competitor was itself. Once they fixed the organisational structure and got more focused, the financials are improving rapidly.

In my Q3 2023 earnings digest, I said the following about Block:

Management is not doing anything tangibly wrong. Nonetheless, for the reasons I explain above, I believe that the company lacks:

Clarity and focus.

A general regard for shareholders.

General Electric was a great case study on how deteriorating organisational properties can rapidly bring down an industrial giant. Block may just become an example of the opposite. In the graph below, it is now apparent that I was right to pick Spotify over Block and Shopify back in 2022. However, Block's evolving organisational properties have put me on high alert. Did I perhaps wrongly assess them back then and confuse a radical focus on achieving scale with shareholder unfriendliness?

I understand that to the average investor this means nothing. But over the years I have developed the conviction that correctly interpreting qualitative datapoints is the difference between making a fortune and making no money at all. I have also been teaching the importance of prioritising scale over profits in the digital economy for longer than the average analyst can endure to track. I may just have made that same mistake and will continue studying Block, to see what I can learn.

⚡ If you enjoyed the post, please feel free to share with friends, drop a like and leave me a comment.

You can also reach me at:

Twitter: @alc2022

LinkedIn: antoniolinaresc

Disclosure

These are opinions only of the individual author. The contents of this piece do not contain investment advice and the information provided is for educational purposes only and no discussions constitute an offer to sell or the solicitation of an offer to buy any securities of any company. All content is purely subjective and you should do your own due diligence.

Antonio Linares makes no representation, warranty or undertaking, express or implied, as to the accuracy, reliability, completeness or reasonableness of the information contained in the piece. Any assumptions, opinions and estimates expressed in the piece constitute judgments of the author as of the date thereof and are subject to change without notice. Any projections contained in the Information are based on a number of assumptions as to market conditions and there can be no guarantee that any projected outcomes will be achieved. Antonio Linares does not accept any liability for any direct, consequential or other loss arising from reliance on the contents of this presentation. Antonio Linares is not acting as your financial, legal, accounting, tax or other adviser or in any fiduciary capacity.

At the current rate of execution, Immunity Bio will be an API you query with your biomarkers (maybe even asynchronously, without you doing anything) and it just ships a personalised reset button for your immune system that clears tumors and pathogens with minimal side effects.

Share

The immune system is an integral component of the human body: without it, our organism cannot fight off pathogens and we die. ImmunityBio is building a platform that brings immune system rejuvenation to the world at a marginal cost. The foundation of this platform is the IL15 superagonist named Anktiva, which instructs the body to produce more natural killer and T cells. These are the basic building blocks the immune system uses to kill, and they can be trained to target specific threats, which is how the platform comes to be.

If you have good natural killer and T cell counts, your immune system can work properly and your mortality trends down, and vice versa. If GLP1s have made Eli Lilly a $1T company, Anktiva can make ImmunityBio a $2T company (200 times bigger than today) or more because everyone will need to rejuvenate their immune system multiple times throughout their lives.

Three things stand out from the Q4 2025 earnings report:

1. Net product revenue is up 700% YoY, which is driven by a 750% increase in unit sales volume. Meaning, both physicians and patients want Anktiva, because it’s solving a real problem.

2. Immunity Bio has demonstrated an organisational ability to drive approvals, ever since obtaining the FDA J code in January 2025. Each quarter of 2025 stronger than the last, with Anktiva now authorised across 33 countries and 4 major regulatory jurisdictions, including the United States, the United Kingdom, the Kingdom of Saudi Arabia, and the entire European Union.

3. The company is tentatively accelerating as it scales.

The vast majority of Immunity Bio’s revenue today comes from the US, where the regulators remain somewhat reticent: this is similar to the situation Grail’s early-cancer-detection test Galleri is in. Anktiva was approved by the FDA in April 2024 for use in combination with BCG (Bacillus Calmette-Guérin, a live weakened strain of the bacterium Mycobacterium bovis that jolts the immune system into action) for BCG-unresponsive non-muscle invasive bladder cancer. However, as discussed, on a first principles basis there’s no difference between this cancer and any other.

If the immune system is strong enough (has enough natural killer and T cells), it can kill any tumor. And if the patient isn’t in enough health to naturally train the NK and T cells Anktiva produces, Immunity Bio has additional platform layers which can help:

1. NaNT (N-803 + Allogeneic NK Cells): built on top of the Anktiva foundation, this arm of the platform combines Anktiva with allogeneic natural killer cells: meaning NK cells derived from a donor rather than the patient.

   1. The Leonardo robotics platform is how ImmunityBio manufactures these off-the-shelf NK cells at scale.

   2. Anktiva then acts as a persistence signal, keeping these infused NK cells alive and active longer than they would otherwise survive in a foreign host, solving one of the core challenges of allogeneic cell therapy.

2. DNA Vaccine (Antigen Targeting): this layer adds specificity to the platform. While Anktiva raises the overall count of NK and T cells and the allogeneic NK cells provide additional firepower, the DNA vaccine trains those cells to recognise and kill specific targets: tumor antigens, viral antigens, and so on.

   1. This transforms the platform from a broad immune amplifier into a precision weapon, capable of directing a now-enlarged and reinforced immune army toward a defined threat.

There’s value in building solutions on top of Anktiva because the pathogenosphere is nuanced. For example, tumors will sometimes hide certain receptors (things that help T cells locate a tumor, like MHC-I) when they sense a T cell coming towards them. For millions of years NK cells have been acting as the backstop in this scenario. NK cells operate on an inverse logic: they kill anything that lacks the markers of a healthy cell. Leonardo can engineer the following NK cell types:

1. CAR-NK: NK cells engineered to express chimeric antigen receptors that lock onto specific tumor antigens. This gives NK cells the targeting precision of CAR-T therapy while retaining their innate killing logic, circumventing tumors’ ability to hide receptors. And without the cytokine storm risk that has historically made CAR-T a difficult therapy to administer safely.

2. PD-L1 NK: many tumors survive by expressing PD-L1, a signal that tells the immune system (T cells specifically) to stand down. PD-L1 NK cells are engineered to target exactly that take out that signal, letting T and NK cells kill the tumor.

Subscribe now

The platform isn’t exhaustive just yet, but powering up the immune system at the root and then deploying additional laters that collapse time to value in specific problems is the path to getting there. At the current rate of execution, Immunity Bio will be an API you query with your biomarkers and it just ships a personalised reset button for your immune system that clears tumors and pathogens with minimal side effects.

For example, consider the “DNA vaccine” layer of the platform. It leverages an adenovirus to deliver a specific antigen to dendritic cells. The adenovirus has an icosahedric shape (a polyhedron with 20 faces) that per its shape, tends to “swim” towards the lungs when injected into the body.

Remember, shape determines function in the proteome.

The adenovirus reaches dendritic cells in the lungs, offloads the antigens, which the dendritic cells then transport to T cells (previously matured in the thymus gland), helping initiate a specific response. As shown in the diagram above, the dendritic cells take up the antigens and internalise them, then follow a chemokine gradient through the lymphatic vessels and migrate to the nearest lymph node.

Upon arrival, the mature dendritic cells present the processed adenoviral peptides on their MHC molecules to naïve T cells. This interaction, aided by co-stimulatory signals like interferon-gamma (this one will come up a lot in my writing), activates the T cells, triggering clonal expansion and differentiation into effector, specific T cells. The result is a targeted adaptive immune response: CD8+ cytotoxic T cells primed to destroy virally infected cells, and CD4+ helper T cells coordinating a broader immunological defense, including the activation of B cells to produce adenovirus-specific antibodies.

If you don’t have T cells, this doesn’t work. But if you take Anktiva before your dendritic cells receive the antigen, your immune system can kill the pathogen successfully. In turn, many cancers stem from the persistence of viruses that meddle with the p53 gene, like HPV. Elephants never get cancer because they have lots of copies of this gene (while we only have one), which expresses p53 proteins which tend to suppress tumors. This is just one example of how cancer is multidimensional chess, that we can play and win by turning immune system rejuvenation and modulation into an API query.

Obviously, this is where Hims Labs comes into the picture, but we’ve talked about that plenty lately.

In the background, two structural shifts are taking place:

1. Consumers have exponentially more information each day and thanks to AI, an ability to process it that’s rising in tandem - meaning, it’s harder every day to gate keep things.

2. The ability to synthesise biochemical structures at will turns health into a compute problem, in which all you have to do is figure out the right key-lock configurations.

I’ve discussed the above at length before many times. If you know exactly what biomarkers are going wrong today, you’re increasingly able to figure out what molecules you need to put inside your body to revert those biomarkers to an optimal state. As always, that’s not medical advice, but rather an observation of where this industry is headed. And as Immunity Bio founder Dr. Patrick Soon remarked during the Q4 2025 earnings call, the FDA knows and they’re changing regulations accordingly:

This concept of the plausible mechanism of action, which is the new policy put forth by Dr. Makary and published in the New England Journal of Medicine, speaks to a very exciting opportunity that the plausible mechanism of action is a pathway.

[…]

He says, “For instance, a single disease with 150 different genetic mutations with the same functional implications may require 150 different therapies and the plausible mechanism pathway would be ideally suited to such therapies.”

So, obviously, with 150 different mutations, 150 different therapies, one wouldn’t be expected to do 150 different trials.

He goes on to say and this is important and this is in the article by Commissioner Makary, “An appropriately designed study with a small sample size can support licensure of a product for which pharmacological effect is aligned with biological plausibility.”

In other words, every pathogen is going to present highly specific circumstances. But if jacking up the immune system and training it to kill the specific pathogen works across the board, it doesn’t make a lot of sense to run a trial for each specific pathogen. The pharmacological effect will tend to track the biological mechanisms on a first principles basis: pointing the platform at a new pathogen is largely a matter of swapping the target antigen, not reinventing the science. In other words, the upside of an immune system platform that goes from general restoration to specialised cytotoxic training is vast.

I continue to buy more $IBRX.

Until next time!

⚡ If you enjoyed the post, please feel free to share with friends, drop a like and leave me a comment.

You can also reach me at:

Twitter: @alc2022

LinkedIn: antoniolinaresc

Disclosure

These are opinions only of the individual author. The contents of this piece do not contain investment advice and the information provided is for educational purposes only and no discussions constitute an offer to sell or the solicitation of an offer to buy any securities of any company. All content is purely subjective and you should do your own due diligence.

Antonio Linares makes no representation, warranty or undertaking, express or implied, as to the accuracy, reliability, completeness or reasonableness of the information contained in the piece. Any assumptions, opinions and estimates expressed in the piece constitute judgments of the author as of the date thereof and are subject to change without notice. Any projections contained in the Information are based on a number of assumptions as to market conditions and there can be no guarantee that any projected outcomes will be achieved. Antonio Linares does not accept any liability for any direct, consequential or other loss arising from reliance on the contents of this presentation. Antonio Linares is not acting as your financial, legal, accounting, tax or other adviser or in any fiduciary capacity.

Longeveron’s Lomecel-B Phase I and IIa trials shed insight on one thing I’ve tested on myself: repairing the mitochondrial>bone marrow>thymus gland triad may be the most upstream lever we have against chronic illness, including Alzheimer’s.

Share

In my own experience, I’ve obtained wonderful results by upgrading my mitochondria, bone marrow and thymus gland. I’ve successfully defeated Lyme and Long Covid, and this “hack” seems to have brought my immune system fully back online for the first time in five years, tilting me back into full recovery. The downstream effects of working on this triad have been broad and deep, spanning improved mental health, performance, sleep and recovery, and have piqued my curiosity about its potential to act as a healthcare platform of sorts.

Longeveron’s Lomecel-B is a mesenchymal cell obtained from young bone marrows. In phase I and IIa trials it’s had the following effect on Alzheimer’s patients:

1. VEGF levels were preserved in treated patients while declining significantly in the placebo group, indicating maintained cerebrovascular integrity in the former.

2. Serum amyloid peptides (Aβ38, Aβ40, Aβ42) trended higher in treated arms, consistent with improved blood-brain barrier function and enhanced clearance of amyloid from the brain.

3. MRI showed both increased hippocampal volume in Phase I and a statistically significant slowing of hippocampal volume loss in Phase IIa, alongside a 49% reduction in overall brain volume loss.

4. And DTI imaging confirmed a durable reduction in neuroinflammation across the hippocampus and temporal lobe.

5. Qualitatively, the end result has been treated patients largely conserving their quality of life.

Lomecel-B is just a natural longevity-associated mesenchymal cell. The latter are connective tissue stem cells that can differentiate into bone, cartilage, fat, and muscle. Regarding the brain, what’s interesting about them is their secretome: the stuff they release which signals the surrounding environment to calm inflammation, stabilise blood vessels, and restore the conditions for the organ in question repair itself. Before we go into the secretome in depth, I’d like to first explore this concept: on a first principles basis, the results of Lomecel-B stem primarily from fixing the blood brain barrier.

The blood brain barrier (BBB) is the brain's border control, a highly selective membrane lining the brain's blood vessels that determines what gets in and what stays out. Structurally, it is formed by endothelial cells bound together by tight junctions, reinforced on the outside by pericytes and astrocyte end feet, creating a physical and biochemical wall between the bloodstream and brain tissue. Small lipid soluble molecules can slip through passively, but glucose, amino acids, large molecules, proteins, and pathogens are either actively transported via specialised carriers or blocked entirely.

This selectivity is not incidental: it’s what keeps the brain's chemical environment stable enough for neurons to function, protecting it from pathogens, toxins, and the inflammatory signals constantly circulating in the blood.

A degraded BBB is a feature of Alzheimer’s. Tight junctions loosen, pericyte coverage degrades, and the BBB becomes leaky, allowing pro-inflammatory molecules into the brain while simultaneously losing its ability to export waste products like amyloid. This is why VEGF preservation and rising serum amyloid levels in the Lomecel-B trials are meaningful: they are proxy signals that the BBB is being stabilised and that its clearance function, pumping amyloid out of the brain and into the blood, is being restored.

Now consider Lomecel-B’s secretome, which releases molecules to act on 4 wide buckets: the aforementioned vascular reparation, neuroprotection, modulating the immune system and flickering epigenetic switches (to turn on genes that drive repair and switch off the ones that don’t). You can see the secretome summarised in the table below.

Subscribe now

VEGF fixes the BBB. HGF reinforces this by reducing neuroinflammation and supporting repair of the BBB and surrounding tissue, while SDF-1 amplifies the overall repair response by recruiting the patient’s own endogenous stem and progenitor cells to sites of injury. PDGF acts more structurally, stabilising pericytes on vessel walls: the same pericyte degradation that, as established above, is a hallmark of BBB breakdown in Alzheimer’s disease.

The neuroprotective layer of the secretome operates in parallel. bFGF supports neural cell survival and promotes recovery from inflammatory damage. IGF-1 provides anti-apoptotic signalling, protecting neurons from death and speculatively supporting an environment conducive to neurogenesis. TGF-β1 acts as the secretome’s most potent immunosuppressive signal, down regulating pro-inflammatory pathways and driving the overall anti-inflammatory shift observed in the trial biomarker data.

Immune modulation is achieved primarily through three cytokines, which Immunity Bio investors will be familiarised with:

• IL-4 shifts macrophages from a pro-inflammatory (M1) to an anti-inflammatory (M2) mode, directly contributing to BBB stabilisation and vascular repair

• IL-10 suppresses chronic neuroinflammation: the low-grade, persistent brain inflammation increasingly recognised as a central driver of Alzheimer’s progression

• IL-6 presents a more complex signal, with its elevation at high dose likely reflecting productive acute immune activation rather than pathological inflammation.

Finally, the secretome operates at an epigenetic level (switching genes on and off) through miRNAs and extracellular vesicles. miRNAs regulate inflammation and repair gene expression in recipient brain cells, modulating behaviour remotely rather than through direct cell contact. Extracellular vesicles and exosomes (a little of corner of biology which I think will also be extremely relevant) serve as the delivery vehicle for this entire cargo: carrying proteins, lipids, and miRNAs across biological barriers including, critically, the BBB itself, reaching brain tissue even when whole cells cannot.

To put it bluntly, if your exosomal system is poisoned, the whole value chain stops working because your brain doesn’t receive the necessary raw material to fully repair. I highlight this because I believe this will also become a platform in itself, with many exosomal systems poisoned worldwide at present. Further, the exosomal system is bidirectional, meaning the brain uses it to both send and receive signals. A compromised exosomal system therefore impairs both the delivery of repair signals inward and the early warning system outward. That double failure may be part of why Alzheimer’s progresses silently for so long before symptoms emerge.

Just like exosomes carry around useful proteins, they can also carry out toxic ones.

If you know, you know.

Moving on, Lomecel-B is acting at the local level by inserting cells that release molecules which act as signals to the local environment. But there’s plenty of reasons to believe that a condition such as Alzheimer’s stems from a deeper causal chain than just blood brain barrier degradation. The following BBB degradation pathways have been identified in research:

1. ApoE4 genotype: activates pericyte destruction cascade.

2. MMP9 enzyme activity: physically cleaves the tight junctions apart.

3. Oxidative stress and ageing: kills pericytes and erodes endothelial integrity.

4. RAGE receptor up-regulation: pulls amyloid into the brain, accelerating its own conditions.

5. Oligomeric tau: deforms endothelial cells and distorts tight junctions.

6. Peripheral macrophage infiltration: once in, amplifies inflammation and accelerates everything above.

Not to my surprise, the root of all the above is cellular energy failure, with mitochondrial disrupting being the upstream event. Everything flows from mitochondria (the energy production engines that power your cells) being damaged over time, be it by wear and tear and/or toxins. When mitochondria in endothelial cells (essential building blocks of the vascular system) and pericytes lose membrane potential and electron transport chain efficiency, ATP production falls. The consequences cascade in sequence:

1. Reduced ATP gives all cells in the body less energy, impairing function across the board.

2. Mitochondrial reactive-oxygen species (ROS) overproduction.

   1. This drives oxidative stress which damages endothelial cell membranes directly, activates TNF-α, initiates the pericyte apoptosis (death) cascade.

3. Impaired mitophagy (failure to remove damaged mitochondria).

   1. Damaged mitochondria accumulate rather than being cleared, amplifying ROS output in a self-reinforcing loop.

4. Reduced NAD+, which affords mitochondria less fuel to work properly.

   1. Sirtuin-mediated DNA repair fails, accelerating cellular senescence (ageing) in pericytes and endothelial cells, driving further degradation of the BBB.

5. Calcium excess builds up in mitochondria, activating cell-death enzymes, causing blood vessel support cells and endothelial cells to die.

   1. Further driving pericyte apoptosis and endothelial cell death.

I learned about the above because at first, I was addressing the immune dysfunction I experienced with Lyme and Long Covid at the thymus gland level. Results were appealing, but not definitive. While maintaining treatment at that level, as well as clearing the pathogens themselves, regenerating the mitochondria and bone marrow is what truly allowed me to recover. Once your mitochondria produce enough energy, the rest of the body tends to fix itself, because all cells have enough energy to do their job properly. One striking property of focusing on this causal chain for me has been a radically improved ability to focus, store information and manage emotions, as well as perform physically in sports.

For example, I’m capable of playing tennis now at speed which I couldn’t previously process. Some balls were too fast for me previously and now they’re not, which I suspect is due to my nervous system having access to more energy.

Once cells have enough energy and the bone marrow can do its job properly, the thymus has a good shot of coming back online. All blood and immune cells originate from the blood marrow, which then differentiate into two lineages:

1. Myeloid.

2. Lymphoid.

In the graph below, you will how in fact a good number of immune cell types are both born and matured within the bone marrow, like natural killer and B cells, which makes the bone marrow a primary immune organ by itself. This is also the case for mast cells, whose primary job is to act as “alarm bells” in barrier tissues such as the skin. Mast Cell Activation Syndrome (MCAS), which I personally experienced over the past two years as a result of Lyme and Long Covid, is the result of these two cells going crazy and releasing histamine all the time. T cell precursors are born inside the bone marrow, but then are differentiated and mature inside the thymus gland.

If you are wondering what the practical application of the above knowledge may be, beyond continuing to understand Alzheimer’s, consider Immunity Bio’s Anktiva. It’s basically an IL-15 agonist, which means it binds onto that receptor and signals the body to produce natural killer and T cells. To maximise the odds of Anktiva succeeding, you need healthy mitochondria, a productive bone marrow and functional thymus to ultimately replenish the immune system.

Mitochondrial dysfunction, as detailed above, does not only degrade the blood brain barrier: it degrades the entire cellular energy infrastructure that immune function depends upon. NK cells and T cells are among the most metabolically demanding cells in the body: their activation, proliferation, and cytotoxic function (ability to kill bad guys) require sustained ATP output and tight ROS regulation.

A mitochondrially compromised immune cell responds weakly to IL-15 stimulation regardless of how well Anktiva binds its receptor. Fixing the mitochondrial layer therefore directly amplifies the signal Anktiva is trying to send. Anktiva sometimes doesn’t work if a patient’s health is overly worn out and I suspect that dialling in on this triad will tend to drastically improve outcomes.

Below you can see Dr. Patrick Soon’s remarks about my Immunity Bio video. He’s the founder and CEO of ImmunityBio, as well as a rock-star-billonaire oncologist. He happens to believe as do I, that immune system rejuvenation is a critical platfom for human well being going forward.

The other practical implication of this knowledge is understanding how important Hims Labs is. You can track the state of the mitochondrial > bone marrow > thymus gland triad with the following biomarkers over time (this is a non-exhaustive, non-prescriptive list which may change over time as our understanding of these biological processes evolves):

Mitochondrial health:

1. NAD⁺ / NADH ratio: direct measure of mitochondrial fuel availability; declining ratio indicates energy production failure.

2. Lactate / pyruvate ratio: levated ratio signals a shift from mitochondrial to anaerobic metabolism, a hallmark of dysfunction.

3. 8-OHdG (8-hydroxy-2-deoxyguanosine): urinary or serum marker of oxidative DNA damage driven by ROS overproduction.

4. CoQ10 (ubiquinol/ubiquinone ratio): reflects electron transport chain efficiency and antioxidant reserve.

5. mtDNA copy number: low copy number in peripheral blood indicates mitochondrial depletion at the cellular level.

Bone marrow output:

1. CBC with differential: baseline measure of red blood cell, white blood cell, and platelet production reflecting marrow productivity.

2. Haematopoietic progenitor cells (HPCs) in peripheral blood: direct count of CD34⁺ cells circulating from marrow; declines with age and marrow dysfunction.

3. Erythropoietin (EPO): elevated EPO with low RBC output signals marrow’s failure to respond to demand.

4. Ferritin and transferrin saturation: iron utilisation markers that reflect the efficiency of marrow-driven erythropoiesis.

Thymus gland output:

1. sjTREC (signal joint T cell receptor excision circles): the gold standard measure of thymic output; directly quantifies newly produced naïve T cells entering circulation.

2. CD4⁺ naïve T cell count: declining naïve CD4⁺ population is the functional consequence of reduced thymic throughput.

3. CD8⁺ naïve / memory ratio: a falling ratio indicates the T cell pool is ageing, with fewer newly minted cells replacing exhausted ones.

4. IL-7 levels: elevated IL-7 is a compensatory signal the body produces when thymic output is insufficient; paradoxically, high IL-7 indicates thymic insufficiency.

By tracking these over a patient’s life, you get a pretty good idea of what the triad is doing. And on a first principles basis, keeping these biomarkers in good state necessary lowers the odds of chronic disease such as Alzheimer’s popping up, let alone unexpectedly. You can optimise this biomarkers (at least in a research setting, I’m obliged to remark) by using certain peptides. For example, SS-31 reverts mitochondrial permeability, which keeps electrons from sizzling out and optimises ATP production, improving overall cellular energy levels. Short sequence organ-specific peptides (known as bioregulators) have been shown to rejuvenate specific organs, like the bone marrow and thymus gland.

The peptideosphere is fully of molecules that can be used to regenerate the triad. It’s obviously up to licensed physician to say what a patient should do or not, but I note that the combination of triad biomarkers plus peptides is the basis for an entirely new medicine format that considerably decreases the prevalence of debilitating chronic illnesses and optimises lifespan. When investors focus on GLP-1s in Hims’s case, the above is just one example of what they fail to see. The body is full of synergistic causal chains that upon degradation can ruin our life and conversely, can be tracked and optimised over the time to keep us healthy.

Until next time!

⚡ If you enjoyed the post, please feel free to share with friends, drop a like and leave me a comment.

You can also reach me at:

Twitter: @alc2022

LinkedIn: antoniolinaresc

Disclosure

These are opinions only of the individual author. The contents of this piece do not contain investment advice and the information provided is for educational purposes only and no discussions constitute an offer to sell or the solicitation of an offer to buy any securities of any company. All content is purely subjective and you should do your own due diligence.

Antonio Linares makes no representation, warranty or undertaking, express or implied, as to the accuracy, reliability, completeness or reasonableness of the information contained in the piece. Any assumptions, opinions and estimates expressed in the piece constitute judgments of the author as of the date thereof and are subject to change without notice. Any projections contained in the Information are based on a number of assumptions as to market conditions and there can be no guarantee that any projected outcomes will be achieved. Antonio Linares does not accept any liability for any direct, consequential or other loss arising from reliance on the contents of this presentation. Antonio Linares is not acting as your financial, legal, accounting, tax or other adviser or in any fiduciary capacity.

Share

This is an update of my original Hims deep dive and Hims Q4 2025 ER Digest.

Grail's test, called Galleri, in combination with Hims' network is a great example of how longitudinal healthcare data is positioned to shift unit economics in the healthcare industry. Indeed, the aforementioned combination promises to shift the cancer business away from reactively treating late stage cancers to preventing them. A few weeks ago the market was spooked by the headline of Galleri's latest NHS trial, but a deeper look reveals cancer unit economics are about to change.

Galleri detects cancer signals early by analysing cell-free DNA (cfDNA) fragments shed into the blood from dying cells, including tiny amounts from asymptomatic tumors. It targets aberrant methylation patterns (epigenetic marks on DNA that cancer cells alter distinctly from healthy cells) across over a million informative sites via targeted sequencing. Machine learning classifies these patterns to identify a shared cancer “fingerprint” across 50+ types and predicts the likely tissue origin, enabling detection before symptoms or routine screens catch many aggressive cancers.

On a first principles basis, there’s no way around the science. If you catch and analyse the cfDNA, you’re going to know what’s going on.

Grail stock is down ~50% because the recently published trial with NHS didn’t meet the primary end point: a significant reduction of Stage III and IV cancers. In the recent Q4 2025 earnings call, analysts were visibly concerned with the results and kept on asking the same questions. However, the trial also revealed findings which would absolutely delight consumers:

1. Galleri led to a substantial and clinically meaningful drop in Stage IV (metastatic, often incurable) diagnoses compared to standard care alone, with further 20% reductions in second and third screening rounds.

2. This was especially true for the 12 cancer types that account for 2/3 of all cancer mortality.

3. Overall cancer detection rate was 4X higher compared to standard screening.

As a consumer, I don’t care about the study’s primary “end point”: a reduction in stage IV cancer is already an appealing value proposal. In the trial Stage III incidence was higher than expected, leading Galleri to miss the end point. But this was likely because they were catching cancers that would’ve progressed to stage IV earlier, while also boosting detections at Stage I and II.

This is not to say that Grail’s management shouldn’t strive to reduce both Stage III and IV, but are already quite valuable to consumers considering the deadly prognosis of Stage IV.

Grail’s revenue is up quite a bit over the past few years and what’s fascinating is it comes mostly from self pay, as the test is not widely covered by health insurance plans (including most major insurers and Medicare at this time). In the US for example, Galleri is marketed as a prescription-only test available available under CLIA (Clinical Laboratory Improvement Amendments) certification as a laboratory-developed test (LDT). Regulators go back and forth, but consumers are voting with their dollars.

Subscribe now

Grail completed its Premarket Approval (PMA) submission to the FDA for Galleri on January 29, 2026. There's no telling what the FDA will do, but the following is clear:

1. The test is more effective as it collects more longitudinal data, per the higher Stage IV reductions with subsequent readings in the NHS trial.

2. Embedded within a more exhaustive biomarker ecosystem, the test could be even more powerful.

Hims Labs gets you 130+ biomarkers today, twice a year. Correlating the Galleri test with those biomarkers will help customers prevent cancers, by nudging key biomarkers in the right direction over time. Perhaps the network (proprietary data + AI) reaches the conclusion that certain biomarkers going wrong have an outsized impact on the odds of a specific tumor showing up later on, for example. By addressing those biomarkers, the odds of that tumor arising go down. And if it does arise, Galleri helps catch it early.

This is the value chain most can’t see.

The healthcare system is going from reactionary to preventative merely by virtue of data. And while the industry quarrels over drugs, this shift is subtle and unstoppable. Value in the healthcare industry will accrue to the platform that gets more and better biomarkers on the most amount of customers over time. This will allow said platform to deliver more value per token than anyone else in the industry and come to hold all the intelligence required to develop and distribute new treatments.

This is a matter of incentives. Only one player makes more money by delivering better patient outcomes per dollar.

Until next time!

⚡ If you enjoyed the post, please feel free to share with friends, drop a like and leave me a comment.

You can also reach me at:

Twitter: @alc2022

LinkedIn: antoniolinaresc

Disclosure

These are opinions only of the individual author. The contents of this piece do not contain investment advice and the information provided is for educational purposes only and no discussions constitute an offer to sell or the solicitation of an offer to buy any securities of any company. All content is purely subjective and you should do your own due diligence.

Antonio Linares makes no representation, warranty or undertaking, express or implied, as to the accuracy, reliability, completeness or reasonableness of the information contained in the piece. Any assumptions, opinions and estimates expressed in the piece constitute judgments of the author as of the date thereof and are subject to change without notice. Any projections contained in the Information are based on a number of assumptions as to market conditions and there can be no guarantee that any projected outcomes will be achieved. Antonio Linares does not accept any liability for any direct, consequential or other loss arising from reliance on the contents of this presentation. Antonio Linares is not acting as your financial, legal, accounting, tax or other adviser or in any fiduciary capacity.

Duolingo has no competition and is on its way to becoming the dominant, personal AI tutor.

Share

Largely seen as a silly bird app, I’ve been anticipating Duolingo’s rise as the world’s personal AI tutor. Seeing management voluntarily sacrifice short term financial performance to materialise this vision is a major thesis milestone. Duolingo’s A/B testing machine has proven itself since IPO, via the results you see below. The machine is now fully focused on bringing superhuman teaching at scale by leveraging AI. The odds of Duolingo bringing a superhuman AI teacher to everyone on Earth are now the highest they have ever been, and so are the odds of exponential free cash flow per share growth.

If Duolingo had 100X potential at all time highs, it now has 600X potential.

The market tends to get upset when founder operators focus on long term value creation. Historic examples include Bezos launching AWS, Reed Hastings killing the DVD business, and most recently, Daniel Ek investing in podcasts and Zuckerberg exponentially increasing AI CapEx before the world was aware of the AI revolution. I believe this instance is no different. As AI capabilities continue to rise exponentially, Luis is focusing all resources on seizing the personal tutor opportunity, as he explained during the Q4 2025 earnings call.

As a long term investor, this is exactly what I want to hear:

In terms of AI, look, we really are in a unique time in history.

AI is going to allow us to teach significantly better.

And I also believe that the way people learn is just going to change. And the reason for that is, I think that in the next few years, we are going to see quality of teaching that is as good as a one-on-one tutor.

[…]

And that just made it so that I have even more conviction that we need to spend really the vast majority of our efforts on DAU growth. And as we built the plan for 2026, this is what came out. I mean the guidance that we’re giving.

As Jensen explained yesterday, AI agents sit on top of software. To create the best agent (personal tutor in this case), you need a platform that gets you more and better data than anyone else. To that effect, it helps that Duolingo has no competition, as Luis highlighted during the call:

The reason Duolingo so far has not really gotten competition at a large scale, again, we have like 85% of the daily active users of language learning apps in the world is because it’s pretty hard to compete with our free product in the sense that we have really huge distribution.

Aforementioned instances of founder operators nuking short term financial performance have one thing in common: their respective platforms started with one vertical and then used the resulting distribution advantage to deploy additional verticals, which then yielded dominant platforms. Amazon started with books, Uber started with black town cars and Spotify started with music. Duolingo MAUs are down QoQ, but:

1. DAUs/MAUs continue to go up in a straight line.

2. The new chess vertical hit 7M MAUs in Q4 2025, just 7 months after launch.

Subscribe now

According to the narrative, humans should have no reason to want to play chess since AI is so good at it and let alone on Duolingo. But the vast majority of chess users are existing Duolingo users. This reveals both their satisfaction with the platform and Duolingo’s ability to deploy and mature new verticals faster as AI gets better. Duolingo is now the internet’s second biggest “chess platform”, as Luis highlighted during the call. This directionally illustrates the fate of Duolingo’s subsequent verticals:

So in particular, chess has grown quite a bit. As I said in our shareholder letter, we have about 7 million daily active users in chess and this is less than a year after we launched.

At this point, we believe we’re the second largest chess thing in the world, which is, by the way, incredible because if you search for chess on the App Store, we don’t appear in the results.

We just haven’t even gotten there to doing App Store optimisation, but we already got to 7 million daily active users.

The chess vertical was instantiated by two Duolingo employees with no chess and/or coding skills, largely enabled by AI. The evolution of the chess vertical outweighs my concerns about the QoQ MAU decline and makes me take Luis’ remarks about maths seriously. We are a few log steps away from Duolingo being able to click entirely new scaled verticals into existence:

And the same is true for our other subjects, certainly with math. Like I said in our shareholder letter, I believe that this year, we’re going to have the best tutor app for math.

Much of the financial weakness management is guiding comes from two things:

1. Sharing economies of scale with customers, in the spirit of Costco, by making advanced AI features available to lower level subscription tiers.

2. Giving themselves room to experiment towards unlocking the AI personal tutor.

Monetisation stems from friction, motivating people to pay to eliminate it. Friction can be introduced so long as the value to price ratio is wide enough, but the drag on long term value creation is high when the reward for distribution is exponential. Every marginal ad weighs on the proprietary data moat, which pushes the AI tutor further out in time. Duolingo has had a good run monetising the value to price ratio they have built over the past decade, but continuing to prioritise monetisation would be a disservice to investors.

One of the AI features they’re moving downstream is Video Call with Lily. During the Q4 2025 earnings call it was clear how management prioritises words spoken per video call, a metric initially highlighted by community member Lluis Fondevila, over Max subscription monetisation. AI video calls are expensive, but inference costs are coming down. Here management had two choices:

1. Leverage declining inference costs to increase Max margins.

2. Make video calls more widely available to deliver more value to users, at the expense of short term financial performance.

As we saw in my last Duolingo post, more words spoken per video call directly translates into better customer outcomes. Luis commented on this during the earnings call:

In terms of video call improvements, I’m very happy with the results for video call improvements. Some -- basically, if you look at the graph of words spoken per user on video call, that is just a monotonically increasing graph over time.

It just keeps getting better and better and better. And basically, we’re just a lot better at getting Lily that you’re talking to, to elicit more conversation. Right now, we’re trying a new thing, we’re actually starting to tell you, hey, in your next response, ask a question or in your next response, use the word because or something.

And that actually really gets people going more and more.

This management team is not focused on the market, but on maximising value delivered per token. The earnings call is littered with examples of Duolingo management shedding weight to accelerate the below flywheel. The market hasn’t yet understood it, per the recent performance of software stocks. But it will be forced to eventually, since proprietary data is the moat of the 21st century.

Duolingo management understands that second best is worth zero in the agentic economy and are thus maximally prioritising customer outcomes to drive more adoption, word of mouth, and proprietary data, and thus value delivered per token. Everything from basically abandoning ads to doing whatever it takes for people to talk to Lily and say more words. As previously mentioned, Duolingo has no competition, so they have blue skies ahead to take the whole market.

The future is binary and the best ontology will take it all.

About investing in general, the average investor tends to be psychologically better off with dynamics that are not optimal for long term wealth creation. With my entry point at $323, the average investor would have seen a steady 10% increase in the stock with management remaining focused on monetisation as a success. However, if my thesis is correct, I will make much more money having doubled my position at $190. And even more so if I continue to buy heavily as the stock continues to drop and management relentlessly focuses on customers.

Risk is not knowing what you own.

Until next time!

⚡ If you enjoyed the post, please feel free to share with friends, drop a like and leave me a comment.

You can also reach me at:

Twitter: @alc2022

LinkedIn: antoniolinaresc

Disclosure

These are opinions only of the individual author. The contents of this piece do not contain investment advice and the information provided is for educational purposes only and no discussions constitute an offer to sell or the solicitation of an offer to buy any securities of any company. All content is purely subjective and you should do your own due diligence.

Antonio Linares makes no representation, warranty or undertaking, express or implied, as to the accuracy, reliability, completeness or reasonableness of the information contained in the piece. Any assumptions, opinions and estimates expressed in the piece constitute judgments of the author as of the date thereof and are subject to change without notice. Any projections contained in the Information are based on a number of assumptions as to market conditions and there can be no guarantee that any projected outcomes will be achieved. Antonio Linares does not accept any liability for any direct, consequential or other loss arising from reliance on the contents of this presentation. Antonio Linares is not acting as your financial, legal, accounting, tax or other adviser or in any fiduciary capacity.