Pharma Is Mis-Pricing Its AI Bet

Learning is not compulsory; it’s voluntary. Improvement is not compulsory; it’s voluntary. But to survive, we must learn.

— W. Edwards Deming

In December 2016, Pfizer and IBM announced that Watson would be turned loose on immuno-oncology drug discovery. The handshake photograph ran in the trade press; the press release named the executive sponsor and the disease. By April 2019 IBM had quietly ended new sales of Watson for Drug Discovery, no named clinical asset had emerged, and the next handshake had taken its place. The contracts were not negligent. The science was real. What was missing was a pricing discipline severe enough to ask, before the announcement, what the firm would have to learn for the partnership to be worth what it cost.

Today the same arc is being run again at ten times the dollar values, on three sides of the same table. Pharma writes ten-figure checks to frontier-model labs (Lilly–Isomorphic, Sanofi–OpenAI). Pharma writes ten-figure checks to AI-bio platforms (Bayer–Recursion, AstraZeneca–Absci). And venture capital writes ten-figure checks into the AI-bio companies themselves (Xaira, Isomorphic). Biopharma has no AI-free future, and the AI of this decade will reshape it more deeply than any wave before. That is precisely why the mispricing matters: the prize is real, and the firms paying for it without earning the learning underneath will not be the ones who collect. Across all three pools, almost none of the bets carry a named therapeutic thesis, a publicly disclosed prewritten exit, or a capitalized build-side. Pharma is not underinvesting in AI. It is mis-pricing AI, and so is its counterparty across the table. Pricing, here, is not valuation. It is the full allocation package behind a bet — capital, data access, decision rights, organizational attention, exit discipline — and the build-side underneath the visible bet that decides whether the package compounds or dissolves. A bet is mis-priced whenever any of those layers is missing, and the most common mismatch now is Commit-grade headlines with no capitalized build-side at all.

This pricing failure has a name. Drift is capital that survives because no one has written down what would make them stop. A bet without a kill condition is not an investment; it is an operating expense in slow motion. Drift is the partnership that renews because nobody wants to kill it; the platform with users but no changed decision; the build-side that was never funded because it never appeared on the slide. We are in inning one of the biology–AI cycle, and its disease is drift dressed as strategy. Almost-good AI is what drift produces: outputs that demo well on curated public data, fail on the firm’s own messy, partially-labeled, GxP-controlled corpus, and rarely change a decision that matters.

What counts as durable advantage in inning one is not the model layer. It is the build-side underneath it. The substrate separates a compounding bet from drift. What separates a real position from a posture is whether that substrate has been capitalized: built inside the firm, acquired, funded at the counterparty under terms that return durable value to the buyer (audit rights on the evaluation harness, co-ownership of fine-tuned weights, IP on output molecules), or owned outright by being the company that is the substrate. Most current pharma AI deployments rent compute and frontier models and stop there, as though the rented capability were itself the bet. The logic is a depreciation schedule, not a slogan: the model layer commoditizes with every frontier release, while a substrate of proprietary data calibrated to the firm’s own biology — and an evaluation harness that compounds with it — appreciates. Rent the part that depreciates; capitalize the part that appreciates. Pricing the build-side is the spine of every honest pharma AI conversation this decade.

The build-side has two layers. A substrate of decision-grade data and evaluation that the firm owns. And an operational layer above it that turns the substrate into decisions the firm acts on.

Two columns make up the substrate. Decision-grade data: proprietary multi-modal longitudinal data, disease biology, genetics, perturbational screens, chemistry, translational biomarkers, patient stratification, clinical response, real-world evidence, GxP-controlled and labeled to a standard the firm’s own decision-making, clinical, regulatory, and commercial, will stand behind. Evaluation harness: infrastructure calibrated to the firm’s own biology and chemistry rather than to public ML leaderboards, asking whether the molecule survives translational filters and whether the biomarker predicts clinical response under the firm’s own data.

Three columns make up the operational layer above the substrate. Decision pipeline and action authority: the apparatus that gets a model output in front of the program leader who can act on it, and the rights for that person to act. Governance and learning loop: the discipline that turns each output into the next sharper bet rather than the next slide, including the re-calibration cadence as biology and assays drift. Adoption surface: the organizational reality on which any of the above either lands or dies. Moderna deploying seven hundred and fifty custom GPTs across the firm within two months of rolling out ChatGPT Enterprise is what an adoption surface looks like; it is necessary evidence, not sufficient. The harder evidence sits at the substrate, where an AI output bends a target, a biomarker, or a Phase 2 design a regulator will see. The build-side is not data ops dressed as AI; it is the apparatus through which any partnership produces a learning loop instead of a subscription.

Biology resists AI more than chemistry, code, or language do, and the build-side has to be sized accordingly. Frontier-model performance on a public benchmark is information about the model, not about whether it will bend a drug-development decision a regulator will see.

None of these line items look like AI on a CFO’s slide. A CFO can sign a five-hundred-million-dollar GPU order and book it as such. A build the substrate so the partnership produces a scientific decision line item is harder, because the substrate is distributed across data ops, MLOps, translational science, governance, and organizational redesign. That dispersion is also why almost no firm discloses what it spends on it. The heuristic does not come from any firm’s disclosure; there is none. It comes from the one place the ratio is actually measured. In general machine-learning practice the model is the cheap part: “only a small fraction of real-world ML systems is composed of the ML code” (Sculley et al., 2015); the surrounding data and infrastructure are the bulk. Biology is the extreme of that case. A decision-grade biological label takes a clinician’s or a PhD’s judgment, slow and hard to crowdsource: one benchmark study turned to crowd workers because expert radiologists needed months to label a few hundred report sentences (Cocos et al., 2015). Add GxP provenance, small-n, and the in-vitro-to-human distribution shift, and the build-side under a model partnership is no rounding error on the headline; it is the larger number, and in biology larger still. Backed out from what survives a serious R&D review, that puts it at no less than the headline deal value and often a multiple of it, over three to five years. The exact multiple matters less than two facts a board can check: most current pharma AI partnerships are funded at a fraction of it, and on most of them there is no disclosed build-side line at all. The absence of the line is the tell.

The industry response sorts into three groups, not two. At least two firms are on long, stepwise journeys in the cohort, and they ran them differently. AstraZeneca built its Cambridge-based AI program into a decade-long sequence: BenevolentAI in April 2019, Schrödinger later that year, Absci in late 2023, Modella AI in early 2026, alongside a parallel internal substrate now visible from the publication trail. Roche bought it: Flatiron and Foundation Medicine in 2018, Aviv Regev to gRED in 2020, Prescient Design in 2021, the PathAI definitive merger in 2026. It is the most expensive forward-carry in the industry, against the deepest data position in oncology. A second group, the bilateral-pact tier, runs from Sanofi to Novartis to BMS, signing frontier-lab partnerships at scale without yet making the internal substrate visible at the same scale. Lilly sits adjacent to this tier with greater execution intensity; Dave Ricks has said publicly that he runs one or two AIs in every meeting. The substrate-side disclosure, though, is still catching up to the deal-side disclosure. A third group, most of the longer tail, has a thin or quiet AI public footprint, which usually means buying black-box capability top-down without a substrate underneath. The decade-defining split sorts more reliably on whether the firm is capitalizing a substrate of its own than on whether it has signed the right partnership.

Behind that split sit four traps that keep the learning from happening, and each currently looks like prudence in the room. They share an underlying error: paying for AI without earning the learning underneath it.

The first is to subscribe without capitalizing the substrate. A firm signs a foundation-model partnership, licenses access to a generative-chemistry platform, buys seats on a tenanted version of a frontier model, and treats that as the AI strategy. Some subscription is unavoidable and rational; frontier capability rents to everyone, and open-weight models lower the rent further. The trap is the posture that treats the rented capability as the strategy, with no substrate capitalized underneath it, whether built, acquired, or capitalized at the counterparty under terms that return value to the buyer. Subscribe-and-capitalize compounds. Subscribe-without-capitalize delivers almost-good AI: outputs that win on curated public data and die on the firm’s own GxP corpus. Most pharma AI deployments that disappointed in 2024–2025 disappointed at exactly this seam. The harder variant of the trap is the pilot that posts a measurable ROI on its own narrow scope while the build-side underneath quietly goes unfunded; pilots run eighteen to twenty-four months, and by the time the firm books the win, the world the pilot was scoped against is two model generations gone.

The second is to call the authority. The AI labs will take the meeting; their native expertise is not IND execution or payer contracting. The internal experts have the reverse problem: deep domain, no model. Either way, importing authority through a deal or a board seat is a 24-month bet. And the underlying difficulty is harder than importing. Frontier AI talent is largely not applying to pharma in the first place, on pay and on perceived research ambition. The question that sorts a legitimate import from the trap is whether the firm is running a parallel internal bet that the import is meant to calibrate, not replace, and whether the build-side is being capitalized at the firm or only at the counterparty. Where neither is true, the import is the trap.

The third is to wait. Waiting sounds prudent in the room and is arithmetic on the patent cliff. Keytruda (a thirty-billion-dollar revenue line under threat from the 2028 IV patent cliff, defended in 2025 by FDA approval of the subcutaneous Qlex formulation, a contained development spend against an outsized revenue pool) is the case every Commit is priced against. Beyond Keytruda, roughly three hundred billion dollars of branded pharma revenue is at risk between now and 2030. Big-pharma R&D returns have collapsed since 2010; the recent recovery is GLP-1-driven and thin underneath. That is the gravity well. And the wait has a second cost the patent-cliff arithmetic does not surface. AstraZeneca’s first AI partnership ran in 2019; Roche’s first AI-relevant acquisition closed in 2018. By 2026 the firms that started have six-to-eight years of compounded build-side. That is a positional gap the next deal cycle cannot close: the build-side a latecomer would buy has already been priced by whoever started earlier. Three clocks run during a wait. Model capability moves every six to nine months. Build-side compounds for whoever started. The only thing that stays still is the firm itself. OpenAI’s own B2B Signals data, released in May 2026, makes the asymmetry measurable: frontier firms now use three and a half times more AI per worker than typical firms, up from two times only thirteen months earlier, and roughly two-thirds of the widening is depth of use rather than volume. The build-side gap is reading itself back through token spend. Wait and see is itself a Commit, by omission, to a timeline somebody else is setting. The disciplined version of waiting runs cheap experiments to update the prior, rather than sitting still.

The fourth looks the most like a strategy and is the hardest to see. It is to partner under conditions that look like access and are actually capture. The pattern is documentary, not speculative: a pharma agrees to provide proprietary data to develop the lab’s AI models; the lab’s evaluation harnesses, prompt patterns, and operational know-how compound across all of its partners; the pharma’s data, once shaped into the lab’s pipeline, does not come back portable. By the time the lab raises its next round at a multiple of its prior valuation, the pharma is the price-taker on a contract written when it was the price-setter. The structure that produces capture is legible on the face of several 2024 deal announcements: who provides the data, who keeps the evaluation harness, and whose valuation steps up on the next round. This is a reading of contract architecture, not of intent. Capture is the inning-one trap that pays for itself in inning two.

The four traps share a tell. A list of initiatives standing in for a portfolio. Twelve names, no disproportionate Commit, no public Kill, no state assignment, no build-side line on any of them, and none of it pressed on the second-order questions, because the vocabulary to ask them is not yet standard on a board pack.

What I do not know is whether the build-side denominator survives the next frontier-model jump or only the current one. The structural claim — that the locus of durable advantage moves down-stack with each capability wave, and that the substrate is where the compounding lives — depends on the frontier flattening fast enough that build-side investments compound rather than getting repeatedly obsoleted. If the next jump instead collapses adjacency to the model layer, the unit of analysis shifts again and today’s build-side denominator survives only as the previous generation’s right answer.

The structural argument survives either way: the firms that capitalize the substrate keep learning faster than the firms that rent it. The size of the prize and the identity of the durable winners do not. The alternative to drift, in either world, is a discipline. Compounding judgment is not produced by spending more carefully; it is produced by holding the portfolio against one question and answering it the same way each time. The question is downstream of a goalpost most AI conversations never name. Pharma exists to produce new interventions (therapies today, more complex regimens across diagnostics and prevention as the lines blur) that change patient outcomes. Pipeline value — better targets, better molecules, higher probability of success, more durable assets at the other end — is the goal stated in operating terms. Every dollar of AI spend either compounds toward pipeline value or it does not. The operating system that produces compounding judgment, and downstream of it pipeline value, has four disciplines.

Map the spend.

The firm cannot price what it cannot see. The map is where dollars buying tools separate from dollars building capability, and where the build-side under each visible bet either appears or is exposed as missing. Many leadership teams cannot produce that slide. Without it, every downstream decision is made blind.

Price the state.

Every material AI bet sits in one of four states. Kill: stop decisively and free the capital. Commit: a visibly privileged bet against a named thesis, with disproportionate capital — measured against the headline plus its build-side, not the headline alone — decision rights, an internal adoption pathway, and a prewritten exit. Constrain: hypothesis-driven continuation at limited capital, exit criteria written down. Drift: spend that continues because no kill condition was ever written down. The kill condition belongs in the original allocation memo: three to five named, observable signals, with the date the board will revisit them. The discipline is not avoiding drift; it is pricing drift as drift, in public, so the capital can move. Almost no AI bet in pharma today meets all of the Commit conditions.

Most AI investment decisions in biopharma are not risk decisions. They are uncertainty decisions: the distribution of outcomes is itself unknown, which is why the usual ROI machinery fails against them. A wrong Commit at scale is costlier than several drifting Constrains, because capital and option value die together when the Commit fails. But a portfolio of Constrains that never converts to Commit is its own failure mode: capital bleeds out in fractions, organizational attention disperses, and the external clock keeps running. The discipline is to write the Constrain-to-Commit conversion test at the same time as the Constrain itself, so the portfolio either resolves into named Commits or releases the capital on a schedule the firm controls.

Test the layer where value actually breaks.

Any AI capability lives on four layers. Model and data: what proprietary learning does this produce, and is the data side capitalized as the build-side that compounds? The model side is the most commoditizing of the four. The data side is the substrate. Drug-development reality: did AI change target selection, patient stratification, biomarker strategy, dose finding, Phase II design, the probability of technical and regulatory success, or the interpretation of real-world outcomes? Each is a kill point in disguise. The honest test asks not whether AI made the workflow faster but whether it bent one of those decisions, with evidence, on a program a regulator will see. Business-model fit: who captures the economics, the firm or the counterparty’s flywheel. A question that gets harder when the unit of intervention is no longer a single molecule but a composite the firm does not yet price as one. Decision authority: who can act on the output, and is that person where the output lands?

A model can be right, a molecule can still fail, the business model can still be wrong, and the organization can still be unable to act. Most AI conversations collapse into the model side of layer one. The valuable ones happen at the data side of layer one and at layers two through four. The first AI-designed drug into the clinic, Exscientia and Sumitomo’s DSP-1181, was discontinued in 2022 after Phase 1 failed to meet the expected efficacy bar in OCD; BenevolentAI’s BEN-2293 missed its Phase 2a efficacy bar in atopic dermatitis in April 2023. The wider cohort tells the harder story: in the small but published AI-discovered cohort tracked by Jayatunga and colleagues for BCG in Drug Discovery Today (2024), molecules show roughly eighty to ninety percent Phase 1 success and then collapse to industry baseline at Phase 2. AI is now reliably shortening discovery-to-IND and beginning to compress trial timelines and operational cost; what it has not yet done is move the Phase 2 success rate. The clinic still kills drugs. The asymmetry matters: speed-to-IND compounds linearly, but a few-point lift in Phase 2 success compounds against an eighty-percent failure rate and a ten-figure cost-per-failure. The largest AI value lever in biopharma is not faster discovery. It is decisions that survive Phase 2 — patient stratification, biomarker selection, dose finding, trial design — the decisions a regimen-grade substrate is built to bend.

Ask whether the business model itself is changing.

AI does not just change how fast and how well biopharma does what it already does. It changes what biopharma is. The unit of intervention is shifting. AI is what makes integrated regimens (diagnostic-stratified, multi-drug, monitored, dynamically adjusted) thinkable as innovations at population scale. Most of what is in the AI-discovered pipeline today was designed for the molecule-as-product world; almost none of it was designed for the regimen world it will increasingly have to compete in. The firms still optimizing for the single molecule inside a 1990s commercial architecture are pricing AI for a product category that is becoming a component, not the whole. The first three disciplines decide what to do with the assets the firm already has. The fourth asks what kind of firm those assets are building toward, in a world where the regimen — not the molecule — is the unit the patient receives, the payer reimburses, and the regulator increasingly evaluates. The transition is staged, not instantaneous. The healthcare system imposes long regulatory and commercial adoption cycles on any capability, AI-native or AI-rented, and even the most capable AI-bio firms must traverse them. That is itself the argument for the build-side: in a world where the system absorbs innovation slowly, the firms that compound the substrate end the decade with positional advantage; the firms that wait for the system to force them do not. Without the build-side underneath, business-model reinvention is rhetoric. The harder question, which this series picks up later, is whether pharma’s own AI capital is being deployed to build regimen-construction infrastructure or just to accelerate molecule discovery — and most boards cannot yet tell which.

Map the spend. Price the state. Test the layer. Ask the harder question. Together they are an answer to one question: what would an AI portfolio look like if every bet were priced correctly, visible bet plus build-side, against the pipeline value it produces? The firms that answer it will own compounding judgment, and the pipeline value it produces. The rest will own a growing AI expense base.

---

The frame is written from the top-twenty pharma seat because that is where the capital sits. It inverts on the other side of the table, and the inversion has its own version of the same disease. A techbio CEO is not choosing among twelve bets; they are one. Their job is to force the buyer to price them as a Commit, by showing that the company is the build-side a pharma would otherwise be paying multiples to construct. Two cases bracket the inning-one shape. Xaira Therapeutics launched in April 2024 with a billion dollars at seed on the Baker IPD lineage and Marc Tessier-Lavigne out of Genentech. Isomorphic Labs closed a 2.1-billion-dollar Series B on May 12, 2026, Thrive-led, on the AlphaFold substrate spun out of DeepMind. Both are capitalization facts at a scale most pharmas have not matched on the build-side line of their public accounts. Whether either translates into a clinical pipeline is the open question for inning two. A meaningful share of that capital arrives from technology investors fluent in the data-and-compute stack but less fluent in the pharma-decision substrate that turns a flywheel into a development decision. The result is a mirrored failure mode, call it substrate without translation: technical substrate beautifully capitalized, translational and regulatory underfunded. Same disease, opposite stack.

Pharma’s historical answer to a wave it could not build internally was to digest the companies that could: biologics in the 2000s, when Genentech, MedImmune, and ImClone were bought at multiples that looked irrational and now look conservative. The disanalogy worth naming is that those targets had Phase 3 assets at acquisition; today’s AI-bio leaders mostly do not. Inning two will sort the AI-bio cohort into those whose substrate translates to clinical pipeline at the speed pharma needs, and those whose substrate looks beautiful in slides and never converts. The price will be set by whoever capitalizes the build-side first, and proves it converts.

---

The framework also runs on a single deal. A large-cap pharma principal returns from a conference convinced the next move is a broad foundation-model deal with a top-tier AI lab: one rolling contract across the value chain. The layer-one demo was genuinely better than the internal team can produce. Run the four layers. Layer one: the model side is strong, but the demo runs on curated public data; production runs on the firm’s own messy, partially-labeled, GxP-controlled corpus. Layer two: not demonstrated; DSP-1181 and BEN-2293 are reminders that layer-one strength does not predict layer-two survival. Layer three: unresolved, and potentially decisive. Foundation-model economics want breadth and royalty across deployments; pharma economics want exclusivity on the molecules the model produces. The right shape is operational: exclusive-on-output, non-exclusive-on-platform, with a step-up trigger tied to a layer-two milestone the firm controls, a build-side capitalization line the pharma owns and funds in parallel, and a kill trigger written into the term sheet at month eighteen. Layer four: the deal lands with the CEO and never lands with the program leaders who would act on the model’s outputs. The right review starts at layers three and four, not one. Many foundation-model deals signed in 2026 risk becoming Drift dressed as a Commit, not because the model is wrong, but because the firm has not capitalized the substrate underneath the price.

The first top-twenty pharma to publicly terminate a flagship AI partnership and reallocate the capital to a named Commit, with a disclosed build-side line, will define the inning-two narrative. The mirror prediction sets the same clock on the other side: the first AI-bio company to capitalize translational and regulatory at parity with technical substrate, and have a pharma price it accordingly, will reset the AI-bio multiple within twenty-four months. The Bitter Lesson the next five years teach will be brutal in its specificity: AI-bio companies that compound longitudinal causal data into a substrate that bends a drug-development decision will hold pricing power across model generations; AI-bio companies whose moat is clever architecture on top of public biology will be flattened by the next frontier-model release. The same logic is starting to extend one layer up. Companies that compound longitudinal regimen-outcome data into a substrate that bends a treatment-and-adjustment decision will, on the timescale of inning two, hold a second kind of pricing power the molecule-only firms do not yet see they are competing for.

At the next AI allocation review, the test is whether anyone in the room can name a bet, price its state, produce its kill trigger, and identify the build-side capitalization line, without leaving the table. The build-side line, drawn honestly, also reveals which world the firm is building toward: whether the substrate it is capitalizing is sized for the molecule the firm has always sold, or for the regimen the firm has not yet learned to compose. Price the build-side, or you have not priced the bet. And the firm that does it on one named bet, with the build-side disclosed, runs into the harder question the next essay takes up: how to price a bet whose distribution of outcomes the standard models cannot produce. The next handshake photograph either ends the way IBM’s did, or it doesn’t.