AI drug discovery is entering a more demanding phase: Big Pharma is buying AI infrastructure at industrial scale, but the value in this race accrues to whoever owns the workflow.
AI is delivering workflow advantage in pharma now. The discovery shortcut is years out.
AI is shortening the loop between hypothesis, experiment, and manufacturing yield. Clinical failure rates have not collapsed, and they will not collapse soon. The clearer the operational loop, the stronger the AI evidence today.
Discovery has the highest upside and the least proof. Operations and manufacturing are where AI is already showing up on the income statement.
Big Pharma is spending on AI as if opting out is no longer on the table. Eli Lilly is building LillyPod with Nvidia, a DGX SuperPOD with 1,016 Blackwell B300 GPUs and the largest computer ever owned by a drug company. In January 2026 the two firms went further, announcing a five-year, billion-dollar co-innovation lab in the Bay Area built on Nvidia’s BioNeMo platform. Roche, GSK, AstraZeneca, Merck, Takeda, and Recursion are working the same question. Can AI rewrite the economics of drug development?
The answer is messier than the slide decks suggest. Drug development still fails roughly nine times in ten. Models can propose molecules quickly. The hard parts — toxicity, trial design, patient response, FDA review, manufacturing scale-up — still happen on the same timelines they always did. A beautiful prediction still has to clear the clinic.
The durable AI opportunity in pharma sits in ownership of the closed loop: every step from data in to drug out.
Why pharma is building its own AI infrastructure
Traditional drug discovery rewards endurance over insight. AI compresses the search across chemical space, protein structure, documentation, and process conditions. Biology stays as hard. The FDA stays as exacting. Cost and tempo are where things change.
Look past the GPU count. LillyPod signals where pharma believes advantage concentrates: proprietary data plus domain-specific compute, owned inside the enterprise. Owning that stack gives a drug company direct control over sensitive data, model training cycles, and scientific iteration in a way that vendor relationships cannot match.
Google DeepMind’s AlphaFold remains the cleanest demonstration of what AI can do when the problem is well-framed and the data is rich. Protein-structure prediction is closer to solved than open. Drug development is harder for a structural reason. Predicting a structure is one step in producing a medicine that is safe, manufacturable, reimbursable, and effective in patients.
Where AI drug discovery proof is strongest today
The evidence in AI drug discovery splits into three categories.
| Area | What AI is actually doing | What remains unproven |
|---|---|---|
| Early discovery | Target identification, molecular design, protein-structure prediction, hypothesis generation, faster design cycles. | Whether AI-originated candidates consistently improve clinical success rates across therapeutic areas. |
| Clinical translation | Patient stratification, trial design support, biomarker analysis, interpretation of complex biological data. | Whether AI can reliably predict human efficacy and safety before the expensive trials run. |
| Operations and manufacturing | Digital twins, process optimization, yield improvement, quality control, sales-force targeting, back-office automation. | Less unproven. This is where AI value is already visible because the feedback loops are tight and the outcomes are measurable. |
Recursion is the live test case. Its platform has generated real biology insights and accelerated experimental design across a broad pipeline. More than a decade in, no Recursion-discovered drug has reached the market. That outcome calibrates the bar for everyone else.
Takeda is closer to the proof line. Zasocitinib, a TYK2 inhibitor designed with physics- and AI-assisted modeling, cleared Phase 3 trials in plaque psoriasis with positive results, and Takeda has signaled regulatory filings in 2026. Other AI-discovered candidates — Insilico Medicine’s rentosertib in idiopathic pulmonary fibrosis, Iambic’s IAM1363 in HER2-altered cancers — are working through human studies. None are approved. All are running the same test: whether AI-derived candidates survive the development funnel and become drugs that physicians prescribe and payers reimburse.
Lilly’s manufacturing case is more immediate. The company used a digital twin and machine learning to optimize process conditions for tirzepatide, the active compound in Mounjaro and Zepbound. That pattern — high-value workflow, measurable physical constraints, proprietary process data, tight loop between prediction and production — is what AI investors should be hunting for. The economics are easier to defend because the savings show up as yield and cost.
The same pattern is playing out one rung up the operational stack, where most of life sciences actually runs. Omega portfolio company Elemental Machines connects the equipment and environment inside research labs and GxP manufacturing facilities — temperature, humidity, vibration, utilization, freezer health — into a single LabOps intelligence platform, with agentic AI sitting on top of the operational data. Customers keep the data, receive contextualized alerts that prevent loss events, and pull compliance reporting under standards including 21 CFR Part 11. Most labs and manufacturing floors will never build a LillyPod. A great many of them need this layer.
The Omega investment lens
For AI investors, pharma is a clean reminder of where moats form. Model capability buys a seat at the table. Adoption, workflow integration, and proprietary data are what keep a company at that table once everyone has access to similar models. The more regulated and scientific the domain, the more that gap matters.
The AI companies that win in healthcare and life sciences sit close to the data and the workflow. Their teams work alongside scientific and clinical experts. The systems they ship improve outcomes the customer already pays for — yield, cost, trial success, revenue. Thin software wrappers on general models do not clear that bar.
The same discipline runs across enterprise AI. Technical advantage becomes company value only when it converts into adoption, contract value, and measurable economics. In pharma, that conversion is harder to fake.
Signals to watch
Evidence will judge the next phase of AI in pharma. The signals to track are concrete:
- Clinical proof: AI-originated or AI-optimized drugs winning large, controlled human studies and reaching approval.
- Cycle-time compression: repeated evidence that AI shortens discovery and development timelines without pushing risk downstream.
- Data advantage: proprietary biological, clinical, imaging, or manufacturing data that competitors cannot replicate cheaply.
- Workflow control: systems embedded in the day-to-day work of scientists, clinicians, manufacturing teams, and commercial organizations.
- Economic impact: measurable improvements in yield, cost, quality, trial design, or commercial execution that show up in financial reporting.
Companies that pass these tests deserve serious capital. Marketing-only competitors fade once a cycle compresses.
Beyond pharma
The pharma race is one of the clearest signals of where the AI cycle is heading. Model demos no longer move the market on their own. The questions now are where the model lives, what data it sees, what decision it changes, and whether any of that flows to the income statement.
This shift is overdue. The valuable AI companies will be the ones where the model sits inside an operating advantage at the workflow level. In drug development, the franchise lives along the full path from candidate to prescription. Finding the molecule is the opening move.
For more Omega perspectives on where AI adoption becomes durable workflow value, read the Omega Insight collection.