Pharma Keeps Renting AI. That Might Be the Point.
Bristol-Myers Squibb announced May 20 it is putting Anthropic's Claude in the hands of 30,000 employees — across research, clinical development, manufacturing, commercial operations, and corporate functions. It is the largest single enterprise AI deployment in pharmaceutical history. And it is, explicitly, a rental.
BMS is not buying Anthropic. It is not building its own large language model. It is paying for access to a shared platform that tens of thousands of other companies also use. The molecules Claude helps BMS discover will belong to BMS. The data those discoveries generate will flow, at least in aggregated form, back into the platform that produced them.
That is the pharmaceutical industry's current bet on artificial intelligence: rent the future, own the outputs, and let someone else figure out the infrastructure.
The deal is new. The strategy behind it is not. BMS is the latest pharmaceutical company to conclude that building world-class AI from scratch — billions in GPU infrastructure, hundreds of specialized engineers, years of iteration — is a distraction from what pharma does best: navigating the regulatory path from molecule to market. Renting keeps you at the frontier without owning the server farm.
The scale of that bet became legible in 2025. Eighty-seven AI partnership announcements that year, according to Mordor Intelligence — a pace that has not slowed in 2026. Cloud-based platforms now account for 82.43% of the AI drug discovery market, expanding at 27.43% per year. Lead optimization cycles that once took eighteen months can now run in six. McKinsey estimates agentic AI can lift clinical development productivity 35% to 45% over the next five years.
"The most sophisticated enterprise AI stops at the chatbot," said Greg Meyers, Bristol-Myers Squibb's chief information officer, in the press release announcing the Anthropic deal. "The real prize is the untapped value still trapped behind decades of data silos, and this collaboration is how we reach it."
Meyers is right about the prize. What he does not say is what happens to that prize — and to the data that unlocks it — when it travels through someone else's infrastructure to claim it.
The Data Problem Behind the Efficiency Story
Here is the structural asymmetry at the center of pharma's AI rental economy.
Every query a BMS scientist runs through Claude — every target ID, every molecular design, every failed assay — generates data. That data, in aggregate with similar queries from Novo Nordisk, Merck, Sanofi, and dozens of other pharmaceutical companies querying the same platform, becomes the most valuable asset in drug discovery: cross-industry biological learning.
No single pharmaceutical company can generate this data at scale from its own internal experiments. The sheer volume of queries, across hundreds of molecular targets and thousands of assay conditions, that accrues to a shared AI platform is categorically larger than what any single company's internal database holds. That is the compounding mechanism that makes the rental model dangerous in ways the efficiency story obscures, as BioPharma Dive noted when it covered the widening gap between pharma's AI rental spending and its ability to retain the resulting learning.
BMS owns the molecules Claude helps it discover. That is not in dispute. What the contracts typically do not specify, at least in their public descriptions, is what the AI platform learns from those molecules in aggregate — which structural features reliably predict activity, which targets are tractable, which chemical series have gone nowhere across dozens of attempts.
"The question is where the compounding returns accrue," said one former pharma R&D head who has watched the wave of AI partnership announcements from the outside. "If you rent the infrastructure and the platform accumulates the learning faster than you can from your own experiments, you are paying subscription fees and training your competitor. That is a structurally bad trade, even if it looks rational quarter by quarter."
The counterargument is real. As IntuitionLabs noted in its build-versus-buy analysis of pharma AI strategy, pharma has always outsourced capabilities it could not efficiently own — CROs for clinical operations, external compound libraries, genomic sequencing. The industry has never collapsed into a single vertically integrated entity. The argument that AI rental is categorically different rests on whether the data that flows back to the platform is categorically more consequential than the data that flowed to, say, a CRO. On that question, the industry is split.
Novartis CEO Vas Narasimhan joined Anthropic's board this year — a signal that at least some pharmaceutical executives believe the AI companies are not just vendors but partners whose long-term trajectory matters to pharma's own. Whether that board seat buys Novartis privileged access to Anthropic's learning, or merely acknowledges that Anthropic now sits at the center of the industry's most important infrastructure, is not publicly known.
The Builders
Not everyone is renting.
Roche spent 2023 through early 2026 constructing what it claims is the pharmaceutical industry's largest announced GPU footprint — 3,500 NVIDIA Blackwell GPUs across on-premise and hybrid-cloud infrastructure, supporting a digital-twin manufacturing network, an AI-accelerated Lab-in-the-Loop discovery platform, and digital pathology scanning at scale. The company is not renting AWS to run its models. It owns the iron.
Eli Lilly partnered with NVIDIA on a $1 billion-plus co-innovation laboratory over five years, building dedicated compute infrastructure for AI-driven drug discovery. Thomas Fuchs, Lilly's chief AI officer, put the strategic intent plainly last year: "Lilly is shifting from using AI as a tool to embracing it as a scientific collaborator." That is a build argument, not a rental argument.
And then there is Isomorphic Labs.
The consolidation pattern in AI drug discovery reinforces the divergence. When Exscientia and Recursion merged in a $688 million all-stock deal, the stated rationale was complementary pipelines — but the combined entity also gained the scale that comes from owning more internal learning infrastructure rather than renting it. Smaller AI biotechs that lack the capital to build are increasingly merging with similarly cash-constrained peers; the survivors are the ones that can afford to own the stack.
Isomorphic is the Alphabet-founded drug discovery company that raised $2.1 billion in a Series B round earlier this year — the largest private AI biotech raise on record. Alphabet did not write that check to become a pharma IT vendor. It wrote it because the path from AI platform to approved drug runs, in Isomorphic's view, through owning the full stack from target selection to Phase I. Alphabet is building because Alphabet believes it can build better than anyone renting.
The tension here is not just strategic — it is existential for the pharma industry. Isomorphic, with Alphabet's compute resources and access to Alphabet's models, is attempting to demonstrate that an AI-native company can out-learn a century-old pharmaceutical company that owns the molecules but not the learning infrastructure. If Isomorphic succeeds — if its drug candidates move through clinical development at materially lower cost and higher success rates than its pharma competitors' — then the rental model will have revealed its structural flaw: the landlord always knows more than the tenant.
What the Landlord Knows
The BMS deal with Anthropic is the clearest articulation of what pharma's rental strategy looks like at scale: 30,000 employees, every function, one platform — a deployment one analyst at Pharmaceutical Commerce called the most comprehensive AI integration shift the industry has seen. CEO Chris Poerner set a target of halving the time from target selection to lead molecule identification. McKinsey's productivity estimates — 35% to 45% improvement over five years — sit behind that ambition.
These are not trivial gains. A ScienceDirect analysis of AI in drug discovery frames the productivity crisis as the structural rationale for the rental model: when the cost of falling behind compounds faster than the subscription fees, the math favors keeping the lab rented. The 40% decline in pharmaceutical R&D productivity between 2010 and 2024 is not a talking point; it is a structural crisis. The average cost to bring a single molecule to market now exceeds $2.6 billion. If AI rental can compress timelines and restore productivity, the subscription fees are worth it, full stop.
But the structural question is not whether AI rental produces better molecules today. It may. The question is who captures the compounding returns as the platform learns faster than any individual pharma company can.
If the current pattern holds — if pharmaceutical companies continue to rent rather than build, and if AI platforms continue to aggregate cross-industry learning faster than any single pharma's internal data can compound — then five to ten years from now, the AI companies will hold the most comprehensive view of biological mechanisms, target tractability, and molecular design principles in history. They will know which chemical series have failed at which targets, which assay conditions predict human efficacy, which clinical trial designs are most likely to succeed. They will know this not from one company's experience but from the aggregated experience of dozens of pharmaceutical companies across thousands of experiments.
Pharma will retain the regulatory relationships, the commercial infrastructure, the patient relationships, the molecules. The question is whether the intelligence that drives which molecules to advance — the judgment call that sits upstream of everything else — will rest inside the companies that own the drugs, or inside the platforms they rent them from.
Roche, Lilly, and Isomorphic are making the opposite bet. They are betting that owning the infrastructure, and the learning it generates, is worth the capital cost and organizational complexity of building rather than renting.
Pharma's "rent, don't build" strategy may be entirely rational at the individual company level. It keeps options open. It avoids $1 billion infrastructure bets. It gets state-of-the-art AI into scientist's hands faster. The subscription fee is predictable; the building cost is not.
But collective rationality and individual rationality are not the same thing. When every pharmaceutical company makes the same rational choice — rent the AI, keep the molecules — they collectively create a new kind of dependency. They pay the subscription fees and, in doing so, fund the compounding asset that makes the platform more valuable to everyone, including their competitors.
The short version: they are renting the laboratory, and the landlord is taking notes.
This story is based on public press releases, earnings filings, and industry reports from Bristol-Myers Squibb, Roche, Eli Lilly, Isomorphic Labs, and third-party market research. The structural argument about cross-industry data accumulation reflects observable patterns in the deal structures and public statements of the companies involved; it is not based on access to non-public contract terms.