Bristol Myers Squibb announced an expanded NVIDIA partnership on July 20: eight DGX SuperPOD systems on Vera Rubin NVL72 architecture, making it the first life-sciences company to acquire that hardware generation. The new system delivers up to 10x the performance-per-megawatt of the cluster it replaces.

It’s the third pharma company in nine months to announce its sector’s largest AI supercomputer. Eli Lilly and Roche came before it. What almost none of the coverage connected is that BMS also announced, two months earlier, a completely different deal with Anthropic. That difference is the part worth understanding.

Two layers, not one

The NVIDIA deal is owned compute. BMS trains foundation models on its own compound, protein, and biology data, for research spanning oncology, hematology, cardiovascular, immunology, and neuroscience, running experimental-design agents under a program it calls “Predict First.” That layer gets built, not rented, because that’s where BMS’s actual competitive edge lives: decades of proprietary scientific data no outside vendor has access to.

The other layer is different. In May, BMS announced a strategic agreement with Anthropic to deploy Claude Enterprise as what its own announcement calls a “shared intelligence platform” across roughly 30,000 employees, spanning research and development, clinical development, manufacturing, and commercial functions. That layer gets bought, because the value there comes from everyone across the company using the same general productivity tool, rather than from the model itself.

Two announcements, two layers, two completely different architecture decisions, and BMS made both in the same quarter without treating it as a single “build or buy AI” question.

Why owned compute, according to the person who decided it

BMS Chief Digital and Technology Officer Greg Meyers was direct about the reasoning in the NVIDIA blog post: neither traditional cloud nor traditional on-premise approaches fully served the company’s scientific needs. The prior cluster was already fully saturated, in his words, “we actually consumed all the space we had.” The stated goal is a single data plane connecting all of BMS’s research sites globally, with guaranteed GPU availability, to train proprietary models on data nobody else has access to replicate.

This isn’t a cost decision. It’s a decision about where the defensible asset actually sits. BMS’s proprietary scientific data, accumulated across decades of research, is exactly the kind of edge a third-party-trained model could never reproduce, no matter how capable that general model is.

The same logic is repeating across the sector

Eli Lilly announced a $1 billion co-innovation lab with NVIDIA in January and its “LillyPod” system, roughly 1,016 Blackwell GPUs delivering about 9 exaFLOPS. Roche built a hybrid AI factory with 3,500 Blackwell GPUs. BMS didn’t disclose a dollar figure for its own deal, but the pattern holds across all three: owned compute for the proprietary scientific asset, licensed commercial models for everything else. STAT News described BMS as the third pharma company in nine months to claim its sector’s largest AI supercomputer, and Sanofi is reportedly exploring a similar move next. This is a sector pattern forming in real time, not a single company’s race.

This connects to something we already wrote about vertical specialization as the defensible edge in AI: once the general-purpose model becomes a commodity available to anyone, the advantage shifts to whoever holds the specific vertical data that model can’t replicate. Pharma is exactly the case study, because the gap between “a capable general model” and “a model that knows what your company knows about your own compounds” is enormous, and that gap is precisely where owned compute gets built.

How we apply this split in discovery

In AI Maestro discovery, we push on this exact question before recommending architecture: is the asset you’re protecting proprietary data nobody else has, or a general capability any vendor could offer just as well? We’ve already written about how AI companies are deciding between building and buying SaaS, and BMS’s pattern sharpens it further: the right answer is almost never “build everything” or “buy everything.” It’s separating the two layers before deciding, exactly the way a pharma company managing research assets worth billions of dollars just did.

Separate what to build from what to buy in your AI architecture