Eli Lilly's AI supercomputer, named LillyPod, went live on February 25. Roche deployed its NVIDIA AI factory on March 16. Bristol Myers Squibb has been running its own AI factory for almost three years. NVIDIA hardware powers all three buildouts.

The three deployments share a common foundation but diverge in timing and approach. BMS holds the longest track record, with nearly three years of continuous operation. Lilly and Roche are recent entrants, bringing their clusters online within the past two months.

NVIDIA uses the term "AI factory" to describe a GPU cluster built to train and run AI models. The label frames the infrastructure as a production facility: raw data flows in, trained models flow out. For R&D organizations weighing whether to build, buy or rent such capacity, the pharma deployments offer three distinct reference points.

The timing gap matters for portfolio planning. BMS has had roughly three years to integrate its AI factory into drug discovery workflows, giving it time to measure what the infrastructure actually delivers. Lilly and Roche, by contrast, are still in early deployment. Their February and March launch dates mean internal teams are only now beginning to route workloads through the new systems.

All three companies chose NVIDIA as their hardware supplier, a decision that consolidates their AI infrastructure around a single vendor's ecosystem. That choice carries implications for procurement strategy, software compatibility and long-term vendor lock-in — considerations any R&D manager evaluating similar buildouts must weigh.

The source material does not specify cluster sizes, GPU counts, power draw or funding figures for any of the three deployments. It also does not detail which specific NVIDIA products each company installed, or how the companies measure return on their infrastructure investment. Those gaps limit how much outside observers can benchmark the buildouts against one another.

What the record does establish: three of the world's largest pharmaceutical companies have concluded that owning dedicated AI compute is worth the capital expense, rather than relying solely on cloud providers or external partnerships. BMS's three-year head start suggests the company saw the case for on-premises GPU capacity well before its peers. Lilly and Roche's near-simultaneous deployments in February and March indicate the calculus shifted industry-wide, or at least among these three competitors.

For R&D leaders tracking the trend, the open questions are operational. How much of each company's model training and inference runs on these clusters versus rented cloud capacity? What workloads justify dedicated infrastructure? And does owning the hardware translate into faster discovery cycles or better candidate selection? The companies have not publicly answered these questions in the available material.

What comes next depends on whether Lilly and Roche report measurable outcomes from their new clusters — and whether BMS's three years of operation have produced results the company is willing to share.