In May 2024, Recursion Pharmaceuticals unveiled BioHive-2, a 504-GPU supercomputer it called the largest in the pharmaceutical industry. Nearly two years later, Roche claims to have almost seven times that number, and none of the companies in this emerging GPU arms race has publicly tied a specific clinical candidate to its AI infrastructure.
Roche deployed 2,176 new on-premise NVIDIA Blackwell GPUs across the United States and Europe, bringing its combined on-premise and cloud infrastructure above 3,500 Blackwell GPUs, what the company calls the greatest announced GPU footprint available to a pharmaceutical company. The expansion, which extends a strategic NVIDIA collaboration that began in 2023, comes less than three weeks after Eli Lilly went live with LillyPod, its own 1,016-GPU supercomputer that it called the most powerful AI factory wholly owned and operated by a pharmaceutical company. (Roche’s claim counts hybrid cloud capacity; Lilly’s claim is limited to hardware it owns outright. Those are technically different benchmarks.)
The combined GPU count across Roche, Lilly and Recursion now tops 5,000 chips, though the raw total obscures meaningful differences: Recursion’s BioHive-2 runs older H100 GPUs, while Lilly’s system uses next-generation Blackwell Ultra chips and Roche’s deployment uses standard Blackwell hardware. Recursion can at least point to a concrete output: the Boltz-2 biomolecular foundation model, trained on BioHive-2, which predicts protein binding affinities with accuracy rivaling physics-based methods. Neither Roche nor Lilly has publicly named a comparable deliverable. Aviv Regev, EVP and head of Genentech Research and Early Development, said the compute expansion will let Roche’s scientists “build more sophisticated predictive frontier models” and compress the timeline from biological discovery to approved medicines through what Roche calls its Lab-in-the-Loop strategy, connecting biological and chemistry experiments directly with AI models, an approach she said Roche has pursued for more than five years. Lilly’s Diogo Rau, EVP and chief information and digital officer, told CNBC in October 2025 that the benefits of AI-assisted discoveries would likely materialize around 2030.
In an NVIDIA pre-briefing ahead of GTC 2026, Kimberly Powell, NVIDIA’s VP of healthcare and life sciences, said nearly 90% of eligible small molecule programs at Genentech, a Roche subsidiary, are now integrating AI, and that one oncology molecule was designed 25% faster while a backup drug candidate was delivered in seven months. That process often takes more than two years traditionally.
Wafaa Mamilli, who joined as chief digital technology officer in February 2025 after more than 20 years at Eli Lilly and a stint as CDTO at Zoetis, said the goal is to “embed AI across the entire value chain,” from drug development and manufacturing to diagnostics and commercialization. The release cites NVIDIA BioNeMo for drug discovery, Omniverse for manufacturing digital twins, Parabricks for genomics, and NeMo Guardrails for healthcare-grade conversational AI. Notably absent: a dollar figure for the investment, any named disease areas, or specific pipeline programs.
For perspective, pharma’s GPU arms race remains a rounding error by the standards of the broader AI industry. Elon Musk’s xAI began 2025 with 200,000 GPUs at its Colossus cluster in Memphis and has since expanded to more than 500,000 across three facilities, with a roadmap to one million. Meta has struck a deal with NVIDIA for millions of chips to fill data centers including a five-gigawatt facility under construction in rural Louisiana. Against that backdrop, Rau’s warning at the LillyPod launch may be the most useful thing any pharma executive has said about the GPU spending wave: “The hype is actually a serious threat to the research itself,” he said, “because if the hype becomes the story, then we’re all going to be disappointed.”
