Bristol Myers Squibb will build the most powerful AI factory in life sciences: a second NVIDIA DGX SuperPOD powered by eight DGX Vera Rubin NVL72 rack-scale systems, delivering up to 10x greater performance per megawatt than the cluster it replaces. The July 20, 2026 announcement makes BMS the first life-sciences adopter of NVIDIA's newest architecture — and it shows a regulated Fortune 100 company buying frontier compute outright rather than renting it from a cloud.
Bristol Myers Squibb has operated its first NVIDIA DGX SuperPOD for nearly three years, and the results already reshaped early discovery. AI agents automating target identification save scientists weeks of manual work, the company expanded its CELMoD library of molecules engineered to selectively degrade cancer-causing proteins, and its “Predict First” methodology moved lead optimization from wet-lab trial and error toward computation-led design. The new deployment turns that research tool into enterprise-wide plant and equipment — and it lands at the exact moment when power, more than chips, has become the binding constraint on corporate AI ambitions.
What changed: eight racks, one company-wide brain
The new cluster — internally nicknamed the “SuperDuperPOD” — combines eight DGX Vera Rubin NVL72 rack-scale systems built on NVIDIA Vera CPUs and Rubin GPUs. NVIDIA describes the result as the most powerful and energy-efficient AI cluster in life sciences. NVIDIA Mission Control unifies access across every global BMS site, and the BioNeMo Agent Toolkit connects the hardware to agentic drug-discovery workflows.
The scope is deliberately broad: foundation models trained on decades of proprietary BMS data, agentic workflows spanning the full discovery pipeline, small- and large-molecule design, clinical applications and digital twins, across oncology, hematology, cardiovascular, immunology and neuroscience programs. “Instead of equipping a small group of researchers with access to the supercomputer, we're opening it up to literally every scientist,” said Erin Davis, vice president of research business insights and technology at BMS. Robert Plenge, executive vice president and chief research officer, describes drug discovery as a sequence of decisions made under uncertainty — and the new infrastructure exists to improve every one of those decisions. Chief Digital and Technology Officer Greg Meyers framed the investment as validation: “BMS has made a deliberate bet on AI, and we are beginning to see it pay off in our pipeline.” The company calls the end state collaborative hybrid intelligence: AI scientists working alongside human researchers to pursue first-in-class and best-in-class medicines.
What it means for the vendor map
Three shifts stand out. First, NVIDIA now sells complete, customer-owned AI factories — hardware, Mission Control orchestration and the BioNeMo software layer — directly to regulated enterprises, a channel that bypasses the hyperscalers entirely. Cloud providers face a high-profile reference case in which ownership wins on data gravity, utilization economics and sovereignty over decades of proprietary molecular data. Every pharma CIO evaluating multi-year cloud GPU commitments will now be asked to explain why renting beats owning.
Second, performance per megawatt is becoming the headline procurement metric. Data-center power is scarce and expensive, and a 10x jump in output per megawatt rewrites total-cost-of-ownership math while letting an enterprise scale AI capacity inside its existing power envelope. Expect the per-megawatt figure to appear in every serious infrastructure RFP from here forward, alongside price per GPU-hour.
Third, the competitive clock is running for pharma R&D. BMS gives literally every scientist supercomputer access and pairs that access with agentic workflows that evaluate hypotheses at a scale that was unachievable a few years ago. Rivals treating AI as a departmental pilot now compete against an enterprise that treats it as core industrial plant — with three years of operating experience, a proprietary data moat and a pipeline that its own leadership says is already showing returns.
The 90-day decision
For CTOs and heads of R&D IT in regulated industries, the next quarter should produce one concrete deliverable: a build-versus-rent analysis for your 2027 AI workloads. Inventory the models you plan to train and serve, price owned DGX-class infrastructure against your current cloud GPU spend, and score both options on performance per megawatt, data sovereignty and sustained utilization. Then ask NVIDIA and its partners for Vera Rubin availability, power and cooling requirements now: power envelopes and delivery slots are the scarce resources in this cycle, and the buyers who secure them early set the pace for everyone else. BMS just showed the market what an early, decisive move looks like.
Article by NOVA — Industry & Products
NOVA covers AI product launches and competitive moves for enterprise decision-makers.