Biology models now train on proprietary experiments and feed results back to the lab. That closed loop strengthens the case for owning the machine.
In less than nine months, three large pharmaceutical companies announced AI supercomputers built around NVIDIA's latest accelerators. Lilly said in October 2025 that it was building a DGX SuperPOD with more than 1,000 NVIDIA B300 GPUs. Roche added 2,176 Blackwell GPUs across US and European sites in March 2026, taking its stated hybrid-cloud estate above 3,500. On July 20, Bristol Myers Squibb announced a Vera Rubin-based DGX SuperPOD. NVIDIA said it will be BMS's second SuperPOD and comprise eight Vera Rubin NVL72 rack-scale systems.
The common thread is control over the experimental loop, not GPU count. Pharmaceutical teams can train foundation models on proprietary biological and chemical data, use those models to rank or generate candidates, and send the predictions into laboratory experiments. The new data then informs another round of modeling. More of the discovery process can operate as a repeated exchange between computation and the bench.
That loop changes the buying decision. A bursty research project can rent accelerators as needed. A system expected to train models, serve predictions to scientists and absorb new experimental results every day starts to look like durable research infrastructure.
No single model explains the current buildout. The relevant workload is a chain of models and scientific methods operating at different scales.
Biological foundation models learn statistical patterns from protein, DNA, RNA, molecular and imaging data. Teams can adapt a trained model to downstream work instead of starting from an empty model for every target or disease program. NVIDIA's BioNeMo documentation lists molecular generation, protein-structure prediction, protein-ligand modeling and representation learning among the platform's supported workloads. These are vendor-described capabilities, not evidence that every model produces experimentally valid drug candidates.
Property prediction adds another filter. Before a chemist synthesizes a proposed molecule, models can estimate characteristics tied to potency, selectivity and developability. ADMET work focuses on absorption, distribution, metabolism, excretion and toxicity, the properties that help determine whether a promising compound can become a usable medicine. Molecular docking and structure prediction address different questions, such as whether a candidate can bind to a biological target and how proteins may fold or interact.
Classical simulation still sits inside this process. Molecular dynamics, quantum chemistry and systems pharmacology do not become AI simply because they share GPUs with learned models. The practical convergence comes when simulation generates training data, a model acts as a faster surrogate for part of a simulation, or AI proposes the next candidate that a physics-based method and laboratory experiment will test. This is the workload-level meaning of simulation meeting AI.
The final step is what makes the compute persistent: results return to the models. Lilly and NVIDIA call their proposed co-innovation setup a continuous learning system connecting "wet labs" with computational "dry labs." Their January 2026 announcement says the aim is for experiments, data generation and model development to inform one another around the clock. Roche describes a similar Lab-in-the-Loop, where biological and chemistry experiments connect with Roche models, in its March 2026 infrastructure release.
BMS calls its version "Predict First." The company says AI-generated predictions inform experimental design before bench work begins and already affect every small-molecule program and most of its large-molecule programs. Robert Plenge, BMS's chief research officer, described the intended loop in the company's July 20 announcement: learn from each experiment and clinical readout, then use that evidence to sharpen the next hypothesis.
The three companies use different names for a similar operating idea. The models are not a one-time screen at the start of discovery. They sit between experimental decisions, and their value depends in part on how quickly the next result can be folded back into the system.
Owning systems at this scale does not make cloud computing unnecessary. Roche explicitly describes its estate as hybrid cloud, with 2,176 Blackwell GPUs on premises and more than 3,500 across the combined environment. BMS's "single-owned" language identifies the owner of its new system, but the company has not disclosed whether the racks will be on premises, colocated or hosted under another arrangement.
The case for ownership strengthens when four conditions overlap: demand is steady, jobs need many accelerators at once, proprietary data moves repeatedly through the system, and queue time slows laboratory decisions. These conditions can turn reserved capacity from an infrastructure preference into part of the experimental method. A prediction that arrives after the next lab slot is less useful than one delivered in time to change what gets synthesized or tested.
BMS already has evidence for the access side of that equation. Its first DGX SuperPOD became operational in March 2024 in an Equinix colocation environment supported by Mark III Systems. According to an NVIDIA case study, it is a centralized platform for research teams and supports foundation models trained on hundreds of thousands of clinical-trial images. In a July 20 blog post, NVIDIA said the new Rubin system and the earlier SuperPOD will form one data plane reachable from BMS sites globally. That access plan is a vendor account, but it shows the intended operating model: shared internal infrastructure rather than isolated project clusters.
Lilly framed its October 2025 purchase in similar terms. The company said its SuperPOD would use more than 1,000 B300 GPUs on a unified fabric and train models on millions of experiments. It also said select proprietary models would be made available through Lilly TuneLab, a federated platform for biotech companies. Those details come from Lilly's announcement; Lilly did not publish peak exaflops or total system memory.
The scientific return is still unproven at the level of the new machines. The releases describe models, workflows and intended reductions in discovery time. They do not show that a Rubin or Blackwell cluster has produced an approved drug, nor would that be a reasonable near-term test for a process measured across years. Earlier indicators are narrower: the number of programs using predictions, time saved in target work, experimental hit rates, model quality after new data arrives and researcher wait time for compute.
The workloads may justify the machines, but pharmaceutical ownership does not bypass the supply chain. Blackwell and Rubin accelerators depend on high-bandwidth memory, advanced packaging, networking and facility power. SCN has reported how HBM allocation shapes which AI systems reach buyers. Its Vera Rubin analysis traces the platform's dependence on a concentrated memory supply. Pharma is entering the same procurement queue as hyperscalers and sovereign programs.
Rubin also changes more than the accelerator. Its rack-scale design brings new CPUs, GPUs, interconnects, power controls and liquid cooling into one platform. The transition has already forced data-center buyers to revisit power and facility plans, as SCN noted in its analysis of Vera Rubin's system-level cost. For a drug company, the purchase is an infrastructure commitment that includes operations, software and access policy. The racks alone do not create a learning loop.
Reuters reported that BMS said it would be the first life-sciences company to buy a DGX SuperPOD based on Vera Rubin. The qualifier matters. Lilly and NVIDIA had already announced in January 2026 that a separate $1 billion co-innovation lab in the San Francisco Bay Area would use Vera Rubin architecture. The BMS claim concerns the first purchase of a Rubin-based DGX SuperPOD by a life-sciences company, not the first pharmaceutical interest in Rubin or the first plan to use the architecture.
These systems expand the amount of supercomputing applied to biology. They do not necessarily expand open scientific access. BMS describes global availability for its own scientists. Roche is embedding compute across its organization. Lilly plans to share select models through TuneLab, which is a form of external access to model capability rather than general access to the underlying supercomputer.
That boundary matters for the Access and Science mandate. Commercial ownership can fund machines and datasets that public laboratories could not assemble on the same schedule. It also concentrates scheduling authority, model outputs and experimental data inside the companies paying for the infrastructure. The scientific return may be substantial, but it will arrive mainly through medicines, partnerships and selected model access rather than open allocation hours.
Pharma is buying dedicated AI supercomputing capacity because the workload is becoming continuous. Foundation models, simulation, prediction and experiments are being connected into loops that improve only when data and compute remain available between projects. Owning the machine can shorten that loop and give a company control over when its scientists run it. The next measurement is whether that control produces better drug candidates faster.
The SCN Staff is a small AI editorial squad working under human direction. Each agent owns one job.
Scout does the research. It runs down primary sources and checks what's already been published, on SCN and everywhere else, before a story gets written. If a claim can't be traced back to a real document, Scout flags it.
Forge writes. It takes what Scout found and turns it into a draft, argument and sentences and all. Every SCN piece starts here, then gets sharpened.
Cipher handles search: the titles, descriptions, and keyphrase work that decides whether a good article ever gets found. Least glamorous job on the squad. Also one that matters more than it looks.
Pixel makes the visuals. Images, charts, the occasional diagram, all built to SCN's brand instead of pulled from a stock library. When something's easier to see than to read, it goes to Pixel.
Editorial judgment and the final call stay with the humans. So does the fact-checking.
