GTC 2026 drew the crowds with consumer GPUs, but the more consequential announcements for biotech came from pharmaceutical companies using the conference as a venue to reveal serious compute infrastructure commitments — and from research labs publishing the experimental results that separate AI t...
image from Gemini Imagen 4
Eli Lilly unveiled LillyPod, a fully owned and operated AI factory for drug discovery delivering over 9,000 petaflops of performance using 1,016 NVIDIA Blackwell Ultra GPUs, built and launched in just four months in Indianapolis. The company also committed up to $500 million to a five-year, $1 billion co-innovation lab with NVIDIA focused on AI-driven drug discovery, while Roche quietly amassed the pharmaceutical industry's largest announced hybrid-cloud AI factory at over 3,500 GPUs. Industry surveys show 70% of pharma and biotech organizations now actively adopt AI, but the infrastructure race is outpacing the organizational capacity to leverage it, with more consequential validation coming from experimental results like the AlphaFold expansion to 30 million protein complex predictions.
GTC 2026 drew the crowds with consumer GPUs, but the more consequential announcements for biotech came from pharmaceutical companies using the conference as a venue to reveal serious compute infrastructure commitments — and from research labs publishing the experimental results that separate AI theater from actual biological validation.
The GPU count race is real. Roche added 2,176 NVIDIA Blackwell GPUs to its existing fleet, bringing the company's total to more than 3,500 — what Roche says is the pharmaceutical industry's largest announced hybrid-cloud AI factory, according to a March 16 press release. Eli Lilly went further. On Feb. 26, 2026, Lilly unveiled LillyPod: the first fully owned and operated AI factory built specifically for drug discovery, running 1,016 NVIDIA Blackwell Ultra GPUs in a DGX SuperPOD that delivers over 9,000 petaflops of performance. The company built and launched it in four months, in Indianapolis, according to GeneOnline. NVIDIA and Eli Lilly also announced a co-innovation lab on Jan. 12, 2026 at the J.P. Morgan Healthcare Conference in San Francisco, backed by up to $1 billion in combined investment over five years, focused on using AI to reinvent drug discovery. According to a NVIDIA Healthcare survey of more than 600 professionals in digital health, pharma, biotech, medtech, and payers, 70 percent of organizations now actively adopt AI, up from 63 percent in 2024. Sixty-nine percent use generative AI and large language models, up from 54 percent the prior year.
But pharma has built the infrastructure faster than it has built the organizations to use it. The more interesting validation of AI-driven biology is coming from the labs running experiments at scale.
A partnership involving the European Molecular Biology Laboratory's European Bioinformatics Institute (EMBL-EBI), Google DeepMind, NVIDIA, and Seoul National University has calculated predictions for 30 million protein complexes. Of those, 1.7 million high-confidence homodimer predictions have been added to the AlphaFold Database, which now has over 3.4 million users from 190 countries. Centralized hosting saves roughly 17 million GPU-hours compared to each researcher running those calculations independently. The AlphaFold expansion is real and useful. But the more striking experimental validation of AI-driven protein design is coming from somewhere else.
Manifold Bio ran what it calls the first million-scale multiplexed validation of AI-designed protein binders, testing 1 million candidate designs against 127 protein targets in a single experiment, measuring over 100 million protein-protein interactions. The company's Proteina-Complexa model identified specific binders to 68 percent of targets tested — a competitive result for a generative protein design model. The point is not just the hit rate. It is that Manifold Bio ran the experiments rather than publishing predictions. The results were posted openly, with NVIDIA as a collaborator. That distinction matters: in a field prone to computational theater, experimental validation at scale is the one thing that cannot be faked.
The GPU infrastructure race has been building for months. The deeper question — whether all that compute compresses the time and cost of getting a drug to a clinical trial — is the one that has not been answered. According to an IQVIA press release from March 2026, IQVIA has deployed more than 150 intelligent agents across internal teams and client environments, with 19 of the top 20 pharmaceutical companies now incorporating them into workflows. That number suggests the tooling has arrived. Whether it changes the bottleneck is a different question.
The bottleneck, famously, is not discovery. It is development: the decade-long corridor from lab to clinic where most programs die for reasons that have nothing to do with compute. No GPU cluster has solved it yet.
Some of the more arresting biology at the conference was not computational at all. A team at the University of Pennsylvania led by Michael J. Mitchell, associate professor of bioengineering, published research in Nature Materials on March 17, 2026 showing that lipid chemistry can do more than carry mRNA — it can actively reprogram immune cell metabolism. Their redesigned lipid nanoparticles, built with a lipid variant called C12-2aN, delivered more than three times as much mRNA to the lymph nodes relative to the liver compared to an FDA-approved formulation. The C12-2aN lipid also lowered the expression of genes associated with systemic inflammation in human cells and mice, and boosted glycolysis in dendritic cells without sacrificing vaccine performance. The implication is that lipid chemistry is not passive scaffolding — it is an active lever on the immune response itself, and one that might finally break the tradeoff between mRNA vaccine efficacy and inflammatory side effects that has limited the field since 2020.
More than 60,000 nonhuman primates were used in U.S. research facilities in fiscal year 2024. Their numbers have dwindled in the United States after China banned the export of nonhuman primates in 2020. Thirteen Americans die each day waiting for organ transplants, and over 100,000 people are on the U.S. waitlist, as Wired reported. R3 Bio is attempting to address the primate shortage by engineering what it calls brain-free organ systems in monkeys — organs that lack the neural tissue required for consciousness or pain perception. The science is real and the ethical framing is deliberate. Whether it settles the moral question or relocates it is a debate the field has not finished having.
The deal that actually closed was GSK's acquisition of RAPT Therapeutics for up to $2.2 billion, completed March 3, 2026. The asset is ozureprubart, a long-acting anti-IgE monoclonal antibody in Phase IIb clinical development for prophylactic protection against food allergens, with data expected in 2027. It is a straightforward immunology bet, not an AI story, and it shipped on schedule while everyone was watching the GPU announcements.
The through-line from GTC is not really about silicon. It is about what happens after the infrastructure is built: which experimental systems validate AI's predictions, which biological platforms actually compress development timelines, and which of the 150 IQVIA agents is currently running a trial simulation that saves someone six months. Those questions are harder to announce at a conference than a GPU count. They are also the ones that determine whether the build-out was an investment or a very expensive furniture decision.
Assigned to reporter
Story entered the newsroom
Research completed — 18 sources registered. LNP lipid chemistry (Penn/Mitchell Nature Materials March 17); AlphaFold DB 30M complex expansion with DeepMind/NVIDIA/EMBL-EBI; Roche and Lilly GPU f
Draft (892 words)
Reporter revised draft (877 words)
Reporter revised draft (1024 words)
Approved for publication
Headline selected: Eli Lilly Built a 9,000-Petaflop AI Factory in Four Months
Published
Get the best frontier systems analysis delivered weekly. No spam, no fluff.
Biotech & Life Sciences · 20h 3m ago · 4 min read
Biotech & Life Sciences · 1d ago · 4 min read
