Powering Pharma AI with NVIDIA H100 and Blackwell GPUs

Recent breakthroughs in generative AI and large language models (LLMs) have shown that scaling up models with more data and compute can yield dramatic performance gains. Pharma companies are keen to apply similar approaches to biological and chemical data. Training AI models on millions of compounds or genetic sequences demands extensive parallel computing power, which modern GPUs excel at. NVIDIA's H100 GPU, for example, is currently one of the most powerful chips for AI, delivering exaflop-scale performance when many are used in tandem (^[3]) (^[4]). As Recursion's CTO Ben Mabey put it, "we see AI models in the biology domain improve performance substantially as we scale our training with more data and compute horsepower" – ultimately leading to better outcomes for patients (^[5]). This need for scale has driven the industry to adopt GPU-based supercomputers and cloud GPU clusters.

Key AI use cases driving GPU adoption in pharma/biotech include:

• Generative AI for drug design: e.g. models that generate novel molecular structures with desired properties (akin to how LLMs generate text). Companies use generative models to propose new drug candidates (small molecules or even biologics) and optimize them (^[6]).
• Protein structure and interaction prediction: GPUs enable protein-structure prediction (including AlphaFold-style models) and prediction of how drugs bind to targets (docking). For instance, NVIDIA's BioNeMo service offers models like DiffDock for predicting ligand-protein binding (^[7]).
• Molecular dynamics (MD) and simulations: High-end GPUs can perform MD simulations of molecular systems at high speed, helping researchers observe how proteins and compounds behave over time (^[8]).
• Analyzing 'omics' and imaging data: AI models for genomics or microscopy images (e.g. high-content cell imaging) require GPUs to train on terabytes of data. Vision transformers applied to cell images or models analyzing genomic sequences are used to discover disease biomarkers and new targets (^[9]) (^[10]).

To support these efforts, pharma and biotech firms are either procuring on-premises GPU supercomputers or leveraging cloud HPC with GPUs. Below, we profile several leading companies that have publicly confirmed deployments of NVIDIA H100 (and plans for Blackwell) in their AI infrastructure, along with the specific use cases enabled by this hardware.

One standout example in biotech is Recursion Pharmaceuticals, a tech-driven drug discovery company. Recursion completed BioHive-2 in May 2024, described as "the largest system in the pharmaceutical industry", which debuted on the TOP500 supercomputer list at #35 (^[11]). This AI supercomputer is powered by 63 DGX H100 systems containing 504 NVIDIA H100 Tensor Core GPUs connected via NVIDIA's Quantum-2 InfiniBand network, configured as an NVIDIA DGX SuperPOD (^[12]). The result is about 2 exaflops of AI performance. Recursion said BioHive-2 delivered four times faster speeds than BioHive-1 in benchmark performance tests.

How Recursion uses this GPU powerhouse: Their mission is to accelerate drug discovery by combining biological datasets at massive scale with AI models. Recursion has generated over 50 petabytes of biological images and data from automated experiments. With BioHive-2's compute muscle, they train foundation models on these data – for example, the "Phenom" family of vision-transformer models that turn high-resolution cellular images into useful biological representations. One such model, Phenom-1, was trained on 3.5 billion microscopy images to learn patterns of disease and treatment effects. These models can predict how cells react to new compounds, helping identify promising drug candidates faster.

Recursion's H100 cluster also enables enormous virtual screening tasks. In a collaboration with NVIDIA, Recursion demonstrated that combining BioHive-1 (earlier system) with cloud GPUs could screen ~36 billion chemical compounds in under 30 days, predicting potential protein targets for each. With BioHive-2's expanded capacity, such analyses are now even faster. The company has integrated this AI supercomputer into an end-to-end workflow (called "LOWE") with natural language interfaces for scientists.

Recent advances (2025): Recursion partnered with MIT's Jameel Clinic and CSAIL to develop Boltz-2, a cutting-edge model trained on BioHive-2. Boltz-2 approaches the chemical accuracy of physics-based free-energy perturbation (FEP) simulations while returning results in about 20 seconds on a single A100 GPU – a dramatic speed improvement. According to Recursion's CTO, "With AI in the loop today, we can get 80% of the value with 40% of the wet lab work, and that ratio will improve going forward." (^[13]) The release describes benchmark performance for Boltz-2, but does not establish realized discovery-timeline reductions or return on investment for Recursion or Amgen.

Global biotech leader Amgen has made a bold move to integrate AI at scale into its R&D. In January 2024, Amgen announced it will deploy an NVIDIA DGX SuperPOD AI supercomputer – nicknamed "Freyja" – at its deCODE Genetics subsidiary in Iceland (^[14]) (^[10]). Freyja will consist of 31 NVIDIA DGX H100 nodes (248 H100 GPUs total) and will be used to train state-of-the-art AI models in a fraction of the time previously required (^[10]). This essentially gives Amgen a dedicated, full-stack AI data center platform on-premises (^[14]).

What Amgen aims to do with Freyja: A key goal is to leverage deCODE's unique human genomic data for drug discovery. deCODE (acquired by Amgen) has amassed over 200 petabytes of de-identified human genetic and health data from ~3 million individuals, including a large portion of Iceland's population (^[15]). Amgen plans to use Freyja to build a "human diversity atlas" – AI models that can find drug targets and disease biomarkers by analyzing genetic variation across this vast dataset (^[16]) (^[10]). For example, they will train generative AI models on genomic and clinical data to uncover patterns that predict disease progression or drug response (^[9]). This could enable more precise identification of therapeutic targets and even aid in developing personalized medicines (by finding patient subgroups with particular biomarker signatures).

Amgen said its planned 248× H100 GPU Freyja system would train state-of-the-art AI models in days rather than months, enabling researchers to analyze and learn from data more efficiently (^[17]). This prospective capability was intended to support complex AI research, including generative models for drug discovery. David Reese, Amgen's CTO, described the union of technology and biotechnology as a "hinge moment" for the industry.

Amgen said Freyja would be installed at deCODE genetics' headquarters in Reykjavik, Iceland, to train state-of-the-art AI models in days rather than months and support analysis of its human data for drug-discovery research (^[17]).

German biotech BioNTech (famous for its mRNA vaccine) has also placed AI at the center of its R&D strategy. In 2023, BioNTech acquired InstaDeep, an AI startup, to bolster its machine learning expertise. As part of this effort, BioNTech/InstaDeep built a new in-house AI supercomputing cluster called "Kyber", unveiled in late 2024 (^[18]). Kyber is equipped with 224 NVIDIA H100 GPUs, 86,000 CPU cores, and high-speed networking, delivering on the order of 0.5 exaFLOPs of AI performance (^[19]). This makes it one of the world's top 100 most powerful computer clusters and among the top 20 H100 GPU clusters globally (^[20]).

Applications of Kyber at BioNTech: The cluster is intended to accelerate BioNTech's pipeline in areas like immunotherapy and vaccine development by enabling large-scale AI research. At the BioNTech AI Day 2024, InstaDeep showcased how Kyber powers new AI innovations in biology (^[21]). For example, they introduced Bayesian Flow Networks (BFN) – a novel class of generative AI models for biotech applications. BFNs can generate biological sequences (such as protein or antibody sequences) in a continuous, controllable manner, which could be useful in designing new proteins or optimizing vaccine immunogens. Unlike standard diffusion models, these BFNs allow researchers to more precisely guide the generation process, potentially speeding up the discovery of drug candidates or vaccines by exploring sequence space more efficiently.

2025 developments: At AI Day 2025, InstaDeep reported that its latest InstaNovo model was trained on 63 million labeled spectra and delivered a 10–15% increase in accuracy, up to twice as many peptide identifications, and 50-fold faster inference. InstaDeep said the model was being applied at BioNTech to uncover novel targets and biomarkers within the "Dark Proteome" (^[22]).

On top of Kyber sits AIchor, InstaDeep's orchestration platform that makes the supercomputer accessible to scientists through a streamlined GitOps workflow. InstaDeep reported that, in 2025, AIchor facilitated an average of 15,000 experiments per month across its research teams. Its CEO, Karim Beguir, described AI progress as a combination of data, compute and models (^[22]).

In Japan, several pharma companies are banding together to access world-class AI infrastructure. A prominent example is Astellas Pharma, one of Japan's top pharmaceutical firms, which is participating in the "Tokyo-1" initiative (^[23]). Tokyo-1, announced in 2023 by Mitsui & Co. in collaboration with NVIDIA, is Japan's first generative AI supercomputer for the pharmaceutical industry (^[24]). In its first iteration, Tokyo-1 includes 16 NVIDIA DGX H100 systems, each with eight NVIDIA H100 Tensor Core GPUs (128 GPUs total), hosted as a shared resource for pharma companies and startups in Japan. The effort is poised to accelerate Japan's $100 billion pharma industry, the world's third largest following the U.S. and China.

Astellas is using BioNeMo (NVIDIA's drug discovery AI platform) on Tokyo-1 to accelerate its research (^[25]). Concretely, Astellas plans to leverage this H100-powered supercomputer for tasks like molecular simulations and large language models applied to drug discovery (^[25]). By tapping into Tokyo-1, Astellas scientists can run high-resolution molecular dynamics to study how drug molecules behave, or train LLMs on chemical and biomedical text data to aid in drug design. The Tokyo-1 infrastructure supports generative chemistry models, allowing users to create novel molecular structures in silico, and also enables running quantum chemistry calculations faster than before (^[8]). In short, it provides Japanese pharma companies a competitive AI-as-a-service platform with H100-scale performance.

For Astellas, which might not have built its own giant supercomputer in-house, participating in Tokyo-1 is a way to still get access to state-of-the-art GPU hardware. According to NVIDIA, "the project will provide customers with access to DGX H100 nodes" for all these advanced applications (^[8]). The collaborative model (with Mitsui's Xeureka unit operating the facility (^[26])) means multiple pharmas can benefit. We know Astellas is on board; other major Japanese pharmas like Takeda or Daiichi Sankyo could join as well. This reflects a trend in some regions to invest in shared AI infrastructure for pharma R&D, rather than each company building from scratch. Nonetheless, it underscores that even in Japan (historically more focused on wet-lab research), there is now a push to embrace AI and GPU-accelerated computing in drug development (^[27]).

Beyond the examples above, numerous other pharma and biotech players have signaled serious investments in AI hardware:

Eli Lilly (USA) – World's Largest Pharma AI Factory with 1,016 Blackwell Ultra GPUs

In October 2025, Eli Lilly announced the world's largest pharmaceutical AI factory at NVIDIA GTC Washington, D.C. – a system powered by 1,016 NVIDIA Blackwell Ultra GPUs delivering 9.3 exaflops of AI performance (^[1]). NVIDIA described it as the first DGX SuperPOD with DGX B300 systems wholly owned and operated by a pharmaceutical company.

At the January 2026 J.P. Morgan Healthcare Conference, NVIDIA and Eli Lilly announced a $1 billion co-innovation AI lab focused on applying AI to tackle some of the most enduring challenges in the pharmaceutical industry (^[28]). The lab will be built on the NVIDIA BioNeMo platform and the upcoming NVIDIA Vera Rubin architecture, with select AI models to be made available on Lilly TuneLab – an AI platform providing biotech companies access to drug discovery models built on $1 billion worth of Lilly's proprietary data.

Using NVIDIA NeMo software, Lilly can create AI agents that reason, plan, and act across digital and physical labs, with goals including generating new molecules, designing treatments in silico, and testing them in vitro. With NVIDIA Omniverse and NVIDIA RTX PRO Servers, Lilly can create digital twins of manufacturing lines to model and optimize entire supply chains before making physical changes.

Novo Nordisk (Denmark) – Gefion Supercomputer Partnership

In June 2025, Novo Nordisk announced a multiyear partnership with the Danish Centre for AI Innovation (DCAI) to use the Gefion sovereign AI supercomputer, powered by 1,528 NVIDIA H100 Tensor Core GPUs (^[29]). This collaboration provides Novo Nordisk access to unparalleled computational capabilities for processing vast datasets and transforming drug discovery.

Novo Nordisk will use NVIDIA BioNeMo for generative AI-powered drug discovery, NVIDIA NIM and NVIDIA NeMo microservices for building customized agentic workflows, and the NVIDIA Omniverse platform to create physically accurate simulation environments. Researchers will focus on using single-cell models to predict cellular responses to drug candidates and designing models to build molecules with drug-like properties. As Lars Fogh Iversen, Senior Vice President at Novo Nordisk noted: "Gefion will allow us to tackle compute-heavy challenges, like AI-based protein engineering and biological models... the opportunities are immense."

AstraZeneca (UK) – Quantum-Accelerated Drug Discovery

AstraZeneca continues to be a pioneer in pharma AI. They utilized Cambridge-1, NVIDIA's £40M supercomputer launched in 2021 with 80 DGX A100 nodes, to develop the MegaMolBART generative model for molecules, trained on about 1.45 billion compounds (^[30]).

In June 2025, IonQ, AstraZeneca, AWS, and NVIDIA demonstrated a quantum-accelerated drug discovery workflow achieving over a 20-fold improvement in time-to-solution for the Suzuki-Miyaura reaction, a widely used method for synthesizing small-molecule pharmaceuticals (^[31]). The hybrid system integrates IonQ's Forte quantum processor with NVIDIA CUDA-Q and AWS infrastructure. AstraZeneca's Centre for Genomics Research has also set an ambitious goal to analyze up to 2 million genomes by 2026, supported by advanced AI and machine learning tools.

Insilico Medicine (USA/Hong Kong) and Others

Insilico Medicine continues to leverage NVIDIA GPUs via cloud services and NVIDIA's Inception program. Its generative chemistry platform (Chemistry42) and target discovery tools run on NVIDIA Tensor Core GPUs. NVIDIA reported that Insilico reached Phase 1 clinical trials with an anti-fibrosis candidate two and a half years after beginning the project, with generative AI used across its preclinical discovery process (^[32]).

IQVIA, a leading CRO, is using NVIDIA's AI Foundry to train custom AI models on its 64 petabyte real-world data repository (^[33]). Illumina is partnering with NVIDIA to accelerate genomic analysis (e.g. DNA sequencing data processing with DRAGEN). Google's Isomorphic Labs is partnering with Novartis and Eli Lilly to discover new treatments using AI.

NVIDIA BioNeMo Platform Expansion (January 2026)

At the J.P. Morgan Healthcare Conference in January 2026, NVIDIA announced a major expansion of the BioNeMo platform, an open development platform enabling lab-in-the-loop workflows for AI-driven biology and drug discovery. BioNeMo provides a development platform to generate and process data, train, optimize and deploy models – enabling the industry to turn data into a competitive engine for discovery and maximize the probability of success while minimizing R&D costs, currently estimated at $300 billion per year industry-wide.

New BioNeMo features include BioNeMo Recipes for accelerating biological foundation model training, nvMolKit (a GPU-accelerated cheminformatics tool), and new NVIDIA Clara open models including RNAPro for RNA structure prediction and ReaSyn v2 for ensuring AI-designed drugs are practical to synthesize. Model builders using BioNeMo now include Basecamp Research (EDEN family of AI models), Boltz PBC (Boltz Lab for AI-driven molecular design), Chai Discovery, Apheris, Dyno Therapeutics, OpenFold, and Terray Therapeutics.

The Vera Rubin Architecture – Next Generation (2026+)

Looking ahead, NVIDIA's Vera Rubin architecture, announced at CES 2026, represents the next evolution of AI compute for pharma. NVIDIA said Rubin-based products from partners were expected in the second half of 2026 and described up to a 5× increase in inference performance and a 10× reduction in inference token cost versus Blackwell (^[34]). The platform features:

• Rubin GPU: 336 billion transistors, up to 50 PFLOPs of NVFP4 inference (5× higher than Blackwell)
• HBM4 memory: Up to 288GB per GPU with 22 TB/s bandwidth
• Vera Rubin NVL72: 72 Rubin GPUs and 36 Vera CPUs delivering 3.6 EFLOPS of inference performance

The Lilly-NVIDIA co-innovation lab will be built on Vera Rubin architecture, and major pharma companies with established AI programs are expected to upgrade as systems become available in the second half of 2026. NVIDIA is framing Vera Rubin as ideal for agentic AI, advanced reasoning models, and mixture-of-experts (MoE) models – capabilities increasingly critical for complex drug discovery workflows.

In summary, the pharma/biotech sector worldwide is "getting serious" about AI, evidenced by substantial investments in NVIDIA's top-tier GPUs. From startup biotechs like Recursion building record-breaking supercomputers, to biopharma leaders like Amgen and BioNTech standing up their own AI clusters, and consortium efforts in Japan and Europe – the trend is clear. These organizations are not just dabbling; they are establishing AI as a core competency and backing it with the necessary infrastructure. The table below compiles some of the key players and initiatives, along with their known NVIDIA GPU deployments and AI use cases, as documented by publicly available sources.

Company	Country	Type	NVIDIA GPU Used	AI Use Case(s)	Source / Evidence
Eli Lilly	USA	Pharma	Blackwell Ultra (1,016 GPUs via DGX B300 SuperPOD)	World's largest pharma AI factory; AI agents for drug design, digital twins for manufacturing, $1B co-innovation lab with NVIDIA	([1]) (1,016× Blackwell Ultra GPUs, 9+ exaflops)
Recursion Pharmaceuticals	USA	Biotech (TechBio)	H100 (504 GPUs via DGX SuperPOD)	AI-driven drug discovery; training vision foundation models on cellular images; massive virtual screening; Boltz-2 model development	([12]) (504× H100 supercomputer, TOP500 #35)
Amgen (deCODE Genetics)	USA (Iceland site)	Biotech/Pharma	H100 (31× DGX H100 nodes = 248 GPUs)	Generative AI on 200+ PB human genomic data; creating "human diversity atlas" for target discovery and precision medicine models	([17]) (SuperPOD with 248× H100 for drug discovery AI)
BioNTech (with InstaDeep)	Germany	Biotech/Pharma	H100 (224 GPUs in "Kyber" cluster)	Scaling AI for immunotherapy; Bayesian Flow Networks for protein generation; InstaNovo for proteomics	([21]) (Kyber cluster with 224× H100); ([22]) (InstaNovo application)
Novo Nordisk (via Gefion)	Denmark	Pharma	H100 (1,528 GPUs in "Gefion" supercomputer)	BioNeMo for generative drug discovery; single-cell models; agentic AI workflows; biomedical LLMs	([29]) (Multiyear Gefion partnership for drug discovery)
Astellas Pharma (via Tokyo-1)	Japan	Pharma	H100 (16× DGX H100 nodes = 128 GPUs)	Molecular dynamics simulations, LLMs and generative chemistry via Tokyo-1 shared supercomputer	([24]) (Using Tokyo-1 H100 AI supercomputer)
Danish Centre for AI Innovation (DCAI)	Denmark	AI infrastructure operator	H100 (1,528 GPUs in "Gefion" supercomputer)	Owns and operates Gefion, providing AI-supercomputing access for healthcare, drug-discovery and other research	([29]) (DCAI owns and operates Gefion); ([35]) (1,528× H100; Foundation funding)

(Table notes: Eli Lilly's Blackwell Ultra deployment (October 2025) represents a major disclosed pharma adoption of NVIDIA's next-generation architecture. Novo Nordisk is listed as the pharma user of Gefion; DCAI owns and operates the system, while the Novo Nordisk Foundation was a funder of DCAI. As of May 31, 2026, NVIDIA said Vera Rubin was ramping into full production, with production shipments set to begin in the fall; this table does not treat that status as a confirmed pharma deployment (^[36]).)

The rapid adoption of NVIDIA's cutting-edge GPUs (H100, Blackwell, and soon Vera Rubin) by pharmaceutical and biotech companies illustrates a paradigm shift in the industry. Drug discovery and development – once the realm of slow, trial-and-error experiments – is being transformed into a compute-intensive, AI-driven process. The January 2026 announcements at the J.P. Morgan Healthcare Conference – including Eli Lilly's $1 billion co-innovation lab with NVIDIA and the massive BioNeMo platform expansion – mark a watershed moment where the industry's transformation from theoretical to operational becomes undeniable.

Companies that invest in powerful AI infrastructure can iterate faster on ideas, sift through vast chemical and biological spaces, and potentially bring therapies to patients more efficiently. The public announcements document substantial compute deployments and AI research programs, while the resulting effects on drug-discovery timelines and return on investment should not be inferred without company-specific evidence. Whether it is a nimble startup whose BioHive-2 marked a May 2024 pharma-supercomputing milestone, or an established pharma giant partnering to stand up a national AI center, these efforts all point to an AI arms race in life sciences.

These announcements provide evidence of GPU purchases, deployments and research projects. Their ultimate effects on drug-discovery productivity, clinical outcomes and return on investment remain company- and program-specific. NVIDIA has said the Vera Rubin platform is designed to reduce inference token cost by up to 10× versus Blackwell; the pace and scale of future pharma projects remain uncertain. The endgame is compelling: a future where "AI factories" for drug discovery run within pharma companies, continuously designing and testing virtual drugs, much like automated fabs. The companies highlighted here have positioned themselves at the forefront of this revolution, and their work is likely to shape how quickly AI delivers new medicines and health breakthroughs to society.

Looking ahead to 2026 and beyond: The industry is expected to see continued consolidation of AI infrastructure, with major pharma companies either building dedicated AI factories (like Eli Lilly) or partnering with sovereign AI initiatives (like Novo Nordisk with Gefion). In January 2026, the FDA's CDER and CBER, collaborating with the EMA, published 10 guiding principles for good AI practice in drug development (^[37]). The principles address AI-generated evidence across the medicine lifecycle, including research, clinical trials, manufacturing and safety monitoring; they are intended to inform future guidance, rather than establish a completed validation framework for drug approvals. The scale of pharma spending specifically on AI compute and the return on investment from these programs are not established by the public sources cited here.

In summary, pharma and biotech firms serious about AI are investing heavily in NVIDIA GPU technology. They are coupling these hardware investments with talent and data to create a new kind of R&D engine – one where computing power and scientific insight go hand in hand. The early adopters have set a high bar, demonstrating that with the right supercomputing resources, tasks once deemed intractable (like searching a chemical universe of billions of molecules) become feasible (^[11]). This is an exciting time at the intersection of computing and biology, and 2026 promises to be the year when pharma AI transitions from competitive advantage to competitive necessity – ensuring companies have the GPUs, AI models, and agentic workflows needed to compete in the next era of drug discovery.

For more insights on AI and data science in the pharmaceutical industry, check out our articles on data science in life sciences, generative AI proof of concepts, and commercial analytics software.