Nvidia brings drug-hunting AI to AWS

Training digital biology models used to mean either spending millions building on-prem GPU supercomputers or spending months fighting cloud infrastructure.

At AWS re:Invent, NVIDIA officially brought its drug discovery AI platform, BioNeMo, directly to Amazon Web Services. Pharmaceutical and techbio teams running on AWS can now deploy BioNeMo through Amazon SageMaker, AWS ParallelCluster, and soon via NVIDIA DGX Cloud on AWS.

For researchers, it removes the headache of setting up distributed AI clusters from scratch and plugs pre-trained biomolecular models right where healthcare datasets already live.

Scaling ESM-2 across 512 H100 GPUs

BioNeMo acts as an end-to-end framework specifically built for generative AI in biology. It bundles pre-trained LLMs, proprietary data loaders, and optimized training recipes for tasks like target identification, molecule design, and protein structure prediction.

Instead of starting from zero, researchers plug in proprietary wet-lab data to fine-tune foundation models.

The headline benchmark here is Meta’s protein language model ESM-2. NVIDIA says ESM-2 hits near-linear scaling on 256 NVIDIA H100 Tensor Core GPUs. Push that cluster to 512 H100s, and a training run that took an entire month in the original research paper finishes in just a few days.

Developers can pull down checkpoints at 650 million or 3 billion parameters. The suite also packs in MegaMolBART for small-molecule chemistry and ProtT5 for protein sequence generation. Startups like Basecamp Research, Evozyne, and LabGenius are already putting the setup into production.

NVIDIA also highlighted its broader Clara suite on AWS: the medical imaging platform MONAI has crossed 1.8 million downloads, while its genomics tool Parabricks can process a whole human genome variant call in about 15 minutes, down from a full day on CPUs.

Why it matters

We think this partnership is a no-brainer for early-stage techbio startups, but it also tightens NVIDIA's ecosystem grip on enterprise healthcare.

Running biological LLMs requires massive distributed clusters that most biotech outfits have neither the budget nor the engineering talent to maintain. Putting BioNeMo inside familiar AWS plumbing like SageMaker lets small biology teams scale compute up and down on demand, without hiring a fleet of distributed systems engineers.

Strategically, NVIDIA is making sure that as biology transitions into an AI-first discipline, its proprietary CUDA stack and specialized frameworks remain the default operating system. Meanwhile, AWS protects its high-margin healthcare cloud business by offering best-in-class hardware workflows.

Still, faster compute only speeds up the in silico sandbox. Generating potential therapeutic candidates in three days instead of thirty is a massive leap for early-stage screening, but wet-lab validation and clinical trials remain the real, stubborn bottlenecks in drug discovery.