Roche scales to 3,500+ NVIDIA Blackwell GPUs for AI drug discovery, diagnostics, and manufacturing

Roche turns AI infrastructure into a core healthcare operating layer

Roche said at NVIDIA GTC on March 16, 2026 that it is deploying more than 3,500 NVIDIA Blackwell GPUs across hybrid cloud and on-premises environments in the U.S. and Europe. NVIDIA described the buildout as the largest announced GPU footprint available to a pharmaceutical company. The scale matters because Roche is not describing a single pilot cluster. It is building an enterprise AI backbone that spans pharmaceutical research, diagnostics, manufacturing, and digital health.

According to the announcement, the new infrastructure will support biological and molecular foundation models, AI-assisted drug discovery, and digital twins for production systems. Roche says the goal is to move accelerated computing from isolated experimentation into an operating capability available across its global scientific organization.

Drug discovery claims already attached to measurable programs

Roche and Genentech say AI is already deeply embedded in their discovery engine. The company said nearly 90% of Genentech’s eligible small-molecule programs integrate AI. It also attached concrete numbers to that claim: in one oncology program, an AI-supported degrader molecule was designed 25% faster, and in another, AI helped deliver a backup molecule in seven months instead of more than two years.

With the larger Blackwell deployment and the NVIDIA BioNeMo platform, Roche says it can train and fine-tune biological and molecular foundation models on proprietary data while expanding AI-driven lab automation. That matters because the company is trying to search much larger regions of biological and chemical space without stretching development timelines in the same way.

Manufacturing and diagnostics are part of the same AI factory story

The announcement also goes well beyond discovery. Roche said it is using NVIDIA Omniverse libraries to build digital twins of production facilities, including a new GLP-1 manufacturing facility in North Carolina. Those simulations are intended to optimize systems before they go live and improve planning in areas such as regulatory documentation, quality assurance, and production scheduling.

On the diagnostics side, Roche said it is using NVIDIA Parabricks for large-scale data insights and NVIDIA NeMo Guardrails for safe and reliable healthcare-grade AI. The company also said it is advancing digital pathology models and evaluating BioNeMo for chemistry and sequencing-related work.

• Roche says it is deploying more than 3,500 NVIDIA Blackwell GPUs across the U.S. and Europe.
• NVIDIA describes this as the largest announced GPU footprint for a pharmaceutical company.
• Nearly 90% of Genentech’s eligible small-molecule programs already integrate AI.
• Roche says AI helped design one oncology degrader 25% faster and deliver another backup molecule in seven months instead of more than two years.

The broader signal is that healthcare AI is moving from model demos to industrialized compute strategy. Roche is treating AI factories as the shared infrastructure for science, manufacturing, diagnostics, and eventually precision medicine rather than as separate departmental experiments.

Source: NVIDIA