NVIDIA bets on open biology models — the bottleneck is FDA approval, not AI capability
At Fortune Brainstorm Tech, NVIDIA's healthcare chief reframes AI drug discovery's core constraint as regulatory acceptance, not scientific capability. Plus: Novartis expands its molecular-glue AI deal to $1.4B, Sanofi locks in Owkin for five years, and The Onion Desk.
The headline from Fortune Brainstorm Tech was about US science funding cuts arriving at an inopportune moment. The more durable insight came from NVIDIA's healthcare chief, who argued that open-source biology foundation models should function as shared infrastructure — and that the real constraint on AI drug discovery isn't the science but the FDA's ability to approve molecules whose training data origins can't be traced. 'We're just not there yet,' she said. That framing inverts most of the sector's conversation, which focuses on model performance rather than the regulatory evidence stack that determines whether model output translates into an approved drug.
Lead — NVIDIA's Powell draws the line between open and closed
At Fortune Brainstorm Tech in Aspen, industry leaders addressed US science funding cuts and their upstream effects on AI drug discovery. NVIDIA's Kimberly Powell advocated for open-source biology foundation models as shared infrastructure, contrasting with proprietary closed systems. The central regulatory question — how the FDA approves molecules designed by AI when training data origins cannot be traced — remains unresolved; Powell's proposed answer, biological digital twin simulations accurate enough for regulatory acceptance, is aspirational by her own admission. The practical analysis: winning teams generate validation data credible to regulators, not the flashiest models. Closed proprietary engines may widen the gap between labs affording validation studies and those without, likely accelerating consolidation. Open models sustain biotech diversity but require wet-lab-plus-simulation loops at scale to produce FDA-persuading evidence.
On our radar
Novartis expanded its molecular-glue collaboration with Orionis Biosciences to a potential $1.4B, paying $40M upfront for continued access to the Allo-Glue platform — which combines small-molecule chemistry with AI systematically profiling target-ligase combinations to reach historically undruggable proteins. The renewal of a 2020-originating relationship signals Novartis confidence in round-one results; molecular-glue degraders are increasingly where AI represents the licensed intellectual property rather than the molecule itself. Separately, Sanofi licensed Owkin's K Pro platform for a five-year autonomous AI agent deployment across R&D workflows. K Pro integrates multimodal patient data — pathology imaging, genomics, clinical records — with agentic systems designed for pharma. The five-year commitment tracks the deepening arc of the Sanofi-Owkin relationship: target identification (2021), drug positioning (2024), autonomous agents (2026).
Quick signals
GlobalData highlights manufacturing as AI's near-term pharma opportunity over discovery — better regulatory comfort and clearer ROI on production speed and quality control. Radical Numerics raised a $50M seed round for frontier models spanning AI-biotech and biosecurity, with backing from Patrick Collison and deep-tech funds. Cellares is expanding automated cell-therapy manufacturing capacity to address throughput bottlenecks constraining cell therapy scalability and economics.
The Onion Desk
This week: a pharma company's AI system surfaces actionable insights from its own historical data — insights the company generated, stored, and then rediscovered at significant consulting expense. A media outlet attributes a formulation optimization to 'AI drug discovery,' prompting a quiet correction buried in the methodology section. And a clinical-trial AI system, facing trust barriers from 47% of surveyed executives, begins a personal growth journey toward being trusted. Full satirical dispatches on Substack.