NVIDIA unveils cuPotential CUDA-X library for AI agents in chip fabs at SEMICON West

cuPotential targets interatomic potential calculations, the core of materials and process simulation. Speeding them up lets fab AI agents run more virtual experiments on deposition, etch and defect formation before touching real wafers. NVIDIA pairs the library with NanoVDB for sparse volumetric data and with its broader digital-twin stack for autonomous defect inspection.

The partner list covers equipment makers (Lam, KLA) and EDA vendors (Cadence, Synopsys), positioning NVIDIA's software as the common substrate from chip design through manufacturing. A companion NVIDIA developer post this week showed NeMo Gym training environments reducing agents' tendency to produce plausible but incorrect artifacts in chip design and making tool calls more efficient.

NVIDIA also contributed its Vera Rubin AI-factory architecture to the Open Compute Project at the OCP summit, extending its strategy of making its hardware the reference design for data centers.

Caveats: the 2-3x figure comes from early tests, and the 30x claims are partner-reported for specific workloads. Watch for production deployments at leading fabs.