Nvidia and SK hynix use AI physics to speed chip design

This NVIDIA–SK hynix story isn’t about a new chatbot, but it hits the race‑to‑AGI where it quietly matters most: hardware R&D velocity. By using PhysicsNeMo to train surrogate models that emulate complex etch and process simulations, SK hynix can iterate on advanced memory designs orders of magnitude faster than with traditional TCAD alone. Faster, cheaper exploration of process ‘recipes’ directly impacts how quickly next‑generation HBM and memory nodes reach production, and those memory stacks are now the lifeblood of AI accelerators. ([developer.nvidia.com](https://developer.nvidia.com/blog/using-ai-physics-for-technology-computer-aided-design-simulations/))

In the broader competition, this kind of AI‑for‑science loop tightens the feedback between model demand and chip supply. As leading labs push ever larger reasoning models, they need better bandwidth and more energy‑efficient memory; AI‑accelerated TCAD is one way the semiconductor side keeps up. NVIDIA also benefits strategically: PhysicsNeMo becomes part of the default toolkit for chipmakers designing around its GPUs, deepening platform lock‑in well beyond CUDA. If more fabs adopt AI physics workflows, the cadence of hardware generations that power AGI research could shorten, nudging the effective compute frontier forward faster than conventional EDA alone would allow.