Peter McMahon

Infrastructure2 hours ago

Building Fixed Hardware Implementations of Neural Networks

Researchers from Yale, Cornell, Boston University, and NTT Research published "Physical Foundation Models: Fixed hardware implementations of large-scale neural networks," arguing that the ~1-year release cadence of trillion-parameter foundation models justifies building fixed, special-purpose hardware where the network is realized directly in physical structure (such as 3D nanostructured glass) rather than executed on general-purpose digital inference chips.

Why it matters

The economic argument here matters more than the optics demo. The authors are saying that once a single foundation model gets reused across thousands of downstream tasks, the amortization math flips: it becomes rational to tape out hardware whose weights are physically baked in, on the same yearly cadence as model releases. That is a direct challenge to the Nvidia general-purpose GPU thesis for inference, and it is the same logic Groq, Etched (Sohu), and MatX are already pursuing in digital silicon. PFMs push it one step further into analog physics.

For operators, the near-term read is not glass holograms running GPT-class models. It is edge inference: trillion-parameter capability in a power envelope a robot or drone can carry, which today caps out at far smaller models. If even a fraction of the claimed energy-efficiency gain materializes in nanoelectronic form, the on-device vs cloud calculus for humanoids and AVs changes. Watch NTT Research specifically; they have been quietly funding coherent optical compute for years and are the most likely party to move this from arXiv to a fab partnership.