Memory , Fakes , and Thermodynamic Compute : CHIPS Act Bets $874M on Six AI Stack Gaps

The Department of Commerce signed letters of intent on July 29 with seven companies to provide up to $874 million in CHIPS and Science Act incentives — and the most revealing part of the announcement is not the dollar total but the six companies that are not GlobalFoundries. Most coverage of AI hardware bottlenecks focuses on the GPU and the optics that connect GPUs to one another. Those bottlenecks are real, and Washington has already committed heavily to them: GlobalFoundries received the round's largest award — up to $300 million — for silicon photonics and co-packaged optics, a story TechTimes covered in depth on July 29. But the other six awards in this round point at a layer of the AI hardware problem that has received almost no public attention: the memory architecture that feeds AI accelerators, the insulating film that governs signal integrity between chip layers, the electron-beam tool that makes fine-pitch chiplet assembly possible, the probabilistic compute paradigm that could replace GPU inference at a fraction of the energy cost, the scanner that catches counterfeit chips before they enter defense systems, and the compound-semiconductor substrate that removes a geographic bottleneck from photonic manufacturing. Taken together, the six non-GlobalFoundries…