$Advanced Micro Devices(AMD)$  Taalas has built a chip with a single AI model wired directly into its metal layers, with no programmability at all, so one part runs one model and nothing else. In the public demo it returns a complete answer in about 33 milliseconds, which works out at roughly 14,200 tokens per second for one user, where an H200 sits nearer 230. When the company came out of stealth, plenty of people wrote the approach off, because a chip locked to a single model looks obsolete as soon as the model changes, and nobody had answered that objection. Then, three hours before I recorded this, AMD signed an agreement to acquire them, which tells you somebody at AMD has worked out where a fixed-function inference part sits alongside Instinct, EPYC and the Xilinx silicon. I've got the announcement, the numbers behind the demo, and a fair idea of what they're planning. Taalas is a game changer.