Taalas HC1 — 17,000 Tokens/Second on Custom Silicon

Taalas Blog·high signal

Canadian startup Taalas unveiled HC1 chip: hardwired Llama 3.1 8B generating 17K tok/s per user, 73x faster than H200 at 1/10th power. 250W per card, air-cooled. Over M raised, 25 employees. Second chip for reasoning model by summer.