AMD bets on affordable AI chips amid financing challenges

AMD announced on August 6 its acquisition of Taalas, a Toronto-based startup that builds AI inference chips so specialized they can only run a single model. The upside: the company’s HC1 accelerator claims throughput of over 16,000 tokens per second per user, roughly 45 times faster than optimized Nvidia hardware. The downside: if that one model falls out of favor, the chip becomes a very expensive paperweight. Speed and cost that look almost too good Taalas was founded in 2023 and raised $219 million in venture funding before AMD came knocking. Its signature product, the HC1 accelerator, takes a radically different approach to chip design. Instead of building general-purpose silicon that can run any AI model, Taalas integrates the weights for a single specific model directly into the chip itself. The architecture relies on an older TSMC process node combined with on-chip SRAM. The performance numbers back that up. Taalas claims the HC1 delivers inference at roughly 0.75 cents per million tokens. For comparison, Nvidia hardware runs about 3.79 cents per million tokens for similar workloads. AMD did not disclose how much it paid for the acquisition, though the deal is expected to cl...