AMD says it has entered an agreement to acquire Taalas, a Toronto-based startup that develops chips for AI inference by hardwiring specific models into custom silicon. Multiple reports describe Taalas’ approach as less flexible than general-purpose AI accelerators, because the resulting chips are customized for particular AI models rather than designed to run many models unchanged. The company positions this design trade-off as enabling lower-cost hardware and faster output for targeted use cases.
Taalas is reported to have raised a total of $219 million since it was founded in 2023. The startup’s technology includes accelerators tailored to individual models; one example mentioned is a chip built for a smaller version of Meta’s Llama 3.1. Taalas also says it can turn newly received AI models into hardware in about two months. Reporting also notes that at least some chips are manufactured using an older Taiwan Semiconductor Manufacturing Co. process and make use of on-chip SRAM.
AMD did not disclose a purchase price in the reported account, and an AMD representative declined to provide transaction details beyond confirming the acquisition agreement.
