London startup Fractile raised a $220 million Series B led by Accel, Factorial Funds and Founders Fund to attack the inference bottleneck with a memory-centric chip design.
變化在哪裡
Fractile says current inference hardware leaves frontier models stuck around tens of tokens per second for difficult reasoning workloads.
CEO Walter Goodwin is pitching a design that avoids both HBM and conventional on-chip SRAM, moving toward in-memory compute rather than a better GPU clone.
為什麼重要
The company has not yet disclosed customers, production timing, benchmark numbers or quantified power savings, so the round is still a bet on architecture.
接下來看什麼
真正值得看的不是標題上的數字，而是產品、合約或平台轉向能否在接下來兩個季度形成可重複的曲線。
參考來源: CocoLoop