Fractile Raises $220M Series B for Agent-Scale Inference Chips
UK startup Fractile just closed $220M Series B led by Accel, Factorial Funds and Founders Fund, with Conviction, Gigascale, O1A, Felicis, Buckley Ventures and 8VC piling on. The pitch is direct: the next bottleneck for AI is not training, it is inference at agent scale.
Agents now chew through tens of millions of tokens per task. Reasoning models think before they speak. Multi-step tool use compounds. Every additional turn is more latency, more electricity, more dollars. Fractile thinks the GPU stack will choke on this, and they are building chips with in-memory compute so the math happens where the weights live instead of shuttling between memory and compute. The claim: 25x faster than current alternatives at 10% of the cost. First silicon ships to enterprise customers in 2027.
The founder Walter Goodwin came out of Oxford robotics. The round, the largest UK AI chip raise to date, lands in the same week that reports surfaced of Anthropic looking at UK silicon partners. Read that signal however you want.
The interesting frame is this: voice agents, deep research, computer-use agents, continual-learning agents - all four hot categories of 2026 are token-hungry by design. If Fractile delivers anywhere near the numbers, the cost curve for running agents in production drops a digit, and that changes what is economically feasible to deploy. Watch this one.
https://fractile.ai
← Back to all articles
Agents now chew through tens of millions of tokens per task. Reasoning models think before they speak. Multi-step tool use compounds. Every additional turn is more latency, more electricity, more dollars. Fractile thinks the GPU stack will choke on this, and they are building chips with in-memory compute so the math happens where the weights live instead of shuttling between memory and compute. The claim: 25x faster than current alternatives at 10% of the cost. First silicon ships to enterprise customers in 2027.
The founder Walter Goodwin came out of Oxford robotics. The round, the largest UK AI chip raise to date, lands in the same week that reports surfaced of Anthropic looking at UK silicon partners. Read that signal however you want.
The interesting frame is this: voice agents, deep research, computer-use agents, continual-learning agents - all four hot categories of 2026 are token-hungry by design. If Fractile delivers anywhere near the numbers, the cost curve for running agents in production drops a digit, and that changes what is economically feasible to deploy. Watch this one.
https://fractile.ai
Comments