FundingFoundation Models & Infrastructure
Fractile Closes £577M to Build AI Inference Chips
Fractile closed £577M Series C to build next-generation AI inference chips for data centers.

The big picture: Fractile, a London-based AI chip startup, has closed a £577 million Series C round, according to Companies House filings. The company, founded by Walter Goodwin, is developing specialized chips for AI inference, aiming to significantly accelerate long AI workloads.
Why it matters:
- Scarcity drives investment: The AI chip market is experiencing intense activity due to high demand and limited supply, with competitors like Cerebras and Etched also securing significant funding.
- Performance bottleneck: Current chip architectures struggle with the scale of modern AI models, leading to slow inference times for complex tasks.
- Strategic customer wins: Fractile has an initial agreement to supply Anthropic with $250 million in chips, signaling strong market interest despite the hardware not shipping until 2027.
How it works:
- Inference specialization: Fractile’s chips are designed specifically for AI inference, the process where trained models generate responses to prompts.
- Integrated architecture: The company’s architecture keeps data close to the processing transistors, eliminating bottlenecks caused by moving data between separate memory and processor chips.
- Speed improvement: Fractile aims to cut the processing time for long AI workloads from approximately a month to a single day by achieving 1,200 tokens per second, up from 40.
The catch: Fractile has yet to ship a production chip or disclose revenue, with deployment not expected until 2027. The reported Anthropic deal is an initial agreement, and the company faces a highly competitive market with established players and well-funded startups. Its valuation currently reflects anticipated technology and market scarcity rather than proven deployment.
Key Facts
AIPressRoom Weekly
Every round we reported this week, the people behind the companies, and the events coming up. One email, every Tuesday.