Anthropic Seeks Fractile Chips to Cut AI Costs by 90% by 2027
Anthropic is negotiating with UK start-up Fractile to secure inference chips using memory-compute fusion architecture that claims 100× faster performance and 90% lower costs. The talks aim to reduce reliance on Nvidia H100 units, with Fractile chips launching by 2027 after a $200 million funding round at over $1 billion valuation.
1. Anthropic Explores Fractile Partnership
Anthropic is exploring a partnership with British semiconductor start-up Fractile to secure a steady supply of inference chips and reduce the overhead of existing solutions. The talks reflect Anthropic's goal to minimize reliance on Nvidia H100 units that currently power its custom inference workloads.
2. Fractile's Architecture and Timeline
Fractile's memory-compute fusion design integrates SRAM directly on-chip, eliminating separate memory modules. The start-up claims performance up to 100× faster and operational costs 90% lower, though its first commercial chips are not planned for data-center deployment until 2027.
3. Implications for Nvidia
If adopted, Fractile's chips could divert a meaningful portion of Anthropic's future hardware spending away from Nvidia, potentially impacting H100 unit volumes and revenue growth. The negotiations coincide with Fractile's effort to raise $200 million at a valuation above $1 billion.