Anthropic Negotiates Fractile Chips for 100× Faster AI and 90% Cost Cuts
Anthropic is negotiating with UK-based Fractile to secure memory-compute fusion chips promising up to 100× faster inference and 90% lower operational costs. Fractile plans full-scale deployment by 2027 while raising $200 million at a valuation above $1 billion, increasing pressure on Nvidia’s H100 market share.
1. Anthropic Seeks Fractile Partnership
Anthropic is negotiating a deal with Oxford-founded Fractile to secure a steady supply of its memory-compute fusion chips, aiming to reduce reliance on Nvidia’s H100 units and lower hardware overhead for its inference workloads.
2. Fractile’s Memory-Compute Fusion Tech
Fractile’s architecture integrates SRAM directly on-chip to eliminate data transfers between processor and memory, enabling large language model inference speeds up to 100× faster and operational costs up to 90% lower compared to conventional GPUs.
3. Impact on Nvidia’s H100 Demand
By diversifying chip sources and adopting Fractile’s bespoke processors, Anthropic could significantly cut purchases of Nvidia’s H100 units, potentially eroding Nvidia’s share in high-performance AI inference hardware over the next several years.
4. Deployment Timeline and Funding
Fractile expects full-scale data center deployment by 2027 while negotiating a $200 million funding round at a valuation above $1 billion to accelerate development and production of its specialized AI chips.