Eyeing faster, cheaper AI processing, Anthropic explores Fractile’s novel SRAM-based chips, potentially adding a fourth supplier as demand surges and infrastructure costs intensify globally.
Anthropic is in preliminary discussions with London-based semiconductor startup Fractile to explore the purchase of next-generation AI inference chips, according to a report by The Information.
If formalised, the move would introduce Fractile as a fourth supplier of server-grade AI silicon for the company, alongside existing partners NVIDIA, Google and Amazon.
The proposed chips are not expected to reach commercial deployment until around 2027, placing them outside Anthropic’s immediate procurement cycle. The timeline broadly aligns with the company’s longer-term infrastructure plans, including its collaboration on tensor processing units with Google and Broadcom.
Founded in 2022 by Oxford PhD Walter Goodwin, Fractile is developing an inference architecture that integrates memory and compute on a single chip using SRAM. This approach is designed to reduce reliance on external DRAM, where data transfer between processors and memory often limits performance in large-scale AI workloads.
By positioning data closer to processing units, the company aims to address latency and efficiency constraints associated with current GPU-based systems.
In earlier statements, Fractile indicated that simulations suggested significant performance and cost improvements over conventional designs, though physical prototypes have yet to be manufactured. The startup has raised $15 million in seed funding and is reportedly seeking further investment that could value it at over $1 billion.
Anthropic’s interest in alternative chip technologies comes amid rising demand for its AI services. The company has adopted a multi-supplier strategy to manage compute capacity, relying on a mix of in-house and third-party infrastructure rather than building dedicated data centres at scale.
