According to Insight Beating monitoring, Anthropic is currently in talks to acquire the inference chip developed by London-based startup Fractile, with the chip set to enter data centers as early as next year. Founded in late 2022, Fractile's core solution involves using SRAM (Static Random-Access Memory) to replace GPU-dependent HBM (High-Bandwidth Memory), reducing the back-and-forth data movement between the chip and external storage to lower inference power consumption and costs. Similar players following a comparable path include Cerebras and Groq.
The deal is in its early stages, with the scale undisclosed and a risk of falling through. However, this potential order has become a key selling point for Fractile's latest funding round. Fractile, valued at over $1 billion, is seeking funding of over $100 million, with Founders Fund, 8VC, and Accel all in discussions. Prior to this, Fractile had only raised $15 million in funding from investors such as Kindred Capital, NATO Innovation Fund, and Oxford Science Enterprises.
Anthropic has consistently diversified its chip supply. The company has leased cloud servers from Google and Amazon, made a $30 billion commitment last fall to rent NVIDIA servers from Microsoft Azure (with NVIDIA contributing $10 billion and Microsoft contributing $5 billion as part of the deal), and recently agreed to purchase Google's in-house-designed chips for external use in Google Cloud. Reuters previously reported that Anthropic is also considering designing its own inference chips, in a strategy similar to that of OpenAI and Meta.
Inference cost is currently a pain point for Anthropic. Last year, the company's inference business gross margin was lower than internal expectations, and a recent surge in demand for Claude Code has caused a shortage of computing power, leading to some users being throttled during peak hours, prompting public outcry from developers. NVIDIA CEO Jensen Huang recently expressed regret for not investing in Anthropic earlier, believing that doing so might have prevented its significant shift towards Google and Amazon chips.
