Anthropic in Early Talks to Buy DRAM-Less AI Inference Chips From UK Startup — Fractile's SRAM Architecture Reduces Need for Pricey Memory During Extreme Pricing and Shortage Crunch

Why It Matters

Diversifying its chip supply helps Anthropic curb inference costs and mitigate DRAM‑related bottlenecks, protecting margins as its revenue exceeds $30 billion. Successful SRAM‑based accelerators could reshape the AI hardware market by offering cheaper, faster alternatives to traditional GPUs.

Key Takeaways

• Anthropic evaluates Fractile as a fourth AI inference chip supplier
• Fractile's SRAM‑based chips eliminate DRAM bottlenecks, cutting latency
• Simulations claim 100× speed and 10× lower cost vs Nvidia GPUs
• Commercial readiness projected for 2027, aligning with Anthropic's long‑term roadmap
• Fractile seeks $200M at >$1B valuation, backed by top venture firms

Pulse Analysis

Anthropic’s recent move to explore a fourth silicon partner underscores the growing strategic importance of supply‑chain diversification in the AI era. With Nvidia, Google and Amazon already supplying inference hardware, the company is hedging against both pricing volatility and the chronic DRAM shortage that has throttled large‑model deployments. By courting a startup that promises a DRAM‑less architecture, Anthropic aims to reduce its per‑token compute cost—a key lever for protecting margins as its revenue run‑rate surges past $30 billion.

Fractile, founded in 2022 by Oxford PhD Walter Goodwin, is betting on an SRAM‑centric design that co‑locates memory and compute on a single die. In theory, this eliminates the data‑shuttle latency that dominates traditional GPU pipelines, allowing the chip to run large language models up to 100 times faster and at roughly one‑tenth the cost of Nvidia’s GPUs, according to internal simulations. Although the silicon is not expected to ship until 2027, the startup has already closed a $15 million seed round and is courting $200 million at a valuation north of $1 billion, with investors such as Founders Fund and Accel on board.

The interest from Anthropic signals that SRAM‑based inference accelerators could become a mainstream alternative to GPU‑centric stacks, especially as cloud providers and AI firms scramble for cost‑per‑token efficiency. Nvidia’s recent $20 billion acquisition of Groq and its own launch of the Groq 3 LPX inference chip illustrate the competitive pressure to innovate beyond traditional memory hierarchies. If Fractile’s performance claims materialize, the startup could force a recalibration of pricing models across the AI hardware ecosystem, accelerating the shift toward near‑memory compute architectures.