Anthropic in talks to buy breakthrough AI chip from Fractile, claims 100x speed, 90% cost cut over Nvidia

Anthropic in talks to buy breakthrough AI chip from Fractile, claims 100x speed, 90% cost cut over Nvidia

Anthropic is in early discussions with London-based chip startup Fractile to purchase its novel compute-in-memory AI inference accelerator, which the company claims delivers 100x performance and 90% cost reduction versus Nvidia GPUs for large language model workloads. Fractile, founded in 2022 by Oxford PhD Walter Goodwin, integrates SRAM storage and compute on a single die, eliminating the data-movement bottleneck between GPU and external DRAM that typically dominates inference costs. The chip is not expected to reach commercial production until approximately 2027, and Fractile has raised only $15 million in seed funding, though it is now pursuing a $200 million Series A at a $1 billion+ valuation, with potential investors including Founders Fund, 8VC, and Accel.

Why it matters: This position signals Anthropic's deliberate strategy to diversify beyond Nvidia, Google TPU, and AWS Trainium as its fourth chip supplier, directly targeting the inference-cost crisis that is the single biggest structural pressure on foundation-model margins. Anthropic's annualized revenue has surged from ~$9 billion in late 2025 to over $30 billion as of March 2026, but high inference costs continue to drag on gross margins. By locking in a compute-in-memory architecture that promises 90% cost reduction at the token level, Anthropic gains a potential escape from the hyperscaler-pricing leverage that currently caps its profitability, while also deepening its position in the ongoing 'context-engineering moat' race where per-token economics determines competitive viability.

Grounded expert take: This is a textbook example of the 'acqui-licensing' pattern where a frontier-model lab pre-commits to a startup's unproven silicon to secure architectural differentiation at scale. The 100x/90% claims are from simulation-only data — the chip has not yet been taped out — but the strategic signal is clear: inference cost has overtaken training compute as the binding constraint on model-lab business models. Fractile's timeline (2027) roughly aligns with Anthropic's Google-Broadcom TPU co-design, suggesting Anthropic is hedging between a custom TPU path and a radical compute-in-memory path. If Fractile delivers even a fraction of its claimed improvements, it could upend the current inference-economy structure where Nvidia holds effective pricing power via its installed base and CUDA moat.

#Anthropic #Fractile #AIChips #InferenceCosts #ComputeInMemory #NvidiaAlternative