The race to optimize AI inference is entering a new phase, where companies are not just adding more GPUs, but rethinking how memory is used.
According to multiple sources, Anthropic has begun early discussions with UK-based startup Fractile to potentially add a new chip supplier. If realized, Fractile would join NVIDIA, Google, and Amazon as part of Anthropic’s AI hardware ecosystem.
Fractile’s key differentiation lies in its architecture. Instead of relying on external DRAM, the company is developing inference chips that integrate compute and memory on the same die using SRAM. This approach directly targets one of the biggest bottlenecks in AI today: the cost and latency of moving data between processors and off-chip memory.
By placing data closer to the compute units, the design aims to significantly improve efficiency. Early simulations suggest major gains in both speed and cost, although the chips are not expected to reach commercial readiness until around 2027.
Anthropic’s move reflects a broader strategy: avoiding dependence on any single vendor. The company already runs its infrastructure across multiple platforms, including NVIDIA GPUs, Google TPUs, and Amazon Trainium chips. Expanding its supplier base could provide greater flexibility and leverage, especially as inference costs continue to pressure margins.
Fractile is not alone in this direction. Other players are also exploring near-memory and SRAM-based architectures to optimize AI workloads. This points to a larger shift in the industry:
AI is no longer just about scaling compute, but about reducing cost per token.
If these new architectures prove viable, competition in AI chips may expand beyond traditional GPUs, opening the door to entirely new system designs.
