London-based semiconductor startup Fractile is in advanced negotiations to raise approximately $600 million at a $6.5 billion pre-money valuation, according to reports from Bloomberg. The funding round represents a steep escalation in valuation from the startup's previous round in May 2026, which valued the company at roughly $1 billion.
The capital raise follows an initial supply agreement with Anthropic valued at approximately $250 million. Under the commercial arrangement, Fractile will supply its custom inference silicon to Anthropic, positioning the British startup as an additional hardware provider alongside Anthropic's existing compute partners, which include Nvidia, Google, and Amazon.
In-Memory Compute Architecture for LLM Serving
Fractile, founded in 2022 by Oxford PhD Walter Goodwin, is developing specialized silicon designed specifically to accelerate large language model inference. Unlike conventional accelerators that rely on high-bandwidth memory (HBM) or off-chip DRAM, Fractile's architecture co-locates compute units directly with on-die Static Random-Access Memory (SRAM).
This in-memory computing approach aims to address the memory bandwidth bottleneck inherent in autoregressive token generation. By keeping model weights and key-value cache states directly in fast on-chip SRAM, the design eliminates the high latency and energy expenditure required to repeatedly fetch data across external memory buses.
Fractile claims its architecture can deliver inference throughput up to 25 times faster than standard hardware setups at approximately one-tenth the operating cost. However, because SRAM offers lower memory density per square millimeter than DRAM or HBM, scaling on-chip SRAM capacity to host large multi-billion parameter models requires sophisticated multi-chip interconnects and wafer-scale integration techniques.
Delivery Roadmaps and Market Dynamics
Fractile's engineering ranks include veterans from Graphcore, Nvidia, and Imagination Technologies. Graphcore, once the United Kingdom's most prominent AI semiconductor startup, struggled to compete with Nvidia's CUDA software ecosystem and was acquired by SoftBank in 2024 for just over $600 million, an amount below its total venture funding.
To avoid similar software adoption hurdles, custom silicon startups increasingly seek early commercial validation from frontier AI laboratories. Fractile's commercial roadmap targets customer test samples in 2027, with volume production shipments slated for customer data centers by late 2027 and 2028.
Fractile's funding round is expected to close in the coming weeks as venture capital continues to pour into specialized AI inference hardware.
