FRACTILE raises €187 million to develop chips designed for future AI agents

While the race in artificial intelligence has largely focused on training models, another problem is emerging as models become more capable: execution.

It is precisely on this critical layer that Fractile, a British startup, is positioning itself with the announcement of a $220 million funding round, approximately €187 million. The round was led by Accel, Factorial Funds, and Founders Fund, with participation from Conviction, Felicis, 8VC, Gigascale, O1A, and Buckley Ventures.

Founded in 2022, the startup develops hardware architectures designed to accelerate inference for frontier models,  in other words, the phase during which models actually generate outputs. Long considered secondary compared with training, inference has now become a central issue with the emergence of reasoning models and future autonomous agents.

Fractile’s thesis is built around a simple idea: the most advanced models will soon no longer be limited by their theoretical capabilities, but by the time required to execute their reasoning chains.

“We bet that the most advanced AI systems would ultimately see their impact constrained by the time needed to produce useful outputs,” a company representative explained. “The only way to truly unlock this latent value was to radically reinvent the hardware powering frontier models.”

This evolution is gradually reshaping the economics of the sector. Every query sent to an AI model consumes compute resources. And as models become more complex, inference costs continue to rise. New reasoning systems now generate long processing sequences, sometimes spanning tens of millions of tokens.

Fractile estimates that some models already generate up to 100 million tokens to solve complex problems. At execution speeds close to 40 tokens per second on current architectures, such processing could require nearly a month of continuous computation.

For the company, this constraint goes far beyond a simple performance issue. “Inference is both the revenue engine of the AI industry and the primary factor limiting its expansion.”

Fractile draws a parallel with the systems developed by DeepMind for AlphaGo. The system did not rely solely on a neural network producing an immediate answer, but on successive inference steps exploring multiple scenarios before each decision.

According to the British startup, large language models are now evolving in this direction. “Complex intellectual work involves many sequential steps, each dependent on the previous one,” the company explains, viewing reasoning models as a first step toward systems capable of maintaining long and structured analytical chains.

The main technical bottleneck identified by Fractile concerns memory bandwidth. The company believes current architectures are not progressing quickly enough to absorb the increasing demands created by long-context models and reasoning systems.

“To compress a month of computation into a single day, execution speeds would need to reach around 1,200 tokens per second while handling the complexity and capacity constraints of large models operating on very long contexts,” the company said.

To address this challenge, Fractile is working across the entire technological stack: microarchitecture, system design, manufacturing processes, and hardware optimisation. This vertically integrated approach places the company alongside players such as Cerebras Systems and Groq.

The battle around inference has become one of the AI industry’s main industrial fronts. Several groups are seeking to reduce their dependence on traditional GPU architectures dominated by NVIDIA. AMD, Google, Amazon Web Services, and Intel are accelerating investments into AI accelerators, while startups such as SambaNova Systems, Etched, Tenstorrent, and d-Matrix are developing specialised architectures for reasoning workloads and AI agents.

Europe is also attempting to maintain a presence in this strategic infrastructure layer. In France, SiPearl is developing processors for European supercomputers, while Kalray is working on parallel processing architectures adapted to massive data flows and AI use cases. Scaleway and Mistral AI are also contributing to the emergence of a European compute and inference infrastructure. In the United Kingdom, Graphcore remains one of the main industrial precedents in this segment despite commercial difficulties against NVIDIA.

Fractile nevertheless believes the stakes extend well beyond current generative AI use cases. “The workloads currently pushing the frontier are already transformational. Those beyond that frontier will redefine the entire economy,” the company argues.

The company is currently hiring in London, Bristol, San Francisco, and Taipei.