British startup Fractile's inference chips caught Anthropic's attention, which began preliminary conversations about a possible partnership. The product promises performance up to 100 times faster and costs up to 10 times smaller compared to Nvidia GPUs. The central idea is to overcome one of the main problems in the market: dependency on external memory, which makes language model processing more expensive and slows down.
At the moment, A large part of the execution time of these models is consumed in the constant transfer of data between the processor and external memory, which limits speed and increases costs. That's why, the industry begins to change focus: instead of just more power, the objective now is to generate responses more efficiently and at a lower cost per token. If the agreement progresses, Fractile could become Anthropic's fourth chip supplier, next to Nvidia, Amazon Trainium processors and Google TPUs.
Understand innovation
Founded in 2022 by Walter Goodwin, Fractile brings together engineers with experience in companies like Graphcore, Nvidia e Imagination Technologies. The startup develops a type of chip that keeps data and processing in the same place, avoiding the need to search for information in external components. Simply put, this reduces waiting time.
The logic is to bring the data as close as possible to where the calculations take place, reducing the internal “traffic” of the system. In simulations, This could dramatically speed up performance and reduce costs, but the technology has not yet been validated on a large scale and should reach the market in 2027.
Investments and pressure for efficiency
In your initial round, Fractile has already raised around US$ 15 with support from investors like Kindred Capital, NATO Innovation Fund e Oxford Science Enterprises. Now, a new round, could reach hundreds of millions of dollars, with the potential to raise its valuation to the “unicorn” level.
Anthropic is experiencing accelerated growth that puts pressure on its infrastructure. The company's revenue jumped from US$ 9 billion at the end of 2025 for about US$ 30 billion in March, significantly increasing your operating costs. In this scenario, Finding cheaper and more efficient ways to run your models is no longer just a future strategy and has become a necessity for the sustainability of operations.
