Fractile is a London company established in 2022 that makes AI inference chips. It just raised $720 million in funding, valuing it at $7 billion post-money.

It was valued at about $1.1 billion during its last round in May this year. It has more than quintupled in half a year and is now one of the most highly valued startups in the UK. Investors include a number of Silicon Valley venture capital firms, Google’s independent venture capital arm, and the British government’s sovereign AI fund.

Its technical route is to directly stack and integrate customized memory and computing units on the same chip.

This is done to solve an old problem: during model inference, the chip calculates very quickly, but the speed of moving data from memory cannot keep up, and the computing power spends a lot of time waiting for data. It's called a "memory wall" in the industry. By stacking memory and computing together, the transport distance is shortened, which can theoretically speed up the process significantly. Fractile’s own expectation is that the reasoning speed will increase by 10 times.

Anthropic has expressed interest in becoming one of the first customers. For a company that worries about computing power every day and has just signed a contract worth hundreds of billions of dollars with SpaceX, having one more inference supplier will give it more bargaining power.

But let’s look at it calmly: Fractile’s chips have not yet been mass-produced, and the first commercial product is not scheduled to be launched until next year. The $7 billion valuation is largely based on prototypes and expectations.

There is already a long queue of inference chip companies that want to challenge Nvidia, and very few have truly crossed the three hurdles of mass production, yield and software ecology. The release of the product next year will be the company's first real test.