You hired the smartest person in the world to help your business. They know everything. They can solve any problem. But every time you ask them a question, they take three weeks to answer.
That is not really useful, is it?
That is, in a very real sense, the problem AI is running into right now. The models are getting brilliant. But the hardware running those models is struggling to keep up. And a London-based startup called Fractile just raised $220 million to fix exactly that.
How Fractile Got Started
In 2022, a researcher named Walter Goodwin left Oxford University, where he had been studying robotics and computing systems, and started a company around one specific bet.
The bet was this: AI capabilities would eventually outpace the hardware designed to run them. The world would build increasingly powerful models, but those models would be bottlenecked not by intelligence, but by speed. Every query would take too long. Every output would cost too much. And at some point, that would become the single biggest constraint on how far AI could actually go.
Goodwin co-founded Fractile with Yuhang Song, and the two of them started building chips from the ground up to address that problem. No pivoting to software. No waiting for someone else to solve it. Just a direct, full-stack bet on new hardware.
Three years later, they have raised $220 million from some of the most respected investors in technology, including Accel, Founders Fund, and Factorial Funds.
The Strategy: Attack the Bottleneck at the Source
To understand what Fractile is building, you need to understand one thing about how AI chips currently work.
Right now, when an AI model generates a response, the processor and the memory sit in two separate places on the chip. Every single calculation requires data to travel back and forth between them. That journey, repeated billions of times per second, creates a traffic jam. Engineers call it a memory bandwidth bottleneck. In plain English, it means the chip spends a lot of time waiting.
Fractile's approach is to put the memory and the compute on the same chip, so data never has to travel. The calculation and the information it needs are right next to each other. Less waiting. Faster output. Lower cost per response.
The numbers that Fractile is targeting are striking. Today's leading chips generate around 40 tokens per second for advanced AI models. A token is roughly one word or part of a word. At that speed, running a very long and complex AI task, the kind that requires tens of millions of tokens, can take close to a month to complete.
Fractile is aiming for 1,200 tokens per second. That compresses that same month of work into a single day.
How Fractile Plans to Flip the Game
There is a moment in the history of mathematics that Fractile's own team references when they talk about the future of AI.
Andrew Wiles spent years working on a proof for Fermat's Last Theorem, one of the most famous unsolved problems in mathematics. At a critical point in his work, he realised that a direction he had been exploring that week connected perfectly with an approach he had tried and shelved three years earlier. The breakthrough came from the ability to hold an enormous amount of work in memory, revisit old paths, and synthesise across time.
Fractile's founders believe that is exactly what the most powerful AI systems of the future will need to do. Not just answer one quick question. But reason across long chains of thought, explore dead ends, revise their thinking, and eventually arrive at answers to genuinely hard problems, in drug discovery, in materials science, in software engineering.
That kind of work, done at the frontier of what AI can achieve, requires output in the tens of millions of tokens. And on today's hardware, the time and cost of producing that output are simply not viable.
That is the problem Fractile is selling. And at $1 billion in valuation, investors clearly believe the problem is worth solving.
Lessons for Founders
While there’s no guarantee if Fractile will actually make it not, there’s a lot to learn from their startup model.
Ask where the real constraint is, not the obvious one.
Every growing market has a visible problem and an invisible one. In AI right now, the visible problem is model capability. Everyone is racing to make smarter models. The invisible problem is inference speed and cost. Fractile found the invisible bottleneck and decided to own it. Whatever space you are building in, the best opportunity is often the constraint that everyone feels but nobody is directly solving.
Deep tech is slow. That can be a feature, not a bug.
Fractile's first chips will not ship until around 2027. That is five years after founding. For most software startups, that would be a red flag. But for hardware companies, that timeline is a moat. The technical difficulty that makes it hard for Fractile to ship is the same difficulty that makes it hard for anyone else to copy them. If your product is genuinely hard to build, the runway between your start and your competition's catch-up is often longer than it looks.
Infrastructure bets compound over time.
Fractile is not building a product that users will ever see directly. Nobody will open a Fractile app. But if their chips power the AI tools that millions of people use every day, the value they capture can be enormous. Think about this for your own business:
- Where in your market does the infrastructure sit?
- Who owns the layer that everything else runs on?
Those positions tend to be durable in ways that front-end products often are not.
One Thing to Remember
The AI race is now about who can run that model fast enough, cheaply enough, and reliably enough for the real world to depend on it. Fractile is betting that the hardware layer is where that race will actually be decided. Whether they win or not, the insight is worth keeping: in any technology wave, the picks and shovels can matter just as much as the miners.
