Building the next generation of inference hardware

At Felicis, we've always been drawn to founders who see around corners.

Those who make a bet on a future that isn't yet obvious, and then go build the hardest possible version of it. That's exactly what the team at Fractile has been doing since they founded the company in 2022. Today, I'm proud to share that Felicis has invested in Fractile's series B.

The Insight: Inference is the Bottleneck of the AI Era

When Fractile was founded, the dominant conversation in AI was about training - who had the most compute, the biggest datasets, the most parameters. Fractile made a different bet: that the real frontier of AI value creation would ultimately be defined not by how powerful models are, but by how fast and economically they can run.

They were right.

We are already in a world where the leading AI systems can tackle problems of extraordinary complexity, like drug discovery, software engineering, scientific research, but only if we can afford to let them think long enough to do so. Today's frontier models are generating up to 100 million tokens to solve the hardest problems. At the ~40 tokens per second typical of current hardware, a single output of that length takes a month to complete. That is not a software problem. That is a hardware problem.

Fractile was built from day one to solve it.

A Full-Stack Hardware Moonshot

What makes Fractile so compelling is the scope and conviction of what they're building. This is a ground-up reinvention of infrastructure. The team is working across the entire stack (foundational AI research, chip micro-architecture, and foundry process innovation), all in service of one goal: compressing that month of inference time into a day, or that weekend of lab computation into a coffee break.

To get there, they need to achieve roughly 1,200 tokens per second at scale — roughly 30x where the industry is today — while handling the context complexity that long-sequence reasoning demands. That's a moonshot. And it's exactly the kind of ambitious, technically grounded moonshot that Felicis was built to back.

Why Now, Why Fractile

From my first conversation with Fractile founder and CEO Walter Goodwin, it was clear that this team had a rare combination of deep technical credibility across hardware, AI systems, and semiconductor design, coupled with a genuine clarity of vision about where the industry was heading. They understood, years before it became consensus, that inference speed and unit economics would become the defining constraint on what AI can actually accomplish in the world.

The team is based across London, Bristol, San Francisco, and Taipei — a reflection of where the world's best hardware talent lives, and a signal of how seriously they're executing on a global scale.

The World Fractile is Building Toward

The most exciting thing about Fractile is the world they are unlocking.

When inference becomes fast enough and cheap enough, we won't just do today's AI tasks faster. Entirely new categories of work become economically viable. Agentic systems that engage in extended chains of reasoning. Drug discovery pipelines that can explore vastly more hypotheses. Software engineering at a scale we can't yet imagine. The parallel to how the internet didn't just digitize existing commerce but created entirely new industries is apt here.

Andrew Wiles spent years working toward a proof of Fermat's Last Theorem — filling reams of paper with working, dead ends, and fruitful directions — before a single insight connected everything. Frontier AI is beginning to operate in a similar mode: long, sequential, deeply exploratory chains of inference. Fractile is building the hardware that makes that kind of thinking not just possible, but economically deployable at scale.

Fractile is early in its journey, and there is a great deal of work ahead before their first chips and systems reach customers. But that's exactly where we thrive, alongside exceptional founders at the moment the hard work begins.

If you are a builder working on inference hardware, chip architecture, or AI systems, and you want to work on one of the most ambitious and consequential technical challenges of our time, Fractile is hiring across the UK, US, and Taiwan. I encourage you to reach out.

We couldn't be more excited to be part of this journey!