AI chip startup Fractile raises $720M Series C to build chips with 20x memory bandwidth advantage over GPUs

Fractile raises $720M Series C at $7B valuation to challenge Nvidia with stacked-DRAM chips

Fractile, a chip startup founded in 2022, has closed a $720 million Series C at a $7 billion valuation, co-led by Thrive Capital, Founders Fund, Lightspeed, and Redpoint, with participation from the UK Sovereign AI Fund. The round is funding a full-stack bet: custom silicon, custom software, and power-dense racks designed to run frontier AI models faster than any GPU on the market.

“Fractile's been on this mission for four years now. We've been building chips to run the world's largest models faster — orders of magnitude faster than any other chip on the planet. On this first-gen chip, we're projecting 20 times more bandwidth than you would get to HBM on a GPU or XPU. The capital lets us paralyze a bunch of things and move faster on production capacity planning for second half of 2027 and first half of 2028.”

The architecture

The core technical claim is a 20x memory bandwidth advantage over GPUs, achieved by stacking DRAM directly on top of the processor die rather than placing it alongside as a separate piece of silicon. Walter Goodwin says Fractile has been quietly developing this stacking approach for the past few years and believes no one else is doing anything comparable.

The logic behind the architecture is straightforward. Transformer models at every scale are increasingly bottlenecked by how fast weights and KV cache state can be loaded from memory. By designing around that constraint rather than around a specific model shape, Fractile is making a bet on what Goodwin calls a "macro property" of AI workloads — one that should remain true regardless of how architectures evolve.

That framing is a deliberate rejection of the hyper-specialization approach taken by earlier AI chip challengers. Chips take months through the fab, then need to be useful to customers for three to five years to satisfy amortization requirements. Locking in a specific workload at design time is a risk Goodwin argues doesn't pay.

Manufacturing and timeline

Capital from the round is being deployed to secure manufacturing capacity now for production in the second half of 2027 and into 2028. Goodwin says Fractile teams are already on factory floors in Taiwan and working directly with the supply chain to build what he describes as fundamentally new building blocks in DRAM stacking.

The cybersecurity use case

Goodwin frames raw inference speed as a national security-relevant capability. His argument centers on test-time compute: a cyberattack model that generated 100 million tokens at 50 tokens per second on a GPU would take 23 days. At 1,500 tokens per second — closer to what Fractile targets — the same run takes roughly a working day. That asymmetry matters both for red-teaming AI systems before release and for defensive counter-agent scenarios, where response speed is the competitive variable.

He also points to formal verification at the hardware level as an area where AI-assisted tooling is delivering returns, particularly for identifying safety-critical bugs in silicon.

Competitive positioning

Goodwin cites Anthropic's IPO materials, which noted that onboarding a new chip used to take six to twelve months and recently dropped to three days, as evidence that software-layer friction is falling. His read is that as optimization commoditizes, the hardware's fundamental properties matter more, not less. A chip with a structural bandwidth advantage becomes easier to unlock, not harder, as tooling improves.

Fractile is building across the US, UK, and Taiwan. The $7 billion valuation on a chip startup that hasn't yet reached volume production reflects investor conviction that the DRAM-stacking architecture is both defensible and timed correctly — with manufacturing slots for 2027–2028 now being the gating resource across the industry.