Fractile Raises $720M at a Reported $7B for an Inference Chip That Has Not Shipped Yet

On Wednesday, 11 November 2026, Fractile, a chip startup based in London and Bristol, announced a $720 million Series C. TechCrunch and the Korean outlet Wowtale both report that Thrive Capital, Founders Fund, Lightspeed Venture Partners and Redpoint co-led the round. The UK government put money in through its Sovereign AI fund and the British Business Bank, and it announced the deal on gov.uk the same day. The government said the round was worth “over £500 million” and valued the company at “over £5 billion”. Reuters and the Financial Times, citing people involved, put the valuation at about $7 billion.

That is roughly seven times what the company was reportedly worth in May, when it closed a $220 million Series B led by Accel, Factorial Funds and Founders Fund at a valuation of about $1 billion. Adding up the seed ($15M, July 2024), Series B and Series C gives about $955 million raised by a company that has not yet sold a chip. CEO Walter Goodwin wrote on LinkedIn that the money will let Fractile “ship [its] first chips and systems next year.”

The technical bet: memory bandwidth, not FLOPs

When a large language model generates text, each new token needs most of the model’s weights, plus the growing KV cache, to be read out of memory. At small batch sizes, the speed of that read decides how fast tokens come out. The arithmetic units mostly sit idle. That is why single-stream decoding on a GPU is said to be memory-bound, and it is the problem Fractile was set up to attack.

Today’s accelerators offer two poor choices. On-chip SRAM is extremely fast but holds only hundreds of megabytes per chip, so a frontier-sized model has to be spread across racks of chips. Groq and Cerebras built their designs around this. HBM (high-bandwidth memory, stacked DRAM placed next to the GPU die) holds the whole model but limits bandwidth per user. Fractile says it packages DRAM directly on top of logic, co-designed in 3D, to get “higher memory bandwidth than HBM alongside more capacity than SRAM.” In plain terms, it wants HBM’s capacity with something closer to SRAM’s bandwidth, by shortening the distance between the bits and the multipliers.

That is a change from how the company was first described. Its 2024 seed coverage and a DatacenterDynamics profile said it was building in-memory compute: doing the matrix multiplications inside SRAM cells. TechCrunch now says the design “revolves on DRAM” after a pivot. Fractile has not published an architecture paper, die photos, bandwidth figures or process details beyond saying its production parts will use standard FinFET foundry nodes. So the exact design is still unknown from the outside.

The numbers behind the pitch

Goodwin’s May essay states the case with a simple calculation. He says frontier models already produce “up to 100 million tokens” on hard problems, and existing chips serve them at about 40 tokens per second. At that rate, 100 million tokens take roughly 29 days. At the ~1,200 tokens per second Fractile is aiming for, the same job takes about 23 hours, or “compressing that month into a day.” The arithmetic holds. The premise depends on agentic workloads that run that long, which only a few labs run today.

The headline claims need more caution. Early on, Fractile said its hardware would run LLMs “100× faster and 10× cheaper” than Nvidia GPUs, with 20× better performance per watt. In the 2024 DCD interview, Goodwin admitted the designs had “only been tested in computer simulations.” TNW reported in May that newer investor materials had lowered the comparison to about 25× faster at a tenth of the cost. No benchmark against a deployed GPU cluster has been published, and Nvidia, AMD and Google will ship at least one more generation before Fractile’s first chip arrives.

Why investors paid up anyway

The jump in valuation came after a customer signed up. In August, Bloomberg reported that Anthropic had made an initial agreement to buy about $250 million of Fractile chips. The chips would not be usable until 2027, and there were plans to expand the deal. Both companies declined to comment. If the report is right, a $7 billion valuation rests mostly on an unreleased part and one very large customer.

The money is part of a wider rush into inference silicon. Etched, Positron, d-Matrix and the newly listed Cerebras have all raised large sums in 2026, and each tries to get past the memory wall in a different way. The interest makes sense: inference is now where AI labs earn revenue and where they spend most on serving. Every lab that depends on one GPU supplier wants a second option. For London, the round also has a political side. Ministers presented it as making sure the UK keeps “a greater stake in the hardware” rather than repeating the Arm story.

What it means for developers, and what to watch

If DRAM-on-logic delivers even part of what Fractile promises, serving costs would change. Long-running coding agents, deep-research jobs and big tree searches over model outputs are limited today by tokens per second per stream, not by total throughput. Faster single-stream decoding would make those workloads cheaper and quicker to iterate on, the same way faster CI changes how teams work.

The open questions are the usual ones for hardware. Will the first silicon reach its simulated numbers? Can stacked DRAM stay cool next to hot logic? Is there a compiler and software stack good enough that labs can port models without months of work? Fractile says it is building that compiler in-house. And can a startup get advanced packaging capacity in a market where HBM and CoWoS are already tight? The first chip is due in 2027. Until independent benchmarks exist, the $720 million is funding a plausible engineering idea and one customer’s confidence, not a proven product.

Sources