Fractile's in-memory-compute bet: $15M seed to a $220M Series B for faster AI inference

London chip startup Fractile bets inference, not training, is AI's bottleneck: compute inside memory for 100x faster models — $220M Series B in 2026.

Fractile

它在赌什么AI progress is gated by the cost of running models, not training them — an in-memory-compute chip can run LLM inference 100x faster and 10x cheaper.在做还没上线

做的是什么生意

Designs AI accelerator chips that run model inference with computation baked into memory, avoiding the constant shuffling of model weights.

启动资金：$15M seed (July 2024), $17.5M total at launch; $220M Series B (May 2026)

起因

Walter Goodwin, a 28-year-old AI PhD, founded Fractile in 2022 in London. He saw that every large AI company depends on near-identical chips built for training, while running trained models (inference) stays slow and expensive — a gap that widens as agents generate tens of millions of tokens.

新闻$15M seed launch (tech.eu)

经过

Fractile left stealth in July 2024 with a $15M seed led by Kindred Capital, the NATO Innovation Fund and Oxford Science Enterprises. It hired engineers from NVIDIA, ARM and Imagination and filed patents on its in-memory circuits. The pitch: execute roughly 99.99% of inference operations in memory, targeting 100x speed, 10x cost cuts and 20x TOPS/W.

结果

In May 2026 Fractile raised a $220M (£162.8M) Series B led by Accel, Factorial Funds and Founders Fund, with 8VC and Felicis among participants, to finish its first chips and scale teams in the UK, US and Taiwan. First commercial hardware had not yet shipped as of September 2026.

新闻$220M Series B (UKTN)

背景

Fractile is a London chip startup founded in 2022 by Walter Goodwin, an AI PhD who left academia at 28 to build hardware. His bet: AI's real constraint is not training ever-larger models but the cost and latency of running them, and that a fundamentally different chip architecture can break that wall.

Most AI accelerators optimize training. Fractile instead designed for inference using in-memory compute: computational operations are baked into memory itself, so model parameters never need shuttling between memory and processor. The company targets 100x faster and 10x cheaper inference, plus 20x better performance per watt, while staying compatible with unmodified standard silicon foundry processes.

Fractile exited stealth in July 2024 with a $15M seed led by Kindred Capital, the NATO Innovation Fund and Oxford Science Enterprises, with angels including Hermann Hauser. In May 2026 it raised a $220M Series B led by Accel, Factorial Funds and Founders Fund, to bring the first chips and systems to market and scale teams in the UK, US and Taiwan.

The thesis is still unproven at scale: as of September 2026 no Fractile chip had shipped to a customer, and the company must now execute a full silicon tape-out against an industry where NVIDIA keeps raising the bar.

这件事要成立，得有什么

Serving a trained model, not training it, is where AI's cost is exploding as agents generate tens of millions of tokens.
In-memory compute removes the dominant cost of inference: moving model weights between memory and processor.
Staying on standard foundry processes means the radical architecture can still be manufactured by existing fabs.
The same talent pool that builds training chips was pointed at a different target, letting a small team differentiate on architecture rather than scale.

可借鉴之处

Bet on the bottleneck everyone ignores: the industry optimized training chips while the cost of running models became the wall — the startup attacking the next constraint sets the pace.

后续进展

As of September 2026 Fractile is spending the Series B proceeds to accelerate delivery of its first accelerator chips and systems and to grow engineering across the UK, US and Taiwan. No commercial product has shipped yet, so the in-memory-compute thesis remains unproven; the bet stands or falls on a successful tape-out and real-world inference benchmarks.

资料来源

发现哪里写错了？告诉我们。

轮到你了

你刚读完一家。说说你在做什么，看看谁在赌同一件事。

免费账号 · 3 次免费提问 · 不用绑卡