Nigel Toon has sold a company before, and watched a chip company get absorbed by a bigger one. That history is the quiet fact beneath Graphcore.
He began in chips at Altera, where he stayed about 13 years. In 2002 he co-founded Icera, a wireless chip firm that NVIDIA bought in 2011; reports put the price at $367 million or $435 million, depending on the source. He then ran Picochip and XMOS, and it was inside XMOS that the seed of Graphcore was incubated before the company stood alone in 2016.
Graphcore sold chips, but it was fighting NVIDIA, whose software ecosystem held customers tight.
So when Toon and CTO Simon Knowles started Graphcore in Bristol, they knew the rules. Build a chip designed for one job, machine learning. Skip the cache hierarchy that traditional processors lean on and give the chip its own large memory. Name the ambition after statistician I.J. Good's idea of an ultra-intelligent machine.
The money came quickly. A round led by Robert Bosch Venture Capital in autumn 2016. Atomico in July 2017, then $50 million from Sequoia. By December 2018 a $200 million round with Microsoft, Samsung and Dell Technologies valued it at $1.7 billion. In late 2020, a further $222 million lifted it to $2.8 billion.
What the headlines of those years hid was the revenue line. Graphcore sold chips, but it was fighting NVIDIA, whose software ecosystem held customers tight. In 2022 revenue was $2.7 million against a net loss of $205 million. Meta took Graphcore's AI networking team in early 2023. Then in July 2024 SoftBank agreed to buy the whole company for around $500 million, a fraction of the 2020 valuation. About 450 people worked there by 2024.
Toon is not a stranger to this part of the cycle. His earlier exits taught him that a deep-tech company can have the better engineering and still lose the market to the better ecosystem. Honoured with an honorary Doctor of Science by Bristol in 2023, he remains the face of British chip ambition.
The lesson is blunt. In hardware, a billion-dollar valuation is a promise, not a business. Customers buy software and habit as much as silicon, and a founder who has exited once should plan the second company around the thing that locked them out the first time.