Profile of UK-based Graphcore, which makes “Intelligence Processing Units”, AI chips designed to be better at guesswork and have high energy efficiency
Austin Carr / Bloomberg : Tweets: @atomico , @graphcoreai , and @austincarr Tweets: @atomico : ‘Unlike the number-crunching alternatives, British startup @Graphcoreai has developed a brain for computers that excels at guesswork.’ https://www.bloomberg.com/... by @JeremyAKhan for @BW Graphcore / @graphcoreai : Worth a read. ‘@graphcoreai is developing a brain for computers that, if its co-founders are right, will be able to process information more like a human instead of faking it through massive feats of number crunching.’ https://www.bloomberg.com/... https://twitter.com/... Austin Carr / @austincarr : Why Microsoft, Samsung, and BMW are betting on Graphcore's A.I. chip https://www.bloomberg.com/...
Context & Ripple Effects
This 2019 Bloomberg profile captures Graphcore at peak promise: two years after its $30M Atomico-led raise put the first Intelligence Processing Units on a shipping schedule, the company had just closed a $200M round at a $1.5B valuation with Microsoft, BMW, and Sofina aboard. The pitch was architectural heresy — silicon tuned for probabilistic 'guesswork' and energy efficiency rather than brute-force number crunching.
First-order effects
- The corporate backers named in the profile were also its intended customers, and the bet failed on that axis: Microsoft, an investor, reportedly declined to deploy Graphcore's chips, leaving the company with roughly $5M in 2021 revenue against a $183M loss after raising $730M.
Second-order effects
- With merchant sales stalled, Graphcore's fallback became selling the company itself — sources in early 2024 pointed to sale talks with rumored buyers including OpenAI, SoftBank, and Arm at a valuation far below the 2018 peak, meaning backers like Atomico face steep markdowns while acquirers pick up IPU design talent cheaply.
Third-order effects
- The pattern points toward consolidation of bespoke AI-silicon startups into the large compute buyers and platform owners they once hoped to sell to — specialized-chip independence gives way to vertically integrated stacks, and UK-designed IP migrates to foreign owners if the reported deal shape holds.
The trend: Differentiated-architecture AI chip startups are following an arc from venture-backed challengers to acquisition targets for the very hyperscalers and platform companies their chips were meant to unseat.