Groq, the AI-inference chip company that spent the better part of a decade pitching itself as the scrappy alternative to Nvidia, has raised fresh money at a valuation that quietly concedes how much has changed.
The company announced on Monday that they took in $350m at a $3.5bn valuation, roughly half the $6.9bn it commanded last September.
The round was led by Disruptive, the Dallas firm whose founder Alex Davis is now Groq’s executive chairman, and, in a twist that would be difficult to invent, Nvidia itself joined in.
Only months earlier Nvidia had licensed Groq’s technology and hired away much of its talent, an episode we covered when the company first set about picking up the pieces.
Late last year Nvidia struck a non-exclusive licensing agreement for Groq’s language-processing-unit technology, a deal widely reported at around $20bn and just as widely described as a “not-acqui-hire”.
There was no outright purchase of the company. Instead Nvidia secured the intellectual-property rights it wanted and walked off with founder and chief executive Jonathan Ross, a former Google engineer who had helped build that company’s tensor chips, together with a good part of the senior team.
What remained needed a new plan and new people. Co-founder Doug Wightman stepped up as chief executive, a fresh bench of executives was recruited, and Groq repositioned itself less as a chip designer squaring up to Nvidia and more as a data-centre operator selling AI inference by the token.
The stated ambition now is a “neocloud” that pushes capacity beyond 200 megawatts within a year.
The inference business it leans on, a cloud that already serves millions of developers and processes trillions of tokens a week, came with the pivot rather than the departed chip team.
A $650m raise in June was the opening act of that reconstruction, and this $350m round is the second.
The optics of the company that emptied the building now helping to refurbish it are peculiar even by the standards of the AI-chip boom, where allegiances are fluid and almost everyone is somehow both a customer and a competitor at once.
It is the same tangle visible across the sector, where Nvidia’s grip has grown so complete that would-be challengers increasingly find it easier to partner than to fight. Backing Groq costs Nvidia very little, and it buys a friendly, dependent supplier of inference capacity, plus a stake in whatever the rebuilt firm turns out to be.
Still, the numbers deserve a raised eyebrow. Halving a company’s paper worth inside a year is a brisk correction in a market where valuations have mostly travelled in one direction, and where newer entrants keep minting billion-dollar price tags on the strength of inference demand alone.
Rival inference-chip startups such as Fractile have raised at buoyant valuations, and one London challenger recently tripled its worth to $3.3bn while betting openly against Nvidia.
Groq’s discount is a useful reminder that the gold rush prices in founders and engineers as much as silicon, and that losing both carries a bill.
Whether $350m is enough to matter is the open question. Groq still runs a genuine inference business, with data centres across several continents and developers who use it precisely because it is fast and cheap, and demand for inference is not in doubt.
What is less certain is whether a rebuilt Groq, shorn of the founder who defined it and now part-funded by the giant that hollowed it out, can be much more than a comfortable supplier orbiting Nvidia.
The valuation suggests investors have already made their peace with the smaller ambition. At $3.5bn, they are paying for a going concern, not a giant-killer.
Get the TNW newsletter
Get the most important tech news in your inbox each week.
