"Post-Nvidia-licensing-deal version of Groq." That is how a Groq spokesperson described the $3.5 billion valuation the company just took — down roughly half from the $6.9 billion it commanded last September — and, in the same breath, insisted it wasn't a down round.
Read that carefully. The chip startup that spent a decade pitching itself as the scrappy alternative to Nvidia just raised $350 million at half its former worth, with planned participation from Nvidia itself.
The same Nvidia that reportedly licensed Groq's core LPU IP for $20 billion, hired away founder and CEO Jonathan Ross, president Sunny Madra, and a good chunk of the senior team — then announced its own "Nvidia Groq 3 LPX inference hardware system" at GTC in March, reportedly built on that licensed IP.
Nvidia gutted the building, and now it's writing a check to help refurbish it. The spokesperson would like you to know this isn't a down round.
It is a down round. But calling it that undersells what actually happened, because the 49% haircut is not a valuation error being corrected. It is a price, cleanly set, for a specific thing: everything Nvidia extracted.
The 49% haircut is not a valuation error being corrected. It is a price, cleanly set, for a specific thing: everything Nvidia extracted.
Here is the more useful way to read the number. At $6.9 billion, investors were paying for a giant-killer — a founder-led team building silicon to beat Nvidia at inference. At $3.5 billion, they are paying for a going concern that resells Nvidia's silicon. The difference between those two figures arguably reflects, in this author's reading, the market-cleared value of Jonathan Ross, his engineers, and the ambition they carried out the door.
The gold rush prices in founders and engineers as much as it prices in silicon. Losing both carries a bill, and by this author's own arithmetic that bill is around $3.4 billion.
Now the mechanism, because this is where "pivot" stops being the right word and "template" starts being the right one.
Axios laid out the structure back in May, when the first act — a $650 million round — was still being raised. Groq's investors "already got paid out," with a final cash distribution incoming, via the $20 billion Nvidia deal. Then those same investors were asked to "participate pro rata in Groq 2.0."
Existing backers Disruptive and Infinitum agreed to backstop the round if it didn't fill. The bottom line, per Axios: "Don't be surprised if this becomes a new transaction template in the AI private markets."
Sit with the elegance of that. The VCs got their exit — Groq was described as one of the biggest AI wins for venture so far. Then they were invited to roll a slice of those proceeds back into the emptied shell, which now operates as a customer of the very company that emptied it.
Nvidia converts a rival into a supplier of demand. The VCs get a second bite. And Nvidia's own check into Groq 2.0 costs almost nothing relative to its $5.4 trillion market cap — arguably a rounding error that buys what looks like a friendly, dependent reseller of inference capacity.
That is not a rescue. A rescue is unplanned. This was designed: pay out, poach, re-fund the husk, book the husk as an NVIDIA Cloud Partner. Every step deliberate.
And the reseller status is now formal. Groq's own release describes it as a certified "NVIDIA Cloud Partner (NCP)," authorized to "design, deploy, and operate NVIDIA accelerated computing to NVIDIA's reference architecture and operational standards."
The fresh capital, the company says, supports customers "seeking usage of medium and larger sized clusters of NVIDIA accelerated computing for training and inference."
Read that as a mission statement and the arc is complete. Groq began building LPUs to beat Nvidia on inference. Groq now sells clusters of Nvidia accelerators, to Nvidia's spec, with Nvidia on the cap table.
The company that was built to be the alternative arguably now looks like a distribution channel.
Is this unique? No — and that's the point. Nvidia supplies the GPUs powering CoreWeave, Lambda, and Nebius while also investing billions into some of them. The whole neocloud layer runs on Nvidia silicon and, increasingly, Nvidia money.
Nvidia is also the most active corporate investor in the sector, participating in a record 59 known funding rounds this year, leading or co-leading at least 11. When you supply the picks and fund the miners, "competitor" becomes a temporary condition.
But Groq is the sharpest case precisely because it started furthest from Nvidia's orbit and ended deepest inside it. CoreWeave was always a customer. Groq was a rival with its own chip architecture and a founder who'd helped build Google's TPU. It had the pedigree to actually threaten the franchise.
And the franchise didn't crush it. It absorbed it — bought the IP, hired the brains, then re-funded the leftovers as a customer. That is a more complete victory than crushing, because the leftovers now generate Nvidia revenue instead of Nvidia risk.
The open question was whether chipmakers could ever claw pricing power back from Nvidia. Groq answers it from the reverse angle. What was arguably the most credible inference-silicon challenger of the last decade didn't claw anything back. It got converted into a customer, and its investors thanked Nvidia on the way through.
Which brings us to the number's real audience: everyone still trying to build inference silicon.
Rival startups are still minting buoyant valuations on inference demand alone. Fractile has raised at strong marks. One London challenger recently tripled its worth to $3.3 billion while betting openly against Nvidia.
Groq's $3.5 billion is the counterweight to all of it. It is the market pricing what a de-founded giant-killer is worth once you subtract the people — and the answer is: half, and pointed straight into Nvidia's ecosystem.
For a founder weighing whether to fight Nvidia or partner with it, the Groq template makes "partner" look less like surrender and more like the rational default. Nvidia's gravity is now strong enough that the exit ramp is paved and lit.
The open question is who rides it next. When Nvidia comes for Fractile, or for the $3.3 billion London bet, does the founder hold the line — or take the "not-acqui-hire," let the VCs cash out, and hand back a husk to be re-funded as an NVIDIA Cloud Partner?
The honest answer is that the conveyor is now built, tested, and profitable for everyone who boards it except the original mission. Scale AI survived its Meta version of this and is reportedly on track for $1 billion in revenue, so survival isn't the question.
The question is what survives. Groq lives. Groq's data centers hum, serving six million developers and trillions of tokens a week. But the thing Groq was for — an actual alternative to Nvidia — is the part that didn't make it.
At $3.5 billion, investors have made their peace with the smaller ambition. They are paying for a going concern, not a giant-killer. And Nvidia, having set the price, is happy to chip in.