Nvidia’s $20B Groq Deal Faces DOJ Probe
Nvidia spent $20 billion on Groq without buying Groq. Now the DOJ is asking the obvious question: was that a deal, or a merger wearing a cheap disguise?
Nvidia spent $20 billion on Groq without buying Groq. Now the U.S. Department of Justice is examining whether that was a clever commercial arrangement or a merger wearing a cheap disguise.
That question should make every founder, investor and big-company operator sit up. Because the reverse acqui-hire — buy the brains, license the technology, leave a corporate shell behind — has become Silicon Valley’s favourite way to do a takeover without calling it one.
The deal everyone pretended was not an acquisition
Nvidia announced its arrangement with AI-chip startup Groq in December 2025 as a non-exclusive licensing agreement for Groq’s language-processing-unit technology. Groq remained independent. Nvidia did not buy its shares, customer contracts or existing products.
But Nvidia hired Groq founder and CEO Jonathan Ross, president Sunny Madra and other employees. The reported transaction value was about $20 billion — extraordinary money for an arrangement repeatedly described as something other than an acquisition. ([news.bloomberglaw.com](https://news.bloomberglaw.com/ip-law/doj-probes-nvidias-license-deal-with-groq-on-antitrust-concerns?utm_source=openai))
This week, reporting from Bloomberg and Axios said the DOJ is investigating whether Nvidia structured the agreement to avoid antitrust review. That matters because the transaction was not a couple of senior engineers leaving for a better gig. Nvidia secured rights to a credible inference-chip rival’s technology, recruited its leadership and paid a sum big enough to make most conventional takeovers look like a Friday arvo expense claim. ([axios.com](https://www.axios.com/2026/09/10/doj-nvidia-groq-antitrust?utm_source=openai))
Let’s not get carried away: an investigation is not a finding of illegality. Nvidia has not been accused in public filings of breaking the law, and the company declined to comment to Bloomberg on the inquiry. But the commercial reality is far more interesting than the legal label.
If you remove the founder, president, key technical people and the economic upside, what exactly is left of the competitive threat?
That is the question the DOJ is asking in its own language.
Nvidia’s filings make the accounting look even stranger
Nvidia’s annual report gives the cleanest view of what it actually obtained. The company disclosed that it entered the Groq licence in December 2025 and hired certain Groq employees. It said it bought no customer contracts, existing products or equity interests.
Then comes the bit that should stop people calling this a harmless talent deal.
Nvidia recorded $14.4 billion of goodwill and a $2.5 billion developed-technology intangible asset related to the arrangement. It disclosed $13 billion paid at closing and another $4 billion, including imputed interest, payable within one year. ([sec.gov](https://www.sec.gov/Archives/edgar/data/1045810/000104581026000038/a2026-annualxreportxwebxfi.pdf?utm_source=openai))
The reported $20 billion headline and Nvidia’s disclosed consideration figures are not identical measures. Fine. Accountants can have their fun. But for an operator, the important point is brutally simple: Nvidia put tens of billions of dollars behind Groq’s people and technology while avoiding an ordinary purchase of the company itself.
That is not semantics. That is strategy.
Nvidia has since made clear why Groq mattered. On its August 26 earnings call, it said it was combining Nvidia’s high-throughput systems with Groq’s high-interactivity architecture, and said its first rack-scale Groq LPU system was already in full production. Nvidia expects volume shipments to early adopters later this quarter. ([investor.nvidia.com](https://investor.nvidia.com/files/content_files/TRANSCRIPT_-NVIDIA-Corp-NVDA-US-Q2-2027-Earnings-Call-26-August-2026-5_00-PM-ET.pdf?utm_source=openai))
In plain English: Nvidia did not pay this sort of money to put a licence certificate in a filing cabinet. It bought speed in AI inference — the part of the market where models answer real user requests — and it bought it from one of the few companies trying to offer something materially different.
Why Groq was worth $20 billion to Nvidia
Most people still talk about AI chips as though the whole contest is about training giant models. That was yesterday’s fight.
Training is expensive, glamorous and easy to put in a keynote. Inference is where the product meets the customer. Every time someone asks an AI assistant a question, generates code, transcribes a call or runs an agent through a business workflow, somebody pays for inference.
Groq built specialised language-processing hardware aimed at serving models quickly. Nvidia already owns the broad AI-compute ecosystem, from chips and networking to its CUDA software moat. Groq offered Nvidia another angle: high-speed, interactive inference.
The company had real credibility before the deal. Groq had raised roughly $3.3 billion since its 2016 founding, including a $750 million round at nearly a $7 billion post-money valuation in 2025. Its investors received substantial payouts through the Nvidia arrangement even though no equity changed hands. ([axios.com](https://www.axios.com/2025/12/28/nvidia-groq-shareholders?utm_source=openai))
After the deal, Groq did not disappear. It raised $650 million in June 2026, then another $350 million reported in August, as it pivoted toward becoming a neocloud provider rather than simply a chip challenger. Its August financing valued the post-deal company at $3.5 billion, down from the $6.9 billion valuation attached to its prior major round. ([techcrunch.com](https://techcrunch.com/2026/06/22/ai-chipmaker-groq-confirms-650m-raise-re-staffs-after-nvidias-20b-not-acqui-hire-deal/?utm_source=openai))
That survival is Nvidia’s best defence in the court of public opinion. Groq still exists. It can still use its own technology. The licence was non-exclusive. New investors backed it.
But don’t confuse legal survival with competitive equivalence. A startup can remain alive on paper while losing precisely the people and assets that made it dangerous.
The overlooked angle: this is really about price discovery
Everyone will frame this as a competition-law story. It is. But it is also a capital-markets story.
In a normal acquisition, everyone sees the price, the target, the buyer, the assets and the regulatory process. Competitors can object. Customers can assess concentration risk. Employees can decide whether they want to stay. Investors can work out what a category leader is actually worth.
Reverse acqui-hires muddy all of that.
The buyer says it licensed IP. The target says it remains independent. Key people leave. Investors receive a giant payout. The remaining company gets recapitalised and finds a new story. It may all be lawful. Yet the market has still lost a clear, independently financed competitor in its original form.
That opacity is valuable to incumbents. It reduces the political drama of a takeover and can reduce the friction of getting a transaction done. It can also make it harder for an outsider to tell whether a startup was truly rescued, hollowed out, or both.
For Nvidia, that is useful because it is already the company everyone watches. Its scale means every investment, customer financing arrangement, licensing pact and talent raid receives a second reading. The larger you get, the less benefit of the doubt you deserve. That is not anti-business. That is the price of being dominant.
The contrarian view: regulators should not punish every clever structure
Here is the bit the anti-big-tech crowd will hate: the DOJ should not treat every non-traditional deal as a crime because it looks aggressive.
Founders deserve the right to sell technology. Employees deserve the right to take better jobs. Investors deserve liquidity. A company that has raised billions should not be forced to remain a standalone competitor forever just to satisfy a regulator’s preferred market diagram.
And Groq’s post-deal fundraising shows there is a serious argument that the company has a second act. It is building cloud capacity, says it plans to scale from 54 megawatts to more than 200 megawatts in 2027, and Nvidia is participating in its latest financing plans. ([techcrunch.com](https://techcrunch.com/2026/08/17/groq-raises-350m-to-fuel-its-pivot-from-ai-chips-to-neocloud/?utm_source=openai))
But that is exactly why the DOJ’s scrutiny is healthy.
The test should not be: “Did Nvidia technically purchase shares?” That is lawyer-brain nonsense. The real test is whether the arrangement transferred so much of the target’s technology, leadership and economic value that it functionally removed a competitor without the scrutiny an outright purchase would attract.
If the answer is yes, regulators need a way to assess it. Not ban it automatically. Assess it properly.
What this means for you
If you are a founder, stop treating structure as an afterthought. The way you sell can matter almost as much as the price. A licensing-plus-talent transaction may deliver liquidity without a conventional acquisition, but it can also leave the remaining company with an awkward identity, a depleted leadership bench and investors expecting miracles from what is left.
Before signing one, ask four questions:
1. Who is actually staying? Not names on an org chart — the people who know why the product works. 2. What rights are you licensing away? “Non-exclusive” means bugger all if the buyer gets your best people, your roadmap and the commercial advantage. 3. Can the remaining business raise money on its own story? Groq did. Many won’t. 4. Would you be comfortable explaining the deal’s real economics to customers, staff and regulators? If the answer is no, the structure is probably too cute.
If you are an investor, do not mistake a giant payout for a clean outcome. Work out what remains after the founder and core technical leadership walk. The headline cheque can be spectacular while the residual equity turns into a very expensive lesson in corporate leftovers.
And if you run a large company, learn the deeper lesson from Nvidia: buying capability is often faster than building it. But don’t assume clever deal architecture makes the strategic risk vanish. It can simply move the risk from product execution to regulators, employees and public trust.
That bill always turns up eventually. Usually with more zeroes on it.