Analysis
The Justice Department is investigating whether Nvidia structured its nearly $20 billion agreement with AI chip startup Groq specifically to avoid a formal antitrust merger review, Bloomberg reported September 10. The DOJ has sent Nvidia a formal demand for information, according to the report, and the inquiry began shortly after Nvidia announced the deal in a December 24, 2025 filing CNBC called the largest transaction in Nvidia's history.
The deal itself was built to look unlike a traditional acquisition. Rather than buying Groq outright, Nvidia agreed to pay close to $20 billion for a non-exclusive license to Groq's Language Processing Unit inference technology, while separately hiring founder and CEO Jonathan Ross, president Sunny Madra and other senior leaders directly onto Nvidia's own payroll. Groq itself continues on as an independent company under new CEO Simon Edwards, the firm's former CFO -- a structure that let Nvidia absorb Groq's most valuable asset (the technology) and its most valuable people (the founding team) while leaving behind a corporate shell that, on paper, was never acquired at all.
Why structure matters to antitrust law
Under the Hart-Scott-Rodino Act, a traditional acquisition above a certain size threshold triggers a mandatory pre-merger review, giving the DOJ or FTC a chance to block or condition a deal before it closes. A pure licensing-and-hiring arrangement doesn't clearly meet the legal definition of a reportable transaction, even when the practical effect -- eliminating a competitor's independent technology roadmap and absorbing its leadership -- looks identical to an acquisition. Regulators have grown skeptical of this structure across tech: Microsoft's 2024 licensing-and-hiring deal with Inflection AI drew similar scrutiny from the UK's Competition and Markets Authority and the FTC, both of which ultimately declined to formally block it but signaled the structure itself remained a live legal question, not a settled one.
Nvidia's position in AI chips makes this specific deal harder to wave through than Inflection's was. The company controls the overwhelming majority of AI training-chip compute and has drawn antitrust attention on multiple fronts in 2026, including Chinese regulators examining Nvidia's chip sales under anti-monopoly law and EU officials scrutinizing its CUDA software lock-in. A formal information demand -- not just a preliminary inquiry -- is a materially more serious posture than the reviews Nvidia has faced on prior, smaller deals.
What Nvidia actually got
Groq's Language Processing Unit, a single-core chip architecture built for fast, deterministic large-language-model inference, was widely regarded as the most credible non-GPU alternative to Nvidia's own inference stack -- Pulse covered the deal's original benchmarks and terms when it was announced. Groq was founded in 2016 by Jonathan Ross, who previously led the original design of Google's Tensor Processing Unit, and had raised more than $2 billion across seven prior rounds before the Nvidia deal.
The nearly $20 billion Nvidia paid values that technology at roughly 2.9 times Groq's most recent private mark, a premium reflecting both the license's scope and the fact that Nvidia was removing its most direct inference-chip challenger, not just buying a stake in it. Groq's closest peers in inference chips illustrate why Nvidia cared enough to pay that premium:
- Cerebras -- founded the same year as Groq (2016), builds wafer-scale chips with roughly four trillion transistors per die; raised $1.1B in 2025 at an $8.1B valuation, and is now Groq's closest surviving independent rival.
- SambaNova -- founded in 2017, builds reconfigurable dataflow chips with a tiered SRAM/HBM/DRAM memory architecture; raised $676M in 2021 at a $5.1B valuation and closed a further $1B round in July 2026.
The remaining, Ross-less Groq entity has already raised $650 million of its own capital to keep competing in inference chips without the technology and leadership Nvidia just absorbed -- a detail that undercuts any argument this deal left the competitive landscape unchanged.
What the probe doesn't resolve
None of this means the DOJ will ultimately act. Regulators can and often do close antitrust inquiries without enforcement, and Bloomberg's report notes officials do not currently expect the deal itself to be unwound even if wrongdoing is found -- a fine is the more likely outcome in that scenario. Nvidia has not issued a public response to the DOJ's information demand as of publication, and Groq's own statements since the deal closed have focused on its go-forward roadmap rather than the antitrust question.
For chipmakers and AI labs watching from the sidelines, the practical signal is about deal architecture, not just this one transaction: a licensing-and-acquihire structure that avoids a formal merger filing is no longer a safe harbor from antitrust attention by default, and any company structuring a similar deal to route around HSR review should now expect the DOJ to ask exactly the question it is asking Nvidia.
