In December 2025, Nvidia executed a massive $20 billion licensing and talent-acquisition deal with AI chip startup Groq. By absorbing key engineering staff and licensing inference-as-a-service technology without a traditional corporate merger, Nvidia bypassed premerger notification thresholds under the Hart-Scott-Rodino Act. Now, the U.S.
The Reverse-Acquihire Playbook Under Regulatory Scrutiny
The U.S. Department of Justice formally opened an antitrust probe into Nvidia’s arrangement with Groq, sending a formal demand for information. First reported by The New York Times and independently corroborated by Reuters, this inquiry marks the first time federal regulators have aggressively targeted the “reverse-acquihire” structure. This legal workaround has channeled over $100 billion across the broader AI industry while sidestepping mandatory federal review.
Under the Hart-Scott-Rodino Antitrust Improvements Act, corporate acquisitions exceeding the $119.5 million threshold require mandatory notification to the DOJ and the Federal Trade Commission, followed by a mandatory waiting period. Because Nvidia structured the late 2025 transaction as a non-exclusive technology license paired with a mass human-capital transfer—bringing Groq founder Jonathan Ross, president Sunny Madra, and senior engineering leadership directly into Nvidia—no corporate equity changed hands on paper. Nvidia maintained that the transaction bypassed standard merger triggers entirely.
Federal enforcers are testing a different legal framework. In March 2026, Acting Assistant Attorney General Omeed Assefi flagged structures designed to circumvent merger reviews as a regulatory red flag, explicitly pointing to the Nvidia-Groq arrangement. Legal scholarship from the Yale Law Journal and The University of Chicago Law Review has argued that reverse acquihires function as de facto asset acquisitions under the Clayton Act. When specialized know-how, proprietary intellectual property, and human capital transfer wholesale, the legal label of a “license” loses its regulatory shield.
Congressional pressure has followed closely behind. Senators Elizabeth Warren and Richard Blumenthal sent a direct letter to Nvidia CEO Jensen Huang in March 2026, writing: “By licensing its technology and hiring its most important employees, NVIDIA has effectively acquired Groq in all but name.”
Hardware Realities and the $17 Billion Silicon Transfer
What exactly did Nvidia buy for billions of dollars without executing a formal merger? The transaction secured two critical elements: mature silicon designs and the specialized engineering talent required to iterate on them. Groq built its reputation on SRAM-heavy dataflow accelerators designed to supercharge LLM inference, reaching hundreds and thousands of tokens per second—performance metrics that standard GPU architectures struggled to match independently.
Where traditional GPUs functioned like high-capacity city buses constrained by memory bandwidth limitations, Groq’s Language Processing Units (LPUs) operated like Formula 1 racers for sequential token generation. Combining the two methodologies yielded unprecedented throughput capabilities.
The integration timeline moved at breakneck speed. Just three months after the December 2025 deal was announced, Nvidia unveiled its LPX racks at the GPU Technology Conference (GTC) in March 2026, powered by 256 Groq-3 accelerators. Industry analysts noted that these chips were lightly modified iterations of Groq’s existing Groq-2 silicon, an aggressive tape-out cadence made possible entirely by the acquired engineering team. Jensen Huang promised that pairing Groq-3 with Vera Rubin GPU racks would optimize performance across mixed inference workloads.
Despite the technical prowess of the hardware, disaggregated compute architectures are not exclusive to Nvidia’s ecosystem. Competitors are aggressively pursuing parallel paths:
- Cerebras is building wafer-scale systems alongside AWS and AMD.
- SambaNova is collaborating with Intel on enterprise-grade inference.
- d-Matrix is combining its in-memory compute platforms with Nvidia GPUs to match the performance profile of Vera Rubin-LPX combinations.
Why Undoing the Deal May Already Be Irrelevant
Even if the DOJ successfully litigates and forces Nvidia to unwind the transaction, market observers suggest the legal remedy arrives far too late. Long before the Groq licensing pact was finalized, Nvidia had already engineered an open infrastructure standard designed to anchor its networking hardware at the center of the AI factory ecosystem.
In late 2024, Nvidia contributed its MGX rack designs to the Open Compute Project (OCP). This move allowed third-party chipmakers to physically house their accelerators inside racks originally architected exclusively for Nvidia hardware. By mid-2025, Nvidia opened its high-speed interconnect technology—the proprietary NVLink Fusion licensing scheme that allows dozens of discrete GPUs to operate as a single unified processor—to the broader industry.
These infrastructure standards mean that disaggregated rack design is already decentralized. If regulators sever the direct ties between Nvidia and LPU engineering teams, Groq technology—or competing accelerators built by other firms—can simply slot into Nvidia-compliant hardware racks as third-party components. The structural lock-in of the networking layer ensures that regulatory intervention regarding the talent transfer will alter corporate revenue streams rather than halt the hardware integration trajectory.
Nvidia continues to defend the legality and intent of the arrangement. In an official statement provided to media outlets, the company asserted: “The Groq story is a prime example of the American system working as designed to promote innovation, reward entrepreneurs, and benefit consumers. The law is designed to encourage America’s startup ecosystem and promote the fundamental rights of inventors and workers to pursue their dreams.”
As the DOJ’s formal investigation proceeds, the legal precedent established here will dictate whether future tech giants can acquire engineering capability through talent transfer loops without triggering federal premerger halts. For the marketplace, however, the architectural fusion of LPUs and GPUs is already deployed at scale.
