DOJ가 엔비디아의 그록 인수에 대한 공식 독점금지 조사를 개시했다.

The U.S. Department of Justice has formally opened an antitrust investigation into Nvidia’s $17 billion licensing deal with AI chip startup Groq, issuing a formal demand for information. This marks the first time a U.S. regulator has moved from stated concern to an active probe of the reverse-acquihire structure that has channeled over $100 billion in AI-era capital while bypassing mandatory merger review. First reported by the New York Times and independently corroborated by Reuters, the inquiry presents the broader AI industry with a deferred legal question: whether a technology license paired with a mass talent transfer constitutes an acquisition requiring pre-closing federal reporting.

The investigation centers on whether Nvidia structured its December 2025 arrangement—licensing Groq’s inference chip technology and bringing founder Jonathan Ross, president Sunny Madra, and other senior leaders to Nvidia—as a licensing deal specifically to circumvent premerger notification requirements under the Hart-Scott-Rodino Antitrust Improvements Act. According to sources familiar with the matter, the agency opened its inquiry shortly after the announcement and has since escalated to a formal information request. Sources indicate the DOJ is unlikely to seek to unwind the already-integrated transaction but may impose financial penalties if it determines a filing was legally required.

No violation has been established. Responding directly, Nvidia stated: “The Groq story is a prime example of the American system working as designed to promote innovation, reward entrepreneurs, and benefit consumers.”

The HSR Act’s premerger notification threshold—currently set at $119.5 million—requires companies to notify the DOJ and FTC and observe a waiting period before closing acquisitions above that value. Nvidia structured its Groq deal as a non-exclusive technology license plus a talent transfer, leaving Groq nominally independent under a new CEO. Because no corporate entity changed hands and no equity stake was formally acquired, Nvidia maintains the transaction did not trigger HSR filing requirements.

The DOJ appears to be advancing a different theory: that the economic substance of asset transfers matters more than the legal label. In March 2026, Acting Assistant Attorney General Omeed Assefi told Reuters that designs to circumvent merger review present a regulatory “red flag,” citing the Nvidia-Groq structure as a recent example. Academic legal analysis supports this view. A Yale Law Journal comment by David T. Wong argued that reverse acquihires constitute asset acquisitions under the Clayton Act, as the valuable assets transferring—engineering talent, specialized know-how, and proprietary technology access—represent human capital and associated intellectual property rather than a mere corporate shell. The University of Chicago Law Review reached a similar conclusion, noting that IP licensing agreements in reverse acquihires can be characterized as intangible asset acquisitions subject to HSR review.

The probe places specific scrutiny on a concrete product milestone: less than three months after the December deal closed, Nvidia unveiled a new inference processor incorporating Groq’s chip technology at its GPU Technology Conference in March 2026. Senators Elizabeth Warren (Massachusetts) and Richard Blumenthal (Connecticut) highlighted this timeline in their March 2026 letter to Jensen Huang, writing: “By licensing its technology and hiring its most important employees, NVIDIA has effectively acquired Groq in all but name.”

Understanding why Nvidia paid $17 billion for a chip company it did not formally acquire requires examining what Groq’s Language Processing Unit (LPU) does differently from a standard GPU—and why that distinction is critical in the AI market Nvidia dominates. Conventional AI inference relies on graphics processing units that store model weights in high-bandwidth memory (HBM): dense, off-chip memory stacks delivering approximately 3 to 3.35 terabytes per second of bandwidth. Groq’s LPU replaces HBM entirely with on-chip static random-access memory (SRAM), which sits directly on the processor die without a memory bus, DRAM controllers, or round-trip latency. This architecture eliminates the fundamental bandwidth bottleneck limiting GPU-based inference. The Groq 3 LPX, which entered full production in August 2026, features 500 megabytes of on-chip SRAM and delivers approximately 150 terabytes per second of internal bandwidth—roughly 45 times the memory bandwidth of an H100 GPU.

The performance advantage is decisive for a specific large language model serving bottleneck: the decode phase, where a model generates each output token by attending to all previous tokens in the KV cache. This phase is memory-bandwidth-bound, as every token generation requires reading the entire weight matrix from memory. On GPUs, weights travel across an HBM bus; on the LPU, they never leave the chip. Nvidia’s own data claims the Groq 3 LPX delivers 35 times more throughput per megawatt than HBM-based GPU inference at equivalent model sizes. The architect behind this design, Jonathan Ross, founded Groq in 2016 after previously initiating Google’s Tensor Processing Unit project as a 20-percent-time initiative in 2013. Ross designed and deployed that chip across more than half of Google’s compute infrastructure within 15 months. At Groq, Ross bet that the inference market would outpace training and that the decode-phase bottleneck would demand purpose-built silicon—a wager that appeared prescient by late 2025 as agentic AI applications began demanding sub-millisecond response times at scale. For Nvidia, securing Ross’s expertise and Groq’s LPU architecture through a single arrangement represented a structural acceleration of its inference roadmap. The timing was commercially urgent: the Groq 3 LPX was announced at GTC in March 2026, less than three months after the deal closed.

The practical stakes of the investigation hinge on the DOJ’s legal conclusion and the remedies available for a deal that has already closed and integrated. Under the HSR Act, civil penalties for failure to file reach approximately $50,000 per day of violation, assessed from the closing date until a required filing would have been completed. For a deal that closed in late December 2025, that penalty window has been accumulating for months. However, sources familiar with the investigation consistently indicate the DOJ is likely to pursue a financial penalty rather than unwind the transaction. More significant than any fine is the precedent this case would establish. A DOJ determination that the Nvidia-Groq structure required HSR filing would mark the first time a technology license paired with key-employee transfers is classified as an “acquisition of assets” under the Act, regardless of the target’s nominal independence. Such a ruling would retroactively cast doubt on structurally similar deals across the AI sector: Microsoft’s arrangement with Inflection AI, Google’s partnerships with Character.AI and Windsurf, and Meta’s $14 billion Scale AI deal, in which Meta hired Scale AI’s CEO and acquired a large minority stake. None of those transactions were filed under HSR, and all would face potential retroactive exposure under the same legal theory now applied to Nvidia. This industry-wide implication explains why the probe extends far beyond AI chipmaking. The reverse-acquihire structure is not an Nvidia-specific tactic; it became the primary mechanism through which major technology firms consolidated AI talent and technology starting in 2023. Consequently, the DOJ’s first formal probe functions as a legal audit of the entire playbook.

Nvidia’s defense is substantive rather than merely procedural. The company argues that the deal preserved Groq’s technology rather than extinguishing a competitor—a position supported by factual circumstances. By late 2025, Groq faced significant financial pressure, having conducted layoffs and lost its former chief architect. Multiple accounts indicated the company struggled to secure the compute resources necessary to scale LPU production without external investment. In that framing