Senators Probe NVIDIA's $20B Groq Deal for Antitrust Evasion

Warren and Blumenthal accuse NVIDIA of structuring its Groq licensing deal to dodge merger review while consolidating AI chip dominance

Close-up of a circuit board with integrated chips

Senators Elizabeth Warren and Richard Blumenthal sent a pointed letter to NVIDIA CEO Jensen Huang on March 20, demanding answers about the company’s $20 billion deal with AI inference startup Groq. Their central accusation: NVIDIA structured the arrangement as a “licensing deal” specifically to evade mandatory antitrust review.

The Deal That Isn’t an Acquisition

NVIDIA insists it “did not acquire Groq, which continues to be a separate and independent business.” Technically true. But the deal’s structure tells a different story.

Under the late-2025 agreement, NVIDIA paid $20 billion for a non-exclusive license to Groq’s Language Processing Unit (LPU) technology. In practice, most of Groq’s software engineers and hardware designers joined NVIDIA. CEO Jonathan Ross also moved to the chip giant. Groq’s cloud business remains nominally independent, but the brain trust walked out the door.

Federal antitrust law requires companies to file most acquisitions above certain thresholds for regulatory review. Groq didn’t file. The senators argue this was by design.

“This deal appears to have been structured to evade scrutiny by antitrust regulators,” Warren and Blumenthal wrote. They’ve given NVIDIA until April 3 to explain why the arrangement shouldn’t be classified as an acquisition subject to Hart-Scott-Rodino review.

Why Groq Matters

Groq’s LPU architecture offers a fundamentally different approach to AI inference than NVIDIA’s GPUs. Where NVIDIA’s chips excel at the parallel processing needed to train models, Groq’s technology is purpose-built for running trained models at extreme speed and efficiency.

NVIDIA has acknowledged the strategic value: the company wants Groq’s technology to improve its inference offerings, which represent an increasingly large share of AI compute spending as enterprises move from model development to deployment.

The $20 billion price tag reflects this urgency. NVIDIA paid roughly 2.9 times Groq’s most recent valuation, a premium that suggests the company sees inference performance as a competitive vulnerability.

Pressure From Multiple Directions

The senators’ inquiry comes as NVIDIA faces intensifying antitrust scrutiny globally.

The EU launched a formal investigation into NVIDIA’s business practices following raids on the company’s offices. European regulators are examining whether NVIDIA has abused its dominant position through bundling practices and restrictions on customers.

The timing creates compounding risk. If regulators in multiple jurisdictions conclude that NVIDIA is using unconventional deal structures to consolidate control over AI infrastructure, the consequences could extend beyond this single transaction.

Who Wins, Who Loses

NVIDIA gains access to technology that could close gaps in its inference portfolio, but faces reputational and legal risk if regulators determine the deal structure was intentionally deceptive.

Groq as an independent company is hollowed out. The cloud business continues, but without the engineering team that built the technology.

Venture investors in Groq received a massive return, but the precedent is concerning. If large incumbents can effectively acquire startups through licensing deals that avoid regulatory review, the exit landscape changes for everyone.

Antitrust enforcement faces a test case. If the Warren-Blumenthal probe gains traction, it could establish new scrutiny for talent-acquisition deals structured as licensing agreements. If nothing happens, expect more “partnerships” that look suspiciously like acquisitions.

NVIDIA has until early April to respond. Given the company’s current regulatory exposure, the response will likely be carefully crafted. But the senators have made their view clear: paying $20 billion and hiring most of the team is an acquisition, regardless of what you call it.