- Lead. The Justice Department has sent Nvidia a formal demand for information as part of an investigation into whether its December 2025 arrangement with AI chip startup Groq was structured as a licensing agreement rather than an acquisition specifically to avoid automatic antitrust review.
- Fact. The deal gave Nvidia access to Groq’s custom AI inference chips and brought in Groq’s chief executive Jonathan Ross and chief operating officer Sunny Madra — a combination that regulators say resembles an acquisition in all but legal form.
- Stake. Officials do not currently expect the deal to be unwound, but any finding of wrongdoing could result in fines and would set a precedent shaping how AI companies structure technology-and-talent transactions that have proliferated as the sector consolidates around a handful of dominant platforms.
The arrangement, announced in December 2025 and described by Groq as a “nonexclusive licensing agreement,” drew immediate scrutiny from antitrust researchers and a group of US senators who characterised it as a device to “evade scrutiny by antitrust regulators.” Groq, which had been valued at $7 billion before the agreement, remains technically independent and continues operating its cloud inference services. In August 2026, it raised $350 million in new financing, with Nvidia participating. The New York Times, citing sources familiar with the investigation, reported that the DOJ had sent Nvidia a formal information demand.
The Structural Question
US antitrust law requires companies to file pre-merger notifications when acquisitions cross certain transaction-value and size-of-person thresholds. Licensing agreements — even comprehensive ones that transfer key technology and bring in the founding team — do not automatically trigger those filing requirements. Regulators have grown increasingly alert to this gap as AI companies have structured deals to move personnel and intellectual property without the delays and conditions that formal merger review can impose.
The Nvidia-Groq arrangement fits a pattern that has emerged across the AI sector: a dominant platform company and a well-capitalised startup reach a deal in which the startup licenses its core technology, some of its senior leadership moves to the platform company, and the startup continues operating as a nominally independent entity. Microsoft and Inflection, Amazon and Adept, and now Nvidia and Groq have all attracted regulatory attention for variations on this structure. The DOJ investigation signals that federal enforcers have decided to test whether existing law reaches these arrangements, even without a new legislative mandate.
What Groq Built
Groq’s significance lies in its inference architecture. While the GPU-centric approach pioneered by Nvidia dominates AI model training, inference — running a trained model to generate outputs — has different computational characteristics: it prioritises low latency and energy efficiency over raw throughput. Groq’s Language Processing Unit was designed specifically for inference workloads and demonstrated benchmark speeds substantially faster than comparable Nvidia hardware for certain model architectures.
For Nvidia, access to Groq’s technology represents both a hedge and an offensive move in the inference market. Competitors including Qualcomm — which recently announced a $60 billion partnership with Amazon targeting AWS inference workloads — are investing heavily to challenge Nvidia’s position. Groq’s LPU architecture, integrated into Nvidia’s product roadmap, would strengthen its inference offering precisely as that segment of the compute market expands to dwarf training in total volume.
Investigation Timeline
The DOJ investigation remains at an early stage. The agency has sent a formal demand for documents and information but has not filed any complaint or notified the companies of a specific legal theory. Officials familiar with the matter told the New York Times they do not currently expect the deal to be unwound; the more likely outcome, if wrongdoing is found, is a financial penalty and potentially conditions on how Nvidia uses the licensed technology. The probe is expected to run well into 2027, adding sustained regulatory uncertainty to a chip sector already navigating export controls and the fallout from Nvidia’s parallel $12.9 billion acquisition of Hugging Face.