DOJ Probes Nvidia's $20B Groq 'Acquihire' Over Antitrust Concerns

Nvidia spent the sum late last year to license Groq's AI accelerator technology and hire key members of its engineering team. The New York Times reported this week that the DOJ is now examining the arrangement. Nvidia has defended the deal, calling it 'a prime example of the American system working as designed to promote innovation, reward entrepreneurs, and benefit consumers.' The company further argued that the law is designed to encourage the startup ecosystem and protect the rights of inventors and workers.

The deal gave Nvidia access to Groq's SRAM-heavy dataflow accelerators, which deliver hundreds to thousands of tokens per second in LLM inference—far faster than traditional GPU-based systems. At its GTC conference in March, Nvidia unveiled LPX racks powered by 256 Groq-3 accelerators, which are lightly modified versions of Groq's existing chip designs. CEO Jensen Huang claimed the combination of Groq-3 with Vera Rubin GPU racks would deliver optimal performance across inference workloads.

However, disaggregated compute architectures are not unique to this deal. Competitors such as Cerebras (partnering with AWS and AMD), SambaNova (with Intel), and d-Matrix (with Nvidia GPUs themselves) are building similar systems. Even if the DOJ were to successfully unwind the deal—regulatory torpedoing of Nvidia deals has precedent—analysts suggest it may be too late. Nvidia has already laid the groundwork for an open ecosystem, contributing its MGX rack designs to the Open Compute Project in late 2024 and opening its high-speed interconnect tech to the industry through NVLink Fusion in mid-2025.

Why This Matters

The probe tests whether 'acquihire' deals—structured as licensing agreements rather than outright mergers—can escape antitrust review.
Nvidia's dominance in AI hardware means any deal that consolidates its position could raise barriers for competitors and raise costs for consumers.
The outcome could set a precedent for how regulators treat similar talent-and-IP acquisitions by dominant tech firms.

Background

Groq's technology uses SRAM-heavy dataflow accelerators (LPUs) to achieve extremely low latency for AI inference, outperforming traditional GPUs in that specific metric. Nvidia, already the dominant supplier of AI training and inference hardware, sought to integrate this capability into its own product line. The licensing-plus-talent deal allowed Nvidia to absorb Groq's engineering expertise without triggering standard merger review thresholds—a strategy that has now drawn federal attention.

Key Perspectives

Nvidia: The deal is a lawful, pro-innovation outcome that benefits the ecosystem and respects the rights of inventors and workers. DOJ and critics: The arrangement effectively transfers control of a key startup to the dominant player, potentially reducing competition in AI inference hardware. Industry observers: Even if unwound, the competitive landscape has shifted; Nvidia's subsequent open ecosystem moves (MGX, NVLink Fusion) may have already locked in advantages.

What to Watch

Whether the DOJ escalates the investigation to litigation or seeks to unwind the deal outright.
Competitor responses, particularly whether Cerebras, SambaNova, or d-Matrix accelerate their own disaggregated compute offerings.
Any additional regulatory scrutiny of Nvidia's broader licensing and acquisition strategy.

ZOTPAPER