Nvidia-Groq $20B Deal Under DOJ Investigation, September 2026
The Justice Department has opened a formal antitrust probe into Nvidia's reported $20 billion Groq acqui-hire, casting regulatory uncertainty over an accelerator program that has already entered mass production at Samsung and is scheduled to deploy at Nebius before year-end.
The U.S. Department of Justice has opened a formal antitrust probe into Nvidia's reported $20 billion deal with AI inference specialist Groq, multiple reports from September 2026 confirm. The investigation joins a parallel Federal Trade Commission review of large technology companies' so-called acqui-hire practices — arrangements in which a major incumbent absorbs a target's intellectual property and key talent while allowing the entity to persist as a nominally independent operation. For Nvidia, the scrutiny arrives at a structurally awkward moment: the Groq 3 LPX accelerator that emerged from the partnership has already entered mass production at Samsung's foundry, first-party deployment at Nebius has been confirmed, and the commercial clock is running.
Understanding why Nvidia struck the deal requires a brief account of Groq's trajectory. Founded on a proprietary Language Processing Unit architecture — a deterministic, low-latency chip built from the ground up for inference rather than training — Groq spent years as a technically credible but commercially subscale player. Its core proposition was straightforward: by sidestepping the memory-bandwidth bottlenecks that constrain general-purpose GPU inference, the LPU could deliver faster token generation at lower energy per output. That case gained institutional validation in March 2026, when Nvidia announced its $20 billion commitment — coverage at the time described the move as following the Google playbook of acquiring specialized AI design and IP capability — and when Groq's LPU was integrated into the Vera Rubin platform. By August 2026, Nvidia was presenting the Groq 3 LPX architecture at Hot Chips and releasing what was described as a first third-party benchmark, with the LP30-based rack already reported as in production.
The technical figures now on record are material to understanding the commercial stakes. The Groq 3 LPX reportedly delivers 30 times the throughput per megawatt compared with the preceding generation, and an independent benchmark on the Vera Rubin platform measured 3,431 tokens per second — performance directly addressed at the agentic AI workloads that are driving inference demand growth. Samsung's mass production of the chip is separately cited as a meaningful step in the South Korean foundry's own recovery story. Groq as an operating entity has simultaneously been raising substantial capital: coverage from June 2026 reports a $650 million raise to expand inference cloud capacity toward a 200-megawatt target by end of 2027, while August 2026 reports describe a $350 million Series A at a $3.5 billion valuation with Nvidia joining the round. In August, Groq also reportedly hired Meta's network chief Ola Kasali to deepen its infrastructure engineering leadership. The composite picture is of an architecture moving quickly from prototype to at-scale infrastructure — against a backdrop of rapid-succession capital formation.
The commercial logic rests on an inference economics paradox documented in September 2026 coverage: even as per-token costs have fallen roughly a thousandfold, total AI infrastructure bills have risen because application volumes have grown faster still. Analysis naming Groq alongside Cerebras and SambaNova positions these specialized inference architectures as capable of capturing expanding workloads that commodity GPU clouds struggle to serve efficiently at the frontier of latency and energy density. That framing explains both the capital raised and the talent assembled.
The DOJ probe introduces genuine uncertainty, but its ultimate impact depends on how regulators define the deal and whether they elect to block, condition, or clear it. Groq's continued independent fundraising and the fact that production has already commenced complicate any straightforward unwinding. Observers should track three concrete signals: whether the DOJ issues a formal second request or challenge, which would push deployment timelines materially; whether the first Nebius Groq 3 LPX racks go live on schedule as a proof-of-production milestone; and whether SambaNova's competing raise — reported in September 2026 at a $11 billion valuation on $1 billion in new capital — signals that inference-specialized silicon has accumulated enough independent momentum to sustain itself even in a scenario where the Nvidia relationship is restructured. The case will set precedent not only for Groq but for the wider question of how technology incumbents integrate specialized IP in an era of heightened antitrust vigilance.