The DOJ Is Sniffing Around Nvidia's Groq Deal. It's Probably Already Too Late.
Nvidia dropped roughly $20 billion late last year to license Groq's AI accelerator technology and hoover up most of its engineering team. Technically, Groq's inference-as-a-service business survived intact. In practice, stripping out the engineering talent while leaving the legal shell standing is the oldest trick in the acquihire playbook.
Regulators noticed anyway. The New York Times reported this week that the Department of Justice has opened an antitrust probe into the arrangement.
Nvidia's response was, predictably, to wrap the whole thing in a stars-and-stripes bow. The company told press that the deal represented "the American system working as designed" and invoked the sacred rights of inventors and workers to pursue their dreams. You'd never guess they just absorbed a competitor's entire technical workforce.
Whether that actually damaged competition is a separate, harder question.
What Nvidia Actually Got
For anyone unfamiliar: Groq (zero relation to Elon Musk's Grok, despite the phonetic confusion) built a name for itself with SRAM-heavy dataflow accelerators, called LPUs, that could run LLM inference at hundreds, now thousands, of tokens per second. GPUs, for all their brute-force training power, have historically struggled with inference throughput at that kind of speed.
The analogy that holds up: Groq's LPUs are the Formula 1 car, Nvidia's GPUs are the city bus. Neither is ideal on its own. Together, you get something considerably more useful.
At GTC in March, Nvidia unveiled LPX racks built around 256 so-called Groq-3 accelerators. Given that the deal was only announced a few months prior, these are almost certainly lightly modified Groq-2 designs. Taping out brand new silicon in three months would be physically impossible. Jensen Huang promised the combination of Groq-3 LPUs and Vera Rubin GPU racks would cover the full range of inference workloads. Bold claim. Not an obviously wrong one.
The inconvenient detail for Nvidia's critics is that disaggregated compute architectures are not remotely unique to Groq. Cerebras is building comparable systems with AWS and AMD. SambaNova is working the same angle with Intel. d-Matrix is pairing its in-memory compute platform with Nvidia GPUs to achieve similar ends. The technology Nvidia acquired is real and valuable, but the competitive landscape around it is crowded.
Can the DOJ Actually Do Anything Here?
Several US senators have already argued that deals deliberately structured to duck merger review thresholds should face scrutiny regardless. That principle is reasonable. The DOJ proving actual competitive harm in this specific case is a trickier proposition.
And even if they managed it, even if they forced Nvidia to unwind the deal as they once torpedoed an earlier Nvidia acquisition attempt, the practical consequences might be modest.
Before any of this happened, Nvidia had already been quietly building the scaffolding for an ecosystem that doesn't require owning every piece of silicon inside it. In late 2024 the company contributed its MGX rack designs to the Open Compute Project, allowing any chipmaker to slot hardware into racks originally designed around Nvidia GPUs. Then in mid-2025 it opened its NVLink Fusion high-speed interconnect technology to outside licensees, meaning third-party chips could plug directly into Nvidia's AI factory infrastructure.
If the DOJ unwound the Groq deal tomorrow, Nvidia would lose direct control over LPU development and the associated chip revenues. It would not lose the racks. Groq would simply join the queue of hardware partners designing around Nvidia's platform, same as everyone else.
If anything, the acquihire has accelerated the build-out of that partner ecosystem. Block the deal and there are already alternatives lining up to fill the gap. Nvidia planned it that way, whether deliberately or not.