Nvidia paid roughly $20 billion for a non-exclusive license to Groq's inference chip technology and hired its two top executives — while Groq itself stays a nominally independent company. Here's what was actually licensed, why the structure matters more than the price tag, and what it signals about how dominant tech companies now avoid merger review.
In December 2025, Nvidia agreed to pay roughly $20 billion for a non-exclusive license to Groq's Language Processing Unit (LPU) inference architecture. Groq co-founder and CEO Jonathan Ross and president Sunny Madra moved to Nvidia along with a chunk of the engineering team. Groq itself did not get acquired. It kept its name, its cap table, and GroqCloud kept serving customers without interruption — under a new CEO, Simon Edwards, who wasn't there when the deal was signed.
That structure is the story. Nvidia didn't buy Groq. It licensed Groq's technology and hired Groq's leadership, and by doing it that way, it avoided triggering the merger-review process that would normally apply to a deal of this size. Two senators have since opened a formal inquiry into whether that was the point.
Groq's LPU is architecturally different from the GPUs Nvidia sells by the millions. A GPU is a general-purpose parallel processor with high-bandwidth memory (HBM) sitting off-chip, connected over a memory bus. That bus is the bottleneck for autoregressive LLM inference, which is memory-bandwidth-bound during token-by-token decoding, not compute-bound. Groq's LPU instead keeps weights in on-chip SRAM and uses a fully deterministic, statically scheduled execution pipeline — no runtime scheduler, no cache misses, no branch prediction. The compiler decides exactly which unit does what on which cycle, ahead of time.
The tradeoff is capacity: SRAM is far more expensive per gigabyte than HBM, so a single LPU holds much less model weight than a GPU holds in HBM, and serving a large model means chaining many chips together. What you get in exchange is extremely high, extremely consistent token throughput with none of the tail-latency variance that comes from a GPU's dynamic scheduling. Independent benchmarks through mid-2026 had GroqCloud serving Llama 3.1 8B at roughly 840 tokens/second and Llama 3.3 70B at around 394 tokens/second — multiples of what most GPU-based inference services deliver for the same models.
That speed is exactly what Nvidia wanted for its inference product line. At GTC 2026, Nvidia unveiled LPX, a liquid-cooled inference appliance that incorporates Groq's architecture directly into Nvidia's own hardware roadmap — the first concrete product to come out of the license.
"Reverse acquihire" isn't Nvidia's term, and it isn't a formal legal category — it's the label analysts and now senators have applied to a now-recognizable playbook: instead of buying a company outright, the acquirer licenses its core IP and hires away its key people, leaving a hollowed-out but nominally independent entity behind. Nvidia had already run a version of this play in September 2025, licensing chip-interconnect startup Enfabrica's technology rather than acquiring it.
The mechanism that makes this attractive is the Hart-Scott-Rodino Act, which requires companies to notify the FTC and DOJ before completing mergers or acquisitions above a certain size and observe a waiting period before closing. A licensing agreement plus voluntary hires isn't a merger or acquisition in the technical sense HSR was written to cover, so it doesn't trigger that filing — even when the economic effect looks a lot like one.
| Standard acquisition | Traditional tech license | Reverse acquihire (Nvidia–Groq) | |
|---|---|---|---|
| Target company status | Absorbed, ceases independent operation | Stays independent, keeps operating | Stays nominally independent, leadership departs |
| HSR antitrust filing required | Yes, above size thresholds | Usually no (no change of control) | No — this is the point |
| Key talent | Transfers with the company | Stays with licensor | Transfers directly to licensee |
| IP ownership | Fully transfers to acquirer | Stays with licensor, usage rights only | Stays with licensor, broad license granted |
| Precedent cited | Microsoft–Activision, Amazon–iRobot | Standard cross-licensing deals | Microsoft–Inflection (2024), Nvidia–Enfabrica (2025) |
The Microsoft–Inflection deal from March 2024 is the template everyone points to: Microsoft paid roughly $650 million to license Inflection's models and hired co-founder Mustafa Suleyman to run Microsoft AI, while Inflection continued to exist as a smaller, redirected company. Amazon did something similar with Adept later that year. Nvidia–Groq is the largest instance of the pattern by a wide margin — $20 billion against Groq's $6.9 billion valuation from a Series E just three months earlier, a roughly 2.9x markup that reflects the value of the talent and technology, not a market price for the company.
In March 2026, Senators Elizabeth Warren and Richard Blumenthal opened a formal inquiry into the deal, pressing Nvidia CEO Jensen Huang on whether the structure was designed specifically to sidestep antitrust review. Their letter argued the transaction could entrench Nvidia's roughly 90% share of the GPU market by neutralizing one of the few companies with a credibly different inference architecture, while avoiding the premerger notification process HSR was built to apply to exactly this kind of consolidation. They asked the DOJ and FTC to review the arrangement, and noted that reverse acquihires remain subject to antitrust law after the fact even when they dodge the upfront filing requirement — meaning the deal isn't necessarily safe just because it wasn't blocked before closing.
As of this writing, that review is active and unresolved. No enforcement action has followed, and there's no clear precedent for what a post-hoc antitrust challenge to a reverse acquihire would even look like in practice — which is itself part of why the structure has proliferated. Regulators built HSR around a specific definition of "acquisition," and a sufficiently creative deal can sit just outside it.
Groq's cloud business, GroqCloud, kept running. In February 2026 the company distributed roughly $7.6 billion to shareholders — the first major payout from the Nvidia deal, equivalent to about $64 per share and representing roughly three-quarters of total shareholder value. Then in June 2026, under new leadership, Groq raised a fresh $650 million to reposition itself as what the company is now calling an inference "neocloud" — a cloud provider built specifically around fast token serving rather than general-purpose compute rental.
The scale numbers Groq is now citing are real: the platform reports processing trillions of tokens per week across roughly 5 million developers, running out of 13 data centers on multiple continents, with a stated goal of reaching 200 megawatts of inference capacity by 2027. The company is also developing next-generation LPU silicon on a 4nm process with Samsung Foundry, a jump from the 14nm parts currently in production, aimed at better performance-per-watt and support for larger context windows.
What's genuinely uncertain is whether Groq can keep winning inference workloads on architecture and price alone now that its two most technical leaders and a slice of the team that built the original LPU are inside Nvidia. The speed advantage in the benchmarks is real and independently verifiable. Whether a reorganized company under new leadership can keep extending that lead — especially once Nvidia's own LPX line, built on licensed versions of the same ideas, starts shipping at Nvidia's distribution scale — is the open question the $650 million round is betting on.
The inference market is where the capital is moving right now, not just around Groq. The same week Groq closed its $650 million raise, Baseten closed a $1.5 billion round at a $13 billion valuation — another inference-focused neocloud, not a foundation model lab. Cerebras and SambaNova continue to compete on specialized inference silicon against both Nvidia's mainline GPUs and the hyperscalers' own inference stacks. Training compute got most of the attention through 2023 and 2024; by 2026, serving tokens cheaply and fast at massive scale is where a comparable amount of money is being bet.
For Nvidia, the deal buys two things at once: it neutralizes the most credible alternative inference architecture to its own GPUs, and it does so without the months-long merger review and possible blocking that a straight acquisition of a company Groq's size might have invited given Nvidia's existing market position. Whether regulators ultimately treat that as a distinction without a difference is the question the Warren-Blumenthal inquiry is trying to force into the open — and the answer will likely shape whether "license the tech, hire the team, leave the shell" becomes the standard move for how dominant companies absorb competitive threats going forward.
The takeaway: when a licensing deal comes bundled with the target's founders and comes in at several times the company's last priced valuation, the license is doing the work an acquisition used to do — read the antitrust angle, not just the headline number.