Groq raises $650M to become inference cloud after Nvidia deal

Groq did something unusual on May 28: it raised $650 million from existing investors to fund a business that no longer controls its core intellectual property. Five months earlier, the startup had licensed its language processing unit (LPU) architecture to Nvidia for $20 billion, then distributed $7.6 billion to shareholders, roughly $64 per share, as the first major payout. By February 2026, the founding team had moved into Nvidia's organization. Now Groq is rebuilding itself as an inference neocloud provider, powered by Nvidia chips rather than its own silicon, with backing from Disruptive and Infinitum. The pivot suggests a bet that inference-heavy AI workloads are large enough to support a vertically focused services business, even if the hardware differentiation is gone.

The scale of that Nvidia deal matters here. A non-exclusive license to Groq's LPU architecture does not sound like a acquisition, Groq technically retained its independence, but the mechanics were unambiguous: Nvidia got the IP, founder Jonathan Ross moved to Nvidia's labs, and the core engineering team followed. At GTC 2026, Nvidia unveiled the Groq 3 chip, built on Samsung 4nm, with 150 TB/s of on-chip SRAM bandwidth (7× faster than Nvidia's Vera Rubin GPU for high-bandwidth memory), 315 petaflops FP8 per rack, and 35× throughput per megawatt compared to Blackwell for trillion-parameter models. The chip ships in Q3 2026. What Groq designed, a silicon architecture optimized for inference throughput rather than training flexibility, was valuable enough that Nvidia paid $20 billion for non-exclusive rights and then built its own version. That validator signal matters more than the lip service to Groq's independence.

But independence is also why Groq can now raise $650 million to operate a competing service on top of Nvidia hardware. Groq 2.0, led by interim CEO Adam Winter and Matt Eng as CFO, inherits a customer base of more than 3.5 million developers and Fortune 500 companies that were using its inference cloud before the Nvidia deal. The capital funds infrastructure expansion into a market segment that is already larger and moving faster than model training: inference for deployed applications, real-time personalization, and edge-adjacent workloads. That segment does not care whether the silicon is branded Groq or Nvidia; it cares about latency, throughput per dollar, and whether the service stays live. Groq no longer owns the silicon, so it competes on service quality and cost, not architectural advantage.

There is a regulatory subplot that will determine whether this reconstitution survives intact. In March 2026, Senators Elizabeth Warren and Richard Blumenthal opened a formal antitrust inquiry into the Nvidia deal, arguing it was structured as a reverse acqui-hire designed to avoid Hart-Scott-Rodino filing thresholds. They gave Nvidia an April 3 deadline to respond and urged the DOJ and FTC to investigate. If the inquiry concludes the deal was a disguised acquisition of key personnel to eliminate a competitor, the structure could unwind, forcing Nvidia to divest Groq's returned IP or reversing the licensing arrangement. That risk is material because Groq's cloud service has no defensible moat without exclusive access to Groq LPU silicon or superior customer service that sticks during transition. Right now, Groq holds about 3.5 million developers and some Fortune 500 compute budget. If Nvidia is forced to give back the IP, Groq could rebuild as a chip company. If the deal stands, Groq is a thin cloud services layer on someone else's platform.

The real read: Groq's pivot is not a vindication of the inference cloud thesis, it is an admission that inference-optimized silicon, on its own, does not generate enough margin or scale to survive as a standalone business. The $20 billion Nvidia paid was partly validation and partly an acqui-hire at hardware scale. Groq's next chapter depends on whether 3.5 million developers and Fortune 500 customers stay on the service when the silicon is now sold by Nvidia directly. Watch three things: first, whether Groq's developer count and Fortune 500 customer base hold steady through Q3 2026 when Groq 3 ships and Nvidia begins direct sales of inference capacity; second, whether the antitrust inquiry progresses or stalls; and third, whether other inference-focused startups, Cerebras, Lambda Labs, others, follow Groq's playbook and license their IP rather than scale as chip makers. If Groq loses half its customer base by end of year, the model breaks. If Groq holds steady and the antitrust inquiry fades, Groq becomes a template. If the antitrust case accelerates, everything reverses.