Nvidia’s 256-LPU Groq Rack Has a Buyer

Groq kept its cloud business and granted no exclusive license. Nvidia nevertheless put Groq’s processors in an Nvidia-branded rack, brought that rack into full production, and named Nebius its first customer. The chip had not failed. Its route to the customer had changed.

In December 2025, Nvidia and Groq confirmed a $20 billion non-exclusive inference-technology license; Groq CEO Jonathan Ross and other senior executives joined Nvidia, while Groq said it would continue operating independently. Nvidia could integrate the design into an established product line; Groq still had to finance the data centers behind its own cloud business.

Groq designed an engine; Nvidia brought the rack to market

Groq began as a startup founded by engineers from Google’s TPU project, pursuing a processor built for efficient model execution. Its on-chip memory architecture offered a distinct answer to latency-sensitive inference.

Nvidia’s Groq 3 LPX announcement preserved that distinction. The rack contains 256 Groq 3 LPUs, 128GB of on-chip SRAM and 40 PBps of SRAM bandwidth. Nvidia did not have to make the LPU look like a GPU to make it useful; it could give the LPU a place in a larger system sold under Nvidia’s name.

By August 2026, Nvidia said the rack had entered full production and Nebius had signed on as its first customer. Nvidia had turned Groq’s licensed design into its own product, though neither company had disclosed who controls manufacturing allocations, customer prices or other commercial terms.

Different inference jobs gave Nvidia room for another kind of chip

The inference infrastructure buildout presents hardware designers with more than one optimization problem. During prefill, a system processes the incoming sequence; subsequent token generation places different demands on memory movement and latency. Nvidia’s Rubin CPX design emphasizes compute performance for prefill rather than memory bandwidth. Groq’s SRAM-heavy LPUs keep data close to the processor for speed-sensitive work.

Nvidia can offer customers more than one accelerator inside its system portfolio. With heterogeneous compute, customers must choose which work runs where, and operators must deploy and maintain the resulting racks. Nvidia’s first LPX customer is evidence of a route to market for that rack, not proof that the two architectures have become interchangeable.

Groq’s forecast broke at the data-center door

Groq encountered the other side of specialization before signing the Nvidia agreement. In July 2025, documents showed that it had cut its 2025 revenue projection from more than $2 billion to more than $500 million, citing delays in data-center capacity. Those were projections, not booked revenue. Groq cited a deployment delay, not a failed chip benchmark.

A cloud provider cannot sell capacity that it has not made available. Groq must turn financing into sites, power, equipment and operating service before a processor design can become cloud revenue. After the license, Groq raised $650 million, led by Disruptive and Infinitum, to expand data-center capacity; it set a goal of 200 MW by the end of 2027. The target describes a plan, not capacity already online.

Other specialists have not all taken Groq’s path

Cerebras offers a substantial counterexample to any claim that specialist silicon must end up inside an incumbent’s rack. The company reported $180 million in second-quarter revenue, up 74% year over year, and raised its annual revenue and gross-margin forecasts. It also unveiled its own CS-4 rack, built around three WSE-3 Turbo chips. Cerebras is pursuing the system-supplier role itself, with reported revenue to support that effort.

Tenstorrent’s $1.5 billion Series E at a $6.2 billion valuation likewise shows that investors will still finance an independent chip company. Financing does not prove that Tenstorrent can deploy systems at scale, just as Groq’s 200 MW target does not prove it has reached that capacity. But neither company fits a rule in which every specialist has to license its design to a larger platform.

A non-exclusive license leaves two businesses to judge

GroqCloud remains in operation, and Groq has raised money to expand it. Nvidia has a production rack with a first customer. Both statements can be true without making the license a completed acquisition or Groq’s cloud a vestige. The disclosed terms do not settle manufacturing allocation, pricing authority, customer-contracting rights, cloud access, deployment commitments or revenue recognition. Those omissions matter more to commercial control than the word “non-exclusive” alone can reveal.

For now, the physical distinction is plain: Nvidia has put Groq’s name on a 256-LPU rack in production; Groq has put its name on a 200-megawatt capacity plan that it still has to build.