# Nvidia's Groq 3 LPX Inference Rack Enters Full Production This Week > Source: > Published: 2026-08-26 00:38:07+00:00 *Nvidia turned a $20 billion raid on a rival chipmaker's brain trust into an actual product this week, and the company it gutted is still standing, just a lot smaller.* On August 24, Nvidia said its Groq 3 LPX rack has entered full production. It will ship to customers before the end of 2026. That's the first real hardware to come out of the unusual deal Nvidia struck with Groq back on December 24, 2025, a roughly $20 billion arrangement in which Nvidia licensed Groq's inference chip technology and hired founder Jonathan Ross, president Sunny Madra, and most of Groq's engineering staff, according to CNBC. Nvidia did not buy the company. That structure matters. Nvidia gets the technology and the people who built it, without absorbing Groq's balance sheet, its cloud business, or its remaining obligations. Groq keeps existing, on paper, run now by its former finance chief, Simon Edwards, as CEO. GroqCloud stayed out of the deal entirely and kept operating without interruption. Each Groq 3 LPX rack packs 256 of Groq's LP30 chips, liquid-cooled, wired directly into Nvidia's Vera Rubin NVL72 systems alongside Vera Rubin GPUs. In Artificial Analysis benchmark testing on the Gemma 4 31B model with a 100,000-token context window, the rack hit 3,400 output tokens per second. That's four times faster than the nearest platform on the same test, Nvidia said. At the rack level, Nvidia claims up to 35 times more throughput per megawatt than a standard Blackwell NVL72 setup when running trillion-parameter models, working out to roughly 300 tokens per second per megawatt at about $45 per million tokens, according to reporting from Tom's Hardware. Nebius will be the first cloud to actually run it. The company plans to fold Groq 3 LPX into its Token Factory inference platform, giving developers access to the faster chip through APIs they're already using. That's the part that turns a chip announcement into a business story: Nvidia is routing that silicon straight into the inference market Groq itself was trying to own. [Groq raises $650 million for its neocloud second act after selling its soul to Nvidia for $20 billion](https://startupfortune.com/groq-raises-650-million-for-its-neocloud-second-act-after-selling-its-soul-to-nvidia-for-20-billion/) Groq is raising $650 million to fund its pivot from AI chip maker to inference neocloud provider, following a $20 billion LPU licensing deal with Nvidia that paid out shareholders and handed the founding team to its former competitor. Existing investors Disruptive and Infinitum are backstopping the full round. The raise signals durable investor... - [Groq raises 650 million for AI inference neocloud](https://startupfortune.com/groq-raises-650-million-for-its-neocloud-second-act-after-selling-its-soul-to-nvidia-for-20-billion/) - [AI chip startup licensing technology to Nvidia deal](https://startupfortune.com/groq-raises-650-million-for-its-neocloud-second-act-after-selling-its-soul-to-nvidia-for-20-billion/) ## A Murkier Kind of Scrutiny Straight acquisitions draw antitrust scrutiny. Licensing deals paired with mass hiring draw a different, murkier kind of scrutiny, and lawmakers have already taken notice, according to Barchart's coverage of the arrangement. Nvidia gets Ross and the SRAM-based architecture that made Groq's language processing units fast at inference, without a merger review clock running. It's a structure other chipmakers will study closely: it shows you can hollow out a competitor's technical core while leaving a shell of the company operating in public. That's the playbook now. And Groq's shell has kept moving. Rather than fold, the company pivoted to running its own neocloud, combining leftover LPU capacity with Nvidia systems and selling inference as a service. It raised $650 million in June 2026, then another $350 million on August 17, this time at a $3.5 billion valuation, according to Bloomberg. That's roughly half the $6.9 billion valuation Groq commanded less than a year earlier, before Nvidia licensed away its core technology and hired away its leadership. Disruptive led the new round. Nvidia itself joined in as a participant, an odd bit of symmetry given Nvidia is also, in effect, Groq's biggest competitor now. A halved valuation after your best engineers and your founder leave for the company that just licensed your chip design isn't a sign of health. It's a sign the market thinks the interesting part of Groq already walked out the door. ## What It Means for Everyone Else Still, the technology works. That's the part that should worry Nvidia's actual GPU-only rivals more than it worries Groq. Frankly, a 35x throughput-per-megawatt claim, even discounted for marketing, is the kind of number that reshapes how cloud providers plan their next data center build. Power is the bottleneck in AI infrastructure right now, not silicon supply. A rack that does more work per watt changes the math for every AI cloud deciding what to buy next. Nvidia just gave Nebius a two-month head start on everyone else weighing that decision. What happens to AMD, Cerebras, and the other inference-chip challengers now depends on whether this was a one-off licensing quirk or the template Nvidia plans to reuse. Given how cheaply it got Groq's brain trust relative to a full acquisition, don't bet against a repeat. **Also read:** [Waymo Picks Munich as Its First European Robotaxi City for Late 2027](https://startupfortune.com/waymo-picks-munich-as-its-first-european-robotaxi-city-for-late-2027/) • [Anima Anandkumar Launches Accelerated Understanding to Rival ChatGPT](https://startupfortune.com/anima-anandkumar-launches-accelerated-understanding-to-rival-chatgpt/) • [xAI's Grok Chatbot Glitched Into Gibberish About Cheese and Planets](https://startupfortune.com/xais-grok-chatbot-glitched-into-gibberish-about-cheese-and-planets/)