Nvidia Moves $20B Groq Asset Deal to Production With New Inference Racks

Nvidia Corp. is placing its Groq 3 LPX rack into full production, taking its first commercial step with technology obtained through a $20 billion asset and licensing deal finalized late last year. As reported by CNBC and published by Yahoo Finance, Nvidia Senior Director Dion Harris confirmed to reporters that the systems will ship later this year to neocloud infrastructure firm Nebius, where they will operate alongside Nvidia's Vera CPUs and Rubin GPUs.

The liquid-cooled LPX cabinets house 256 Groq processors each. Independent benchmark data from Artificial Analysis indicates that the rack configuration achieves an inference speed of 3,400 tokens per second. While Nvidia relies on Taiwan Semiconductor Manufacturing Co. to fabricate its flagship graphics processing units, the specialized Groq chips inside the LPX rack are manufactured by Samsung.

Rather than creating an internal alternative to Nvidia's existing graphics processors, the LPX system is designed to broaden the company's coverage across the entire AI compute stack. The Groq architecture targets low-latency inference workloads, allowing Nvidia's core GPUs to concentrate on model training and large-context processing. Furthermore, cloud customers can pair the LPX hardware with Vera Rubin systems without changing their existing CUDA software workflows, maintaining Nvidia's established platform lock-in.

The transition to volume production comes eight months after Nvidia agreed to pay $20 billion to license Groq's technology and hire its workforce, representing the largest transaction in Nvidia's history. Because the deal was structured as an intellectual property license and talent acquisition rather than a full corporate takeover, investors have limited visibility into the ultimate financial return on the expenditure. However, Bernstein analyst Stacy Rasgon told CNBC that Nvidia's strong capital position allows it to absorb a transaction of this scale without balance-sheet strain, granting the company flexibility to invest in alternative chip architectures.

The eight-month timeline between the December agreement and commercial availability demonstrates rapid integration of the acquired intellectual property. On performance metrics, the LPX system provides Nvidia with a concrete speed benchmark against rival approaches; its 3,400 tokens per second output compares to the 750 tokens per second targeted by OpenAI for its upcoming Cerebras-powered "Ultrafast" mode.

The Groq deployment is one component of a broader investment strategy Nvidia has executed across the artificial intelligence sector. The company has pledged up to $100 billion to OpenAI and $5 billion to Intel, alongside smaller venture investments in startups including CoreWeave, Cohere, and Crusoe, as well as a $900 million licensing deal for Enfabrica's engineering team. However, because the LPX chip design is specialized strictly for inference rather than general-purpose compute or model training, its total addressable market remains smaller than Nvidia's core GPU business.

Widespread commercial adoption of the LPX rack remains to be proven, as Nebius is currently the only cloud provider publicly confirmed to deploy the hardware. Even so, institutional backing for Nvidia continues to expand. Data from Insider Monkey shows that 285 hedge funds held Nvidia stock during the second quarter of 2026, up from 275 in the prior quarter, compared to 164 hedge funds holding shares of rival Advanced Micro Devices Inc.