Nvidia issued a public correction this week rejecting what it called inaccurate claims in a report by The Information about the Groq 3 Language Processing Unit and the strategy surrounding it. The LPU business is based on a roughly $20 billion licensing agreement with chip startup Groq announced in late 2025. Introduced at Nvidia’s GTC 2026 event on March 16, 2026, as part of the Vera Rubin AI platform, the Groq 3 LPU is designed for low-latency, high-concurrency inference (running trained AI models), complementing rather than replacing Nvidia’s GPUs. Nvidia said the report mischaracterized the LPU’s specifications and implied strategic intentions related to China, where the company continues to navigate U.S. export restrictions on advanced chips. The chip has 500 MB of on-chip SRAM and 150 TB/s of inference bandwidth, while the LPX rack-scale system integrates 256 Groq 3 LPUs per rack. The technology targets interactive and agentic AI applications requiring predictable per-token latency, adding competitive pressure on AMD and AI chip startups focused on inference.