Via dwglogo.com
Nvidia says a recent article got key details wrong about its language processing unit, a $20B licensed chip designed to complement its GPU lineup
Nvidia issued a public correction this week targeting a report by The Information that, according to Nvidia, contained inaccurate claims about the Groq 3 Language Processing Unit and the strategic direction surrounding it.
The denial is notable for the context around it. Nvidia’s LPU business is built on a roughly $20 billion licensing agreement with chip startup Groq that was announced in late 2025.
What the Groq 3 LPU actually is #
Nvidia’s GPUs excel at high-throughput training and batch inference, processing massive amounts of data simultaneously. The Groq 3 LPU targets something different: low-latency, high-concurrency inference workloads, the kind of tasks where a user is waiting in real time for an AI response.
The Groq 3 LPU was introduced publicly at Nvidia’s GTC 2026 event on March 16, 2026, as part of the company’s Vera Rubin AI platform. Each chip carries 500 MB of on-chip SRAM and delivers 150 TB/s of bandwidth for inference tasks. In rack-scale deployments, the LPX system integrates 256 Groq 3 LPUs per rack, built specifically to handle ultra-low-latency requirements at scale.
Nvidia has been explicit that the LPU is meant to work alongside its GPU lineup, not replace it. GPUs handle the heavy lifting of training and bulk inference, while LPUs manage interactive and agentic AI applications where predictable per-token latency matters most.
What The Information got wrong, and why Nvidia responded #
The Information published its report on or around August 20, 2026. Nvidia’s response singled out the article’s characterization of the Groq 3 LPU’s specifications and, separately, what the report implied about Nvidia’s strategic intentions, specifically around China.
The Information article reportedly touched on the possibility of Nvidia using the LPU as part of a strategy to re-enter the Chinese market, a subject that carries significant weight given Nvidia’s ongoing navigation of US export restrictions on advanced chips. Nvidia pushed back on this framing.
The bigger competitive picture #
Groq, before the licensing deal, had built a following among developers specifically because its chips delivered remarkably fast token generation speeds for inference tasks. The company’s architecture, centered on large on-chip memory and deterministic execution, solved real problems for applications where latency is the bottleneck rather than raw throughput.
By licensing that technology and folding it into the Vera Rubin platform, Nvidia is essentially telling customers they no longer need to go elsewhere for fast inference, putting pressure on alternatives like AMD’s inference-focused efforts and a growing list of AI chip startups competing on the same low-latency pitch.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our