Overview
- Nvidia issued a public correction on Thursday saying The Information’s report was incorrect and that it has no LPU sales in China and no China-specific LPU on its roadmap.
- The Information had reported that Nvidia planned small-batch shipments of a Groq-based language processing unit for Chinese customers by year-end and that several customers had placed orders, a claim Nvidia disputes.
- Nvidia’s LPU work stems from a large late-2025 licensing deal with Groq and the Groq 3 LPU was unveiled at GTC 2026 as a low-latency inference accelerator meant to work alongside GPUs for real-time token generation.
- U.S. export controls that limit advanced memory and packaging access are a key reason the company’s China strategy is constrained, and earlier this year limited approvals for H200 sales to a few Chinese firms were reported.
- The public back-and-forth leaves the market uncertain and strengthens momentum for Chinese alternatives: domestic foundries such as SMIC, RISC‑V designs and near‑memory compute efforts are winning business and investment as suppliers adapt to restrictions.