Nvidia prepares Groq AI chip for China, sources say

Reuters reported, citing two sources familiar with the matter, that Nvidia is preparing a version of the Groq AI chip for sale in China.

Nvidia bought Groq late last year for US$17 billion and showcased a new product lineup at its San Jose developer conference.

Reuters said Nvidia plans to use Groq chips for inference and to pair them with its forthcoming Vera Rubin chips that cannot be sold in China.

The report comes as Jensen Huang, Nvidia CEO, said the company restarted production of H200 chips after obtaining export licenses from US President Donald Trump’s administration and receiving purchase orders from Chinese customers.

While Nvidia dominates AI training, it faces more competition in inference from Chinese firms such as Baidu that already make inference chips.

One of the sources told Reuters the China-ready Groq variant is not a downgraded version or made specifically for the Chinese market. It can be adapted to work with other systems and is expected to be available in May.

🔗 Source: Reuters

🧠 Food for thought

Implications, context, and why it matters.

Nvidia’s Groq LPUs are positioned as inference accelerators alongside Rubin GPUs

Large language model inference runs in two stages. The compute-heavy “prefill” stage processes the prompt, then the memory-heavy “decode” stage generates the response token-by-token 1.
Nvidia’s Rubin GPUs target prefill work. Groq’s language processing units (LPUs) target decode work and add capacity rather than replacing Nvidia GPUs 1.
Each Groq 3 LPU includes 500 MB of on-board memory. Nvidia’s VP ian buck puts its memory bandwidth at about 150 TB/s, versus 22 TB/s of High Bandwidth Memory 4 (HBM4) bandwidth for Rubin GPUs, which makes the LPU fit decode-heavy workloads 1.
Groq’s LPUs do not support CUDA natively yet. Nvidia says it is using the LPU as an accelerator while CUDA runs on the Vera Rubin NVL72 platform 1.

Nvidia’s China chip strategy is a hedge against policy uncertainty and capacity constraints

Nvidia restarted H200 production for China after receiving export licenses and purchase orders. A Financial Times investigation cited by another report says Nvidia also halted H200 production set aside for Chinese customers and shifted TSMC capacity to its Vera Rubin platform 2.
Global demand for Rubin and Blackwell is described as “insatiable,” so steering capacity toward China-compliant chips can carry a real tradeoff 2.
China-compliant parts can bring thinner margins, so supplying restricted markets may cost more than it pays compared with unrestricted sales 2.
A Groq-based inference path could keep Nvidia involved in China inference without selling Rubin systems there, since Rubin chips cannot be sold in China under the export limits described in 1.