Citing two Nvidia employees, The Information reported that the company was preparing to ship a small batch of LPU chips to Chinese customers before the end of this year, with some customers already having placed orders. According to the report, the chip complies with current U.S. export controls and is designed to work alongside processors already available in China.
On Thursday U.S. time, an Nvidia spokesperson pushed back on the report line by line: "The reporting in The Information on Nvidia's LPU is incorrect. We have no LPU sales in the China market today, and no China-specific LPU product in our roadmap."
Nvidia shares closed down 0.33% that day at $216.85, while the Nasdaq fell 1.00% in the same session. The market's reaction to the news stayed at noise level.
Two Sentences, Two Different Timeframes
The spokesperson's statement splits into two parts — one about the present, one about the future. Both are technically true, and neither addresses what the report actually described: a small sample shipment, customer orders already placed, and whether a license can be obtained. This kind of phrasing is common in corporate PR — the denial covers a "product line," while the report was about "a batch of chips."
Nvidia's public position on its China business has kept shifting. In May this year, Washington approved limited H200 sales to companies including Alibaba, Tencent and ByteDance, cracking the gate open. Jensen Huang has separately acknowledged that Nvidia has already ceded a large share of China's AI chip market to domestic vendors. A company that has already given up market share doesn't need to put a comeback product on a public roadmap in advance.
Where the LPU Came From
The name LPU comes from Groq, a company that builds chips dedicated to inference. Its selling point is squeezing the latency of token generation for large models down to a minimum, dividing labor with general-purpose GPUs: GPUs handle training and heavy computation, while LPUs get a chatbot's reply out as fast as possible.
Last December, Nvidia acquired Groq's technology and core team for roughly $20 billion, and unveiled the Groq 3 chip along with a matching CPU server at GTC this year. An acquisition that size needs to be amortized, and amortization needs shipment volume. China is one of the fastest-growing markets for inference demand worldwide, and pushing a latency-oriented, lower-compute-density inference chip into that market is far easier than squeezing in a training card under the ceiling set by export controls.
That division of labor carries extra weight in the China market. The domestic training-chip ecosystem is still catching up, but the bar for inference is much lower — once a model is running, what matters is cost per unit and time-to-first-token. An accelerator that only handles inference, not training, wouldn't directly disrupt the pace of domestic training-chip substitution, and in theory carries a lower degree of policy sensitivity. That's likely the underlying reason this kind of rumor keeps resurfacing.
Export controls focus mainly on metrics like total compute and interconnect bandwidth, and inference chips naturally don't score high on either. The line in The Information's report about the chip working "alongside processors already available in China" points to the same path: not a full system, just an accelerator card, leaving the rest to the local supply chain.
A Rough Tally
Nvidia's China data-center business has gone from a major revenue line two years ago to something no longer called out separately in earnings reports. Using the publicly reported $81.6 billion in revenue for the most recent quarter as a rough baseline, even a recovery to just 5% in the China market would mean a quarterly gain in the range of $4 billion — close to a mid-sized chipmaker's full-year revenue. That scale is enough to explain why rumors about a China-specific chip keep surfacing every few months.
For buyers in China, the math runs the other way. H200 quotas are limited and approval takes a long time, and domestic inference-chip capacity can't fill the gap. If a low-latency, CUDA-compatible inference accelerator actually made it in, it wouldn't replace training clusters — it would land in inference facilities running online services that are sensitive to first-token latency.
So the denial is one thing, but the underlying commercial logic hasn't been denied. What actually blocks it is licensing. Over the past two years, every China-specific chip Nvidia has designed has had to go through a review process in Washington, and approval has often had little to do with the chip's own specifications. A card that only handles inference, not training, is in theory easier to clear than a training card — but the approval standard has long since expanded beyond raw performance figures to whether it helps a rival build independent AI capability, a boundary far blurrier than any number.
Nvidia's denial was clean and direct. What it didn't say: if a license were ever granted, when the roadmap would change — and whether that change would ever be announced.
Sources: The Information, Reuters, CocoLoop, Investing.com; Nvidia's spokesperson statement and closing price move were checked line by line.