Nvidia told Tom’s Hardware on Thursday that a report claiming it would ship a China-specific chip by year-end is wrong. “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,” a company spokesperson said.
The Information’s report, citing two Nvidia employees, described small-batch shipments beginning by the end of 2026 of a variant of the Groq 3 LPU, the decode co-processor Nvidia announced at GTC in March for use alongside its Vera Rubin platform. Because Vera Rubin cannot legally be sold in China, the report said Nvidia had rewritten the software that splits work between GPU and LPU so the accelerator could instead pair with processors already permitted for Chinese customers, and that several Chinese buyers had already placed orders. The Information said Nvidia did not respond to requests for comment over several days before it published.
The chip’s specifications make the claim technically plausible regardless of whether it is true. The Groq 3 LPU is built on Samsung’s 4-nanometer process with 512 megabytes of SRAM per die and no HBM memory stack, and Nvidia said at GTC it would begin shipping to customers, OpenAI among them, in the third quarter of this year. US export thresholds for China are set on compute density and memory bandwidth, and a decode accelerator built entirely around on-die SRAM, with no HBM, is the kind of part that can clear those limits without needing a deliberately cut-down version, the mechanism The Information’s sources described.
This is the second time this year Nvidia has had to deny it. A similar report surfaced in March, and chief executive Jensen Huang called it “totally false” within two days. Thursday’s denial is narrower than Huang’s was, addressing current sales and ruling out a China-specific product without directly saying whether the standard Groq 3 LPU itself might ship there.
Even a green light from Nvidia would not settle the question. China blocked purchases of Nvidia’s H20 last year over national-security objections, and only in the past two weeks did Beijing let ByteDance and Tencent take delivery of H200s it had cleared months earlier, and even then only by routing the chips through Hong Kong rather than the mainland. Chinese firms that cannot buy the hardware outright have also been renting access to it remotely through data centers in third countries, a workaround US regulators are still deciding how to police. In each of these cases, the recent bottleneck has been Beijing’s own approval process, not Washington’s export rules. Huang told CNBC in May that Nvidia had “largely conceded” China’s AI chip market to Huawei, whose Ascend 950DT the Information named as the LPU’s direct domestic competitor.
Whether Nvidia ships nothing, the standard LPU, or the variant The Information described, the chip is caught between two governments that have each, independently, made the decision harder than an engineering roadmap can resolve on its own.
Sources: Tom's Hardware
–
By the Control Plane Editorial Team