Nvidia's Reported China Comeback Plan: A Groq-licensed LPU Inference Chip — and a Swift Denial

Just months after CEO Jensen Huang said Nvidia had “largely conceded” the China market, the chipmaker is reportedly plotting an unlikely route back in — one that doesn’t center on a GPU at all.

On Thursday, The Information reported that Nvidia plans to begin small-batch shipments of a new AI chip specially tailored for Chinese customers by the end of the year. The twist: according to details circulating from the report, the design is built around LPU (Language Processing Unit) technology licensed from Groq, intended to work alongside Nvidia’s GPUs to accelerate AI inference — while staying inside the lines drawn by US export controls.

Nvidia moved quickly to knock the story down. In a statement to Reuters, the company said: “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 the market. But as Wccftech noted in its coverage, much of Nvidia’s statement is anchored to the present status of its LPU plans — wording that leaves considerable room for future maneuverability. For a company that once shipped China-specific H800s, then H20s, then watched each successive carve-out get restricted, keeping its options quiet is almost a reflex.

Why an LPU sidesteps the export rulebook

US export controls on China-bound AI accelerators hinge largely on two levers: how much high-bandwidth memory (HBM) a chip carries, and access to advanced packaging technologies such as CoWoS. Nvidia’s traditional data-center GPUs are drenched in both — HBM stacks account for roughly half the cost of a flagship GPU, and CoWoS packaging is what fuses the compute die to that memory in the first place.

An LPU-based design is a different animal. Groq’s Language Processing Units cluster hundreds or thousands of specialized chips, each carrying giant blocks of matrix-multiply (MXM) and vector (VXM) execution units alongside roughly 230MB of blazing-fast on-die SRAM. Model weights are baked directly into that SRAM, bypassing the conventional memory hierarchy entirely. There are no branch predictors and no hardware schedulers — the Groq compiler plans every calculation down to the nanosecond so data arrives from SRAM in a deterministic, continuously scheduled cadence. The result is inference throughput that is extraordinarily fast and power-efficient per watt, without needing the massive HBM payloads that trigger US licensing thresholds.

Nvidia acquired the rights to this playbook in December 2025, when it struck a technology-licensing and asset/talent-transfer agreement with Groq — a deal reported at roughly $20 billion in value that handed Nvidia a non-exclusive license to Groq’s LPU inference technology plus its chip and systems teams. A China-specific inference chip built on that foundation could, in principle, offer Chinese customers competitive tokens-per-dollar for the inference workloads that dominate 2026’s agentic-AI demand, while carrying little of the HBM and advanced packaging that Washington restricts.

The context Nvidia is fighting against

The timing of the report says as much as the technical details. Beijing is actively discouraging Chinese enterprises from importing the H200 GPUs that the Trump administration specifically authorized for sale in China — a rare alignment of US permission and Chinese refusal that has left Nvidia’s flagship China-legal product commercially stranded. Export licenses exist; demand is being strangled from the other direction.

Meanwhile, domestic substitution is compounding. TrendForce projects that Nvidia and AMD’s combined share of China’s high-end AI chip market will plunge to roughly 10% in 2026, with Chinese domestic chips seizing the remaining ~90% — and Nvidia’s own slice falling to around 8%, down from approximately 40% a year earlier. Separate reporting in July found Chinese firms raising their domestic AI chip budget share from 30% to 46% as they shift away from Nvidia.

The numbers behind China’s import-substitution push are no longer theoretical:

• SMIC, China’s only 7nm-class foundry, posted quarterly revenue of $3.01 billion (up 36.1% year-over-year) with a 93.7% utilization rate, and net profit nearly tripling to $479.2 million.
• Alibaba unveiled the XuanTie C950 in March — a TSMC-built 5nm, 64-core RISC-V server processor with embedded matrix and vector acceleration engines, capable of natively running Qwen models without GPUs.
• DFSX is taking a different route entirely: its 14nm DF1000 stacks memory directly atop the compute die via 3D wafer-level hybrid bonding, and its DF2000 “supernode” — expected in Q4 2026 — arranges multiple stacked memory-compute towers in what the company calls a 3.5D Infinity Chiplet layout, claiming twice the memory bandwidth of Nvidia’s GB200 NVL72.
• Shanghai Aishegna, a secretive lithography venture, has entered the DUV business with plans for 5 machines this year and 20 more next year, as Beijing hedges toward glass substrates and even an EUV prototype.

Against that backdrop, a Groq-derived inference chip is Nvidia’s creative attempt to remain relevant in a market that is closing on two fronts at once — Washington’s export rules limiting what it may sell, and Beijing’s industrial policy discouraging customers from buying.

Denial today, product tomorrow?

The honest read of this week’s news is that we have a report and a denial, and the truth likely lives between them. Nvidia’s statement — “no LPU sales in the China market today, and no China-specific LPU” — does not foreclose a product in development. The Information cited two employees for its claim of year-end small-batch shipments, and the strategic logic is difficult to dismiss: with TrendForce-style projections showing Nvidia’s China share collapsing toward single digits, and with inference (not training) becoming the dominant workload of the agentic-AI era, an LPU-class accelerator that threads the export-control needle may be the only legal product lane left.

There are real obstacles beyond politics. Chinese cloud giants would have to integrate a novel architecture into software stacks they have spent two years re-tooling around domestic silicon like Ascend and the RISC-V alternatives. Groq’s SRAM-heavy design trades HBM cost for die area, which affects pricing. And Washington’s thresholds can be tightened again at any time — the BIS revised its license-review policy for China-bound semiconductors as recently as January 2026, and Congress has repeatedly floated legislation to kill China-specific chip variants outright.

Still, the episode marks a striking inflection: the company that defined the AI-accelerator era is now reportedly reduced to licensing a rival’s architecture to attempt re-entry into its former second-largest market — while the world’s largest foundry customer base inside China consolidates around homegrown chips. Whether or not an “LPU for China” ships this year, the deeper story is already decided: the bifurcation of the global AI chip market is essentially complete, and every actor on both sides is now optimizing within its own half.

Watch for three signals in the coming months: any regulatory filing or supply-chain evidence of small-batch LPU shipments into China, Nvidia’s commentary on China revenue in its next earnings call, and whether Beijing’s “buy domestic” guidance softens for inference-class hardware that Chinese foundries can’t yet match on efficiency.