Reported from two employees by The Information: Nvidia is engineering a US-compliant inference chip for China via a software workaround — not a return, but a narrow attempt to hold a position before domestic silicon locks the market shut.
How We Got Here — Key Moments
Late 2025
Nvidia licenses language processing unit (LPU) technology from Groq in a deal reported at roughly $20 billion — structured as an IP license, not a full acquisition, to avoid merger-review thresholds.
March 2026
Reuters reports Nvidia preparing a Groq-derived chip for China. CEO Jensen Huang calls the report “totally false.”
May 2026
Huang says Nvidia has “largely conceded” China’s AI-chip market to Huawei — framing that matters when evaluating the current reported maneuvering.
Q2 2026
Amkor reported to have finished over one million H200 units — chips subject to US export restrictions — raising questions about the blockade’s permeability. Moonshot AI halts new Kimi sign-ups; DeepSeek raises API prices on its latest models. Both signal an acute inference-capacity shortage inside China.
August 2026 — Reported
The Information, citing two employees, reports Nvidia plans small-batch LPU shipments to China by year-end. Chinese customers have reportedly placed orders. Nvidia declined to comment. Shipments are planned, not confirmed. Beijing approval: pending and uncertain.
The Information, citing two employees with knowledge of the plans, reports that Nvidia is preparing small-batch shipments of a China-tailored AI chip by the end of 2026 — a variant of its LPU, or language processing unit, an inference-acceleration chip that speeds up how quickly an AI system generates responses. The chip is built on technology Nvidia licensed from Groq in a deal reported at roughly $20 billion, structured as an IP license specifically — not a clean acquisition — a distinction that matters for regulatory review. Chinese customers have reportedly placed orders. The chip is said to comply with current US export rules. Nvidia declined to comment.
The hedges belong at the front, because they are the whole story. These shipments are planned for year-end, not confirmed. The orders are placed but not filled. And whether this chip ever reaches Chinese customers depends not only on Washington’s export-control regime but also on Beijing, which has a documented pattern of blocking or discouraging even chips Washington has formally cleared. The prior denial also deserves careful handling: when Reuters reported in March that Nvidia was preparing a Groq-derived chip for China, Huang called it “totally false.” The current reporting points in a different direction — but plans change, denials are sometimes narrowly worded, and Nvidia’s silence now is not a confession. What can be said plainly is that the public messaging and the reported internal maneuvering are pointing in different directions.
Huang’s May statement that Nvidia had “largely conceded” China’s AI-chip market to Huawei is also worth holding onto. A compliant inference chip aimed at a specific capacity shortage is a partial re-entry attempt — an effort to hold a slice — not a reversal of that concession. The reported maneuvering around the blockade’s edges — banned H200s moving through resellers, gaming chips modified to run AI workloads, over a million H200s reportedly finished at Amkor in Q2, and Nvidia said to be developing two or more new China gaming chips that could be adapted for AI — all come from sourced reporting and should be read as reported, not confirmed.
The key insight: Nvidia did not weaken the LPU to clear export rules — the chip already complied. Instead, it reportedly modified the software so the LPU can operate alongside processors available inside China, bypassing the need for Vera Rubin, which cannot be sold there. The compliance happens at the software layer. That is not a workaround in the pejorative sense; it is a precise demonstration of how software-defined systems handle hardware constraints — and it is the same logic Nvidia uses to argue its compute stack is fungible across configurations.
The Structural Read
Read this story not as a chip launch but as a report about where water is finding its way through a dam. Export controls were engineered to wall off the most advanced training silicon — and they have had real effect at that layer. But demand does not disappear when supply is blocked; it routes around. The same reporting documents the routing: reseller channels for restricted H200s, gaming chips quietly adapted for AI inference workloads, and now a purpose-built, export-compliant inference chip aimed straight at the gap the controls left open.
That gap is specifically on inference — and that is the part of the story that carries the most structural weight. China’s bottleneck right now is not training frontier models. It is serving them. Moonshot AI halting new Kimi sign-ups when demand overwhelmed compute and DeepSeek raising prices on its latest API models are not unrelated data points — they are symptoms of the same inference-capacity shortage. An LPU is an inference accelerator. Nvidia is reportedly aiming at exactly the shortage that is most acute.
The timing is the strategy. Huawei’s Ascend 950DT is shipping into the same inference gap. Chinese big tech is reportedly booking domestic compute capacity through 2027. The window before Chinese customers standardize on domestic silicon — before the relationships, the driver stacks, and the operational muscle memory all harden around local chips — is genuinely narrow. Get a compliant chip in now and you preserve the relationship and the software ecosystem hook. Arrive after lock-in and the market is inaccessible for a generation. As explored in Beyond Nvidia’s Moat, the real moat is not the chip — it is the installed base and the software stack that runs on it.
The inference fork unfolding across the chip industry makes this more consequential, not less. As training and inference diverge into distinct hardware requirements, the controls that targeted training silicon leave inference relatively exposed — and that is precisely where this reported chip is positioned. The memory supply chain is following the same logic: policy shapes where capital flows, and capital flows around the blockage points.
The Permission Layer — Business Engineer Framework
“The China AI-chip market is no longer a market in the ordinary sense. It is a permission layer — and every transaction has to clear two governments that do not agree. A chip that complies with US export rules is not a chip Nvidia controls the fate of. Beijing holds a veto that Washington’s rule-writers did not write and cannot lift.”
Three Implications
EXPORT CONTROLS AS A LEAKY DAM — DEMAND ROUTES, NOT DISAPPEARS
The structural lesson from this reported episode is not that Nvidia is circumventing controls — the LPU reportedly complies — it is that controls built around training silicon leave inference exposed, and market pressure fills the gap. Reseller channels, modified gaming chips, and purpose-built compliant accelerators are all expressions of the same underlying dynamic: blocked demand finds a path. The relevant policy question is not whether the dam holds everywhere, but which gaps matter most. Right now, inference is the gap that matters.
SOFTWARE-DEFINED REGULATORY ARBITRAGE — THE HARDWARE IS FIXED, SOFTWARE DECIDES THE SYSTEM
Nvidia reportedly did not derate the LPU’s hardware to comply. It rewrote the software so the chip can operate with non-Rubin processors available inside China. That is regulatory arbitrage at the software layer — the same lever that makes CUDA so durable as a competitive moat. When hardware is constrained by policy, software determines what configuration the hardware plugs into. This is a repeatable playbook, not a one-time move, and it has implications for how export controls will need to be written going forward if they are to reach the software integration layer.
THE TWO-GOVERNMENT SQUEEZE — THE OUTCOME IS GENUINELY OPEN
Neither government has said yes. Washington’s export rules may permit this chip; Beijing’s approval is uncertain and pending. Beijing has previously blocked or discouraged chips Washington cleared, and Chinese big tech is already booking domestic Huawei and homegrown capacity through 2027. Even if the chip ships, the competitive question is open: Huawei’s Ascend 950DT is aimed at the same inference shortage, domestic supply chains are building momentum, and a chip that requires Beijing’s blessing for every batch sale is not a chip Nvidia can scale on its own terms. The window is narrow, both governments hold a veto, and the outcome depends on which lock-in — domestic or foreign — solidifies first.
Business Engineer Framework
The Map of AI Redrawn — The Permission Layer
The Permission Layer framework explains why the AI stack’s most consequential variable is increasingly neither the model nor the chip — it is which governments allow which silicon to move where, and how fast domestic supply chains can standardize before a foreign competitor gets back in. The Map of AI Redrawn traces the nine layers of the AI stack and shows exactly where the permission bottlenecks sit — and where the next competitive windows are opening.
Export controls do not eliminate demand — they redirect it, and the redirection is already underway across reseller channels, adapted gaming chips, and now a reported purpose-built inference accelerator that complies at the hardware level and routes around the Rubin ban at the software level. Nvidia is not back in China; it is attempting to hold a position in a market it publicly said it had largely conceded, through a narrow window, before domestic silicon lock-in makes re-entry structurally impossible. Whether the window is open is a question two governments have not yet answered — and until they do, “planned shipments” and “placed orders” are the honest outer boundary of what the facts support.
Gennaro is the creator of FourWeekMBA, which reached about four million business people, comprising C-level executives, investors, analysts, product managers, and aspiring digital entrepreneurs in 2022 alone | He is also Director of Sales for a high-tech scaleup in the AI Industry | In 2012, Gennaro earned an International MBA with emphasis on Corporate Finance and Business Strategy.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
