Somewhere in Inner Mongolia, in a data cluster that does not appear on any export compliance filing, several thousand Nvidia Blackwell GPUs have been running for months. According to a senior Trump administration official who spoke to reporters on April 10, those chips were used to train DeepSeek V4, the Chinese AI lab's next flagship model, scheduled for public release within days.
The Blackwell GPU is Nvidia's most powerful AI chip. It is also banned from export to China under U.S. restrictions designed to prevent the country from building frontier AI systems. The chips were never supposed to leave their country of origin.
The official said DeepSeek plans to erase the technical traces of its American chip usage before releasing the model, making enforcement after the fact nearly impossible. "DeepSeek will erase technical traces of its American chip usage, making it harder to detect," the official stated.
The allegation, if confirmed, would represent the most significant known violation of U.S. AI chip export controls since their introduction in October 2022.
The Alleged Smuggling Operation Reads Like a Spy Novel
The mechanics of how banned chips allegedly reached China were first reported by The Information and corroborated by Tom's Hardware and multiple outlets citing intelligence sources.
The operation works like this: shell companies in Southeast Asia purchase entire data centers' worth of Nvidia servers, setting up the hardware entirely to specification. Nvidia's OEM partners send contractors to inspect the installations, confirming successful setup and export compliance. The inspectors leave satisfied. Then the smugglers go to work.
After each inspection, the data centers are allegedly disassembled rack by rack. The GPU servers are packed into suitcases and shipped across the border into mainland China, where U.S. restrictions prohibit their sale and use. Sources told The Information that smugglers and their clients prefer 8-GPU rack servers like the HGX B200 over the more powerful GB200 NVL72, because the smaller form factor is easier to transport covertly.
The scale is not trivial. Reports indicate that "several thousand" Blackwell GPUs made the journey. At current market prices, that represents hardware worth hundreds of millions of dollars.
Nvidia's response was terse. The company told The Information it had "not seen evidence of this" and called the smuggling claims "farfetched." Nvidia did not respond to follow-up requests for comment from other outlets.
DeepSeek denied using Blackwell chips entirely. The company claims V4 was trained on Nvidia H800 chips, which were legal to export to China before the restrictions tightened, and Huawei Ascend 910C processors.
V4 Is a Bet on Chinese Chip Independence
Whatever hardware DeepSeek actually used, the model itself represents a major technical ambition.
| Specification | DeepSeek V4 (reported) |
|---|---|
| Parameters | Trillion-level |
| Context window | 1 million tokens |
| Modalities | Text, images, video |
| License | Apache 2.0 (free commercial use) |
| Key feature | Long-Term Memory (LTM) |
| Hardware optimization | Huawei Ascend native support |
| Expected release | Late April 2026 |
The Long-Term Memory system is designed to let the model retain and reason over information across extended interactions, a capability that would distinguish V4 from most existing frontier models.
The most geopolitically significant detail is the model's deep native optimization for Huawei Ascend processors. DeepSeek has framed this as part of China's broader "de-CUDAfication" effort, the push to build AI infrastructure that does not depend on Nvidia's CUDA software ecosystem. If V4 performs well on Huawei hardware, it would be the strongest signal yet that Chinese AI labs can build frontier models without relying on American chips.
A March 29 outage on DeepSeek's platform, which lasted 13 hours, is widely suspected to have been infrastructure testing for V4's deployment. Developers who accessed the system during the outage noted improved coding logic and better SVG generation, hinting at the new model's capabilities.
Stephen Wu of Carthage Capital told AFP that a successful V4 launch "signals a material shift in the geopolitical tech landscape." Wei Sun, a researcher at Counterpoint Research, said the release matters because "at one level, it is a signal of China's AI self-sufficiency trajectory."
The Export Control Regime Has a Credibility Problem
The smuggling allegations land at the worst possible moment for U.S. policymakers who have staked significant diplomatic capital on the idea that export controls can slow China's AI progress.
The controls were first introduced in October 2022 and tightened in October 2023 and again in late 2024. Each round added new chip categories to the restricted list and closed loopholes that Chinese firms had exploited. Nvidia designed the H800 specifically as a China-legal alternative to its more powerful H100, downgrading chip-to-chip interconnect speeds to comply with the rules. When that was restricted too, Chinese firms turned to Huawei's domestically produced alternatives.
The Blackwell smuggling allegations suggest a different approach: ignoring the rules entirely.
If the claims are accurate, the shell company technique exposes a fundamental weakness in the enforcement model. Export controls rely heavily on end-use verification, the process of confirming that chips shipped to approved buyers stay where they are supposed to be. When legitimate-looking data centers are purpose-built to pass inspection and then stripped for parts, that verification becomes performative.
The timing is also awkward. Nvidia reported $35.6 billion in data center revenue for Q4 2025, with China remaining a significant market even under restrictions. The company has consistently lobbied against tighter controls, arguing that they push Chinese firms toward domestic alternatives without actually slowing their progress. The smuggling reports give that argument an uncomfortable new dimension: if Chinese firms are obtaining the restricted chips anyway, the controls may be failing to achieve either goal.
The Other Side: Reasons for Skepticism
Not everyone is convinced the smuggling claims hold up.
Nvidia's dismissal was not the only skeptical voice. Analysts at Digitimes noted that Nvidia had found "no concrete evidence" in its own investigation. The logistics of disassembling entire data centers, packing GPUs into suitcases, and reassembling them in China without Nvidia's monitoring tools detecting the relocation strain credulity for some observers.
DeepSeek's own denial was specific: the company named the H800 and Huawei Ascend 910C as its training hardware. The H800 was legally available in China before the latest rounds of restrictions, and existing stockpiles are not covered by retroactive enforcement. The Ascend 910C, Huawei's most capable AI chip, has shown steady performance improvements. It is not implausible that V4 was trained on this combination.
There is also a motivation question. The Trump administration official who disclosed the Blackwell allegation did so a week before DeepSeek's planned release, raising the possibility that the timing was chosen to undermine the launch's credibility rather than to announce an enforcement action. No seizures, arrests, or sanctions were announced alongside the disclosure.
The truth may be more complicated than either narrative allows. DeepSeek may have used a mix of hardware: legally obtained H800 stockpiles, domestically produced Ascend chips, and a smaller number of smuggled Blackwell units for targeted training runs. The "de-CUDAfication" narrative and the smuggling narrative are not mutually exclusive.
The Competitive Landscape Is Getting Crowded
V4 is not launching into a vacuum. China's AI sector is in the middle of its most intense model release cycle ever.
Alibaba's Qwen 3.6 Plus, released April 2, features a million-token native context window and a hybrid architecture combining linear attention with sparse mixture-of-experts routing. Zhipu AI's GLM-5.1 reached the third spot in Code Arena rankings, surpassing Gemini 3.1 and GPT-5.4 in coding benchmarks. MiniMax open-sourced its M2.7 model with state-of-the-art scores on SWE-Pro (56.22%) and Terminal Bench 2 (57.0%). ByteDance, Tencent, and Baidu all have models in active development.
The demand for hardware is surging alongside the model race. Alibaba, ByteDance, and Tencent have placed advance orders for hundreds of thousands of next-generation AI chips, driving prices up approximately 20%, according to BigGo Finance.
For the U.S. AI companies that recently agreed to share intelligence on Chinese model-copying through the Frontier Model Forum, the V4 launch will be a critical test. DeepSeek's previous model, R1, released in January 2025, triggered a sell-off in U.S. tech stocks and prompted President Trump to call it a "wake-up call." V4, with its trillion-scale parameters and multimodal capabilities, could be a larger shock.
The Bottom Line
A trillion-parameter model is about to go live under an open-source license, built by a Chinese lab that the U.S. government says trained it on chips that were never supposed to leave American-allied soil. The lab denies the allegation. Nvidia calls it implausible. The government has provided no public evidence beyond a single official's statement. And no enforcement action has followed.
What is not in dispute is that DeepSeek V4 exists, that it will be released within days, and that it will be free for anyone to use. If it performs anywhere near its specifications, the question of which chips trained it will matter less to the AI industry than what it can do.
As Counterpoint's Wei Sun framed it, V4 is "a signal of China's AI self-sufficiency trajectory." Whether that trajectory runs through smuggled suitcases in Southeast Asia or domestically built silicon in Shenzhen, the destination is the same.
Sources
- China's DeepSeek Uses Nvidia's Blackwell Chip Despite US Ban — Washington Today (April 10, 2026)
- Nvidia decries 'far-fetched' reports of smuggling in face of DeepSeek training reports — Tom's Hardware (April 10, 2026)
- Waiting for DeepSeek: New model to test China's AI ambitions — TechXplore/AFP (April 11, 2026)
- DeepSeek V4 Launch Imminent: Trillion-Parameter Model to Pioneer Chinese AI Chip Independence — BigGo Finance (April 10, 2026)
- New DeepSeek model to test China's AI ambitions — Taipei Times (April 12, 2026)
- Waiting for DeepSeek: new model to test China's AI ambitions — Hong Kong Free Press (April 11, 2026)
- DeepSeek reportedly using smuggled Nvidia Blackwell chips for AI model development — Yahoo Finance/Investing.com (April 10, 2026)
- DeepSeek is Using Banned Nvidia Chips in Race to Build Next Model — The Information (April 2026)
- DeepSeek V4 and Qwen 3.5: Open-Source AI Is Rewriting the Rules in 2026 — Particula (April 2026)