WASHINGTON — A senior Trump administration official has told reporters that Chinese AI startup DeepSeek trained its forthcoming artificial intelligence model on Nvidia's most advanced commercially available processor, the Blackwell chip, despite comprehensive US export controls that explicitly bar those semiconductors from shipment to China. The disclosure, which surfaced on February 24 and was first reported by Reuters, represents the first time the US government has on-record confirmed the use of restricted hardware for a specific Chinese AI development program, and it arrives at a moment of acute policy tension in Washington over how tightly to police the global flow of high-end computing silicon.

What the Government Says Happened

The official, who spoke without revealing how the US government obtained the information, said American authorities believe the Blackwell processors used to train the model are physically located inside a DeepSeek-operated data center cluster in Inner Mongolia, an autonomous region in northern China. The official added that the US expects DeepSeek to strip out technical metadata and firmware identifiers that might reveal the origin of the chips before the model is publicly released, which was described as potentially imminent at the time of the briefing.

Asked directly how DeepSeek came to possess hardware that is prohibited under the Export Administration Regulations administered by the Commerce Department, the official declined to explain the acquisition pathway. The official did, however, state the administration's standing posture in unambiguous terms: "We're not shipping Blackwells to China." Nvidia declined to comment on the report. The Commerce Department and DeepSeek did not respond to requests for comment, and the Chinese embassy in Washington pushed back on the framing, stating that Beijing opposes what it described as the overstretching of national security concepts and the politicization of trade and technology questions.

The Blackwell Architecture and Why It Matters

The Nvidia Blackwell GPU architecture, which entered commercial production in late 2024, represents the company's most capable AI training platform to date. The chips offer substantially higher memory bandwidth and computational throughput than the preceding Hopper generation and are the processors powering the majority of frontier-scale AI training runs currently being conducted by US hyperscalers. Their classification under strict export control thresholds means they cannot be legally sold to Chinese entities through normal commercial channels.

Earlier in 2025, the Trump administration briefly signaled openness to licensing a scaled-down Blackwell variant for Chinese customers, but reversed course, reaffirming that its most advanced hardware should remain unavailable to Chinese purchasers. A separate December 2024 decision allowed Chinese firms to apply on a case-by-case basis to acquire the H200, which is Nvidia's second-tier chip and one generation behind the Blackwell. Those approvals remain subject to regulatory guardrails, and the H200 remains substantially less capable than the Blackwell for training large-scale foundation models. The distinction between the two chip classes is not merely symbolic: the difference in performance translates into meaningful gaps in training speed, model size, and capability ceiling for frontier AI systems.

The allegations do not stop at hardware. The same administration official said the model that DeepSeek trained on Blackwell chips also likely relied on a technique called knowledge distillation, applied to outputs from several leading US AI firms. Distillation, in this context, refers to a process in which a newer, less-established AI model is iteratively evaluated and refined using the outputs of a more powerful existing model, allowing the newer system to absorb learned representations without having been trained from scratch on raw data. The process can substantially reduce training cost and time while producing models that inherit many of the behavioral and reasoning characteristics of the more capable parent.

The official said the US believes the DeepSeek model in question drew on outputs from US AI companies. This echoes accusations already leveled in separate legal and policy contexts by two of those companies, which have both separately alleged that DeepSeek engaged in unauthorized appropriation of their model outputs. Neither allegation has been adjudicated, and DeepSeek has not formally responded to the substance of those claims. The double-layered accusation, combining alleged hardware smuggling with alleged model distillation from restricted sources, paints a picture of a development process that sidestepped US restrictions on multiple levels simultaneously.

The Policy Fault Lines in Washington

The revelation has sharpened a divide inside the administration that has been building for months. On one side sit national security hawks, who argue that any access by Chinese commercial AI firms to top-tier semiconductors poses a direct military risk. Their case rests on the dual-use nature of the hardware: chips powerful enough to train frontier AI models are by definition capable of accelerating military simulation, signals intelligence, and autonomous weapons research. The unauthorized presence of Blackwell processors in China, they argue, demonstrates that export controls as currently structured are failing to contain the most sensitive technology.

On the other side, voices including the White House's AI policy coordinator and Nvidia's chief executive have contended that blanket restrictions may be counterproductive. Their argument is that denying Chinese firms access to US hardware primarily incentivizes domestic Chinese chip development, benefiting companies like Huawei, which is developing its own competing AI processor line. Under this logic, selective access keeps Chinese AI firms dependent on US supply chains and slows domestic alternatives from maturing. A former senior National Security Council official who worked on technology policy under the Biden administration stated publicly that the DeepSeek case demonstrates precisely the domestic chip shortfall that hawks have warned about, and argued that permitting even lower-tier Blackwell equivalents to reach Chinese firms would constitute a meaningful concession.

Market and Strategic Implications

Nvidia's stock closed at $192.82 on February 24, 2026 on the Nasdaq, within a daily trading range of $187.40 to $193.77. The DeepSeek-Blackwell story broke during regular trading hours but did not produce a sharp negative reaction in Nvidia's share price, suggesting markets interpreted the underlying demand signal, proof that even China's best-resourced AI firm wants Blackwell chips above all else, as broadly supportive of the thesis that Nvidia hardware remains irreplaceable at the frontier. The company had also been expected to report fiscal fourth-quarter 2026 earnings on February 25, and investors were already navigating significant uncertainty in the stock ahead of that report, which historically had generated large post-announcement moves in either direction.

For the broader semiconductor export control framework, the implications may prove more durable than any single-day stock move. If Blackwell chips reached DeepSeek through third-party intermediaries, entrepot jurisdictions, or gray-market resellers, it suggests that the current enforcement apparatus has material gaps. The chips could have transited through Southeast Asian markets, been purchased by shell entities in jurisdictions not subject to equivalent restrictions, or arrived through other pathways not yet documented. Each of those scenarios would imply different legislative or regulatory remedies, ranging from stricter end-user verification requirements to expanded third-party reseller licensing conditions. Commerce Department officials have not publicly commented on how or whether they intend to investigate the specific acquisition pathway.

DeepSeek itself is a Hangzhou-based startup that generated significant market disruption in early 2025 when it released AI models that performed competitively with leading US offerings despite ostensibly operating under tight hardware constraints. That episode prompted a wave of reassessment in Washington about the relationship between raw compute access and AI capability, with some analysts arguing it demonstrated that software efficiency gains could partially offset hardware disadvantages. The latest allegations suggest that the efficiency framing may have been incomplete: if Blackwell chips were in use all along, the performance of those earlier models may owe more to hardware than was publicly understood.

Whether the Commerce Department opens a formal investigation, whether Congress moves to tighten third-party reseller regulations, and whether the administration's internal debate resolves toward tighter or looser controls on the H200 are the three immediate policy questions this episode puts on the table. Each carries direct consequences for Nvidia's addressable market, for the pace of Chinese AI development, and for the durability of the US government's claim to be managing the technology competition with China on terms it controls.

Written by Nick Ravenshade for NENC Media Group, original article and analysis.