OTWopentechwire
Tech Intelligence, Openly Wired
Policy

Blackwell GPUs Power China's Leading AI Models Despite US Export Bans

Forensic analysis of model weights reveals Chinese labs trained frontier systems on chips Washington intended to keep out of Beijing's hands, exposing gaps in export enforcement.

KW
Kenji Watanabe
Hardware & Products Reporter · Tokyo
Nov 16, 2026
5 min read
Blackwell GPUs Power China's Leading AI Models Despite US Export Bans
Credit: Sean Gallup / Getty Images

Technical Fingerprints Expose Hardware Origins

The architecture of large language models preserves traces of the infrastructure used to train them. Will DePue, a former researcher at OpenAI who now leads the nonprofit Unemployment Labs, examined the publicly released weights of models from DeepSeek, Moonshot AI, MiniMax and Ant Group. His analysis found evidence that several frontier systems, including Moonshot's Kimi K3 releases and DeepSeek's V4 series, display configurations consistent with training on Blackwell architecture GPUs, according to the research organisation.

Blackwell chips remain subject to export restrictions that prohibit their sale to China or Chinese-owned entities. The Trump administration permits shipments of the less capable H200 accelerator to proceed via specific routing arrangements that enable the US to collect a 25 per cent tariff.

DePue's methodology relies on identifying structural patterns in model weights that correspond to the tensor configurations and communication topologies of particular hardware platforms. Whilst these fingerprints do not constitute definitive proof of illegal activity, they indicate which chip families were present during training runs.

Offshore Data Centres and the Malaysia Pathway

The presence of Blackwell signatures in Chinese models does not necessarily mean export regulations were violated in a legal sense. One explanation centres on offshore compute facilities, which have become a significant channel for Chinese firms seeking access to restricted hardware without importing it directly into the mainland.

Malaysia has emerged as a key hub for this arrangement. Epoch AI, a nonprofit research organisation, estimates that approximately 275,000 Nvidia AI accelerators entered Malaysia during 2024 and 2025. The same group calculates that roughly two million advanced chips reached China over the same period through schemes designed to obscure final destinations.

White House technology adviser Michael Kratsios stated earlier this year that Moonshot had used export-banned Nvidia chips, citing intelligence sources. The claim aligns with DePue's technical findings but leaves open the question of whether the hardware sat in offshore racks or entered Chinese territory.

The offshore route satisfies the letter of current export rules, which focus on ownership and installation location rather than end users. Yet it undermines the stated policy objective: preventing Chinese AI laboratories from leveraging cutting-edge silicon to close the capability gap with US developers.

Calls to Tighten User-Based Controls

Policy analysts have long argued that location-based export restrictions leave an exploitable loophole. Sam Bresnick, a researcher at the Georgetown Centre for Security and Emerging Technology, questioned the logic of blocking chip sales to entities in mainland China whilst permitting the same organisations to rent compute capacity in neighbouring jurisdictions.

China hawks in Washington have pushed for user-based controls that would require tracking who runs workloads on restricted chips, regardless of where the hardware physically resides. Implementation would demand co-operation from cloud providers, data centre operators and chip manufacturers, raising both technical and diplomatic challenges.

Current US regulations place limited obligations on companies to monitor what customers do with controlled goods once sold. This gap becomes more acute when transactions involve intermediaries or offshore facilities that operate in regulatory grey zones.

Smuggling Networks and Corporate Liability

Megaspeed, a company led by Alice Huang, has attracted scrutiny from both US and Singaporean authorities for its role in chip distribution. The firm operates through its subsidiary Speedmatrix, which has been implicated in schemes to move restricted hardware into China. Investigations remain ongoing.

A recent development added personal dimensions to the corporate compliance questions. Jensen Huang, who departed Nvidia this week, admitted to a relationship with Alice Huang. The connection has intensified focus on whether company executives were aware of potential export violations involving Megaspeed's operations.

Nvidia and its former chief executive had previously vouched for Megaspeed's data centres, stating that company officials visited the facilities to verify their legitimacy. If government investigators establish that smuggling occurred and that senior leadership had knowledge of it, Nvidia could face substantial penalties beyond reputational damage.

Short-selling research firm Culper published analysis in May arguing that Nvidia faced a "China problem" stemming from dependence on Chinese demand, whether through remote data centres or illicit shipments. The report detailed Megaspeed's ties to Nvidia and alleged that Alice Huang's companies received backing from Chinese entities barred from accessing US chips.

Beijing's Domestic Chip Push and Stockpiled Hardware

Culper's thesis did not hinge on US enforcement action. Instead, the firm bet that China's response to export restrictions - a renewed commitment to domestic semiconductor production - would erode demand for Nvidia products if Beijing's fabrication efforts succeed.

An estimated 600,000 H200 chips currently sit in warehouses in Hong Kong, awaiting clearance to enter mainland China. The reason for the hold-up remains unclear. Graham Webster, a research scholar at Stanford who focuses on US-China technology policy, dismissed the theory that Beijing is deliberately restricting imports to spur domestic manufacturing, noting that demand far exceeds local supply.

Alternative explanations include preparations for a trade negotiation that might involve large-scale US goods purchases, or concerns about embedded surveillance capabilities in the chips themselves. The stockpile represents a tangible measure of unmet demand and the friction inherent in current trade arrangements.

Regulatory Scrutiny Turns Inward

The Commerce Department, which administers chip export licences, faces questions about its own oversight. Secretary of Commerce Howard Lutnick reportedly confronts pressure following reports that investigators failed to identify suspicious buyers and maintained overly close relationships with Nvidia officials.

The episode illustrates the difficulty of enforcing technical restrictions in a global supply chain where chips, data centres and compute workloads can be separated across jurisdictions. Forensic analysis of model weights offers a new tool for assessing compliance, but it arrives after training runs are complete and models are already deployed.

For policymakers, the central tension persists: export controls aim to preserve a technological edge, yet the architecture of cloud computing and the fungibility of offshore data centres allow determined actors to route around geographic restrictions. Closing those pathways would require either stricter end-user verification or acceptance that some leakage is inevitable in a market where demand for advanced compute spans borders.

Read next
Policy

Researcher Publishes Windows Privilege Escalation Zero-Day After Disclosure Dispute

Marcus Halloran · 4 min
Policy

Export Control Allegations Turn Scrutiny to Southeast Asian Shell Companies

Mei-Lin Tan · 6 min
Policy

AI Hardware Suppliers Brace for Tighter Export Controls After Nvidia Turmoil

Kenji Watanabe · 5 min
Spot something wrong? Email corrections@opentechwire.com. We log every correction publicly.