Key Takeaways
- Nvidia’s dealings with Megaspeed International have renewed concern about intermediaries that purchase advanced AI processors outside China.
- Physical diversion and cross-border cloud rentals can provide access to restricted computing capacity without a direct sale into China.
- Voluntary hardware monitoring and the withdrawal of the 2025 AI Diffusion Rule leave gaps for regulators, chipmakers, distributors, and cloud providers.
Nvidia’s relationship with cloud computing company Megaspeed International has pushed a familiar export-control problem back into view: restricting a direct shipment does not necessarily restrict access to the computing power behind it.
Megaspeed International was an Nvidia cloud partner and purchased billions of dollars’ worth of advanced computing equipment. The controversy surrounding those purchases has increased scrutiny of how high-end processors move through intermediaries, where the resulting servers operate, and who can remotely use them. It does not, by itself, establish wrongdoing by Nvidia or Megaspeed International.
Still, the scale of the broader market is difficult to ignore. Reuters reported IDC estimates showing that Nvidia accounted for about 55% of China’s AI-accelerator shipments in 2025, representing approximately 2.2 million of four million units. That level of availability suggests U.S. restrictions have constrained formal sales channels without eliminating demand or access.
An AI chip can cross a border, but its computing capacity does not have to. A customer in China might rent GPU resources hosted in another country, submit workloads remotely, and receive results without importing the underlying hardware. Such arrangements complicate the distinction between exporting a physical product and delivering a controlled capability as a service.
Physical diversion remains another concern. Distributors, resellers and shell companies can obscure the ultimate destination or beneficial user of restricted processors. U.S. authorities have identified hundreds of millions of dollars in Nvidia-chip smuggling to China, indicating that enforcement weaknesses can emerge several steps downstream from the original manufacturer.
Commercial data provides another warning signal, although not proof of diversion. Nvidia’s China revenue declined from $7.1 billion in fiscal 2021 to $5.8 billion in fiscal 2022. Its Singapore revenue, meanwhile, increased from $2.3 billion in 2022 to $23.7 billion in 2024. Singapore is a major regional business and billing center, so the figures should not be treated as evidence that chips were improperly redirected. They do, however, show why regulators are examining purchasing entities, deployment locations and end users separately.
The monitoring tools available today only partially answer those questions. According to the Atlantic Council, Nvidia’s “fleet intelligence” location-monitoring system is voluntary and limited to Nvidia Cloud Partners. Hardware sold through other channels may therefore sit outside that visibility. Even participating operators can present a challenge if corporate ownership, customer access or equipment location changes after purchase.
What would more effective oversight look like? Regulators could require stronger beneficial-ownership checks, ongoing end-use verification and records covering remote access to large GPU clusters. Chipmakers and distributors could also apply risk-based reviews when purchasing patterns, financing structures or deployment plans appear inconsistent with a customer’s established operations.
Policy has moved in the opposite direction in at least one important respect. The United States rescinded the 2025 AI Diffusion Rule, which had been designed partly to address third-country diversion and remote access through locations including Singapore and Malaysia. The framework was controversial because broad geographic controls could burden legitimate cloud investment and allied markets. Yet withdrawing it reopened the question of how the Commerce Department should regulate compute that is purchased in one jurisdiction and consumed from another.
The CBC has outlined how rerouting, intermediaries and remote computing can help Chinese users obtain access despite direct sales restrictions. At the same time, controls have encouraged China to strengthen domestic alternatives, including Huawei’s accelerator ecosystem. Tighter restrictions may slow access to particular Nvidia products while also supporting a larger market for Chinese hardware and software.
For enterprise buyers, this is becoming a governance issue rather than a narrow customs matter. Procurement teams may need to document where GPU systems are installed, who administers them, which customers can rent capacity, and whether workloads originate in restricted jurisdictions. Cloud providers face especially difficult questions because infrastructure location and user location can differ.
The Nvidia and Megaspeed International controversy exposes the central weakness in the current approach. The U.S. Export Administration Regulations can restrict products and transactions, but advanced AI compute travels through supply chains, corporate structures and cloud connections. Closing one route may redirect demand into another. The next phase of policy will likely depend less on a simple country-of-sale test and more on tracing ownership, deployment and actual access over the hardware’s operating life.
⬇️