Key Takeaways

  • Chinese artificial intelligence labs train advanced models by renting American graphics processors hosted in third-party hubs like Johor, Malaysia, completely bypassing physical export restrictions.
  • Megaspeed, a Singapore-based neocloud provider, built its core business model around supplying high-end Nvidia computing capacity to international clients without moving physical hardware across Chinese borders.
  • The Commerce Department's export control enforcement team has begun mapping global remote access patterns, despite cloud rentals remaining entirely legal under current United States statutes.
  • The Biden administration drafted an "AI diffusion rule" targeting overseas compute access scheduled for May 2025, but the Trump administration chose not to enforce the measure.
  • Restricting cross-border cloud rentals would threaten revenue streams for American tech giants, including Nvidia and cloud providers like Oracle that count Chinese teams among their buyers.

Johor Powers Chinese Artificial Intelligence

Washington spent years blocking advanced Nvidia accelerators from crossing the Pacific directly into Shenzhen and Beijing. Yet Chinese engineers still run workloads on top-tier American silicon every day. They simply do not buy the boards. They rent the rack space.

As Mackenzie Hawkins explains, “In some ways the backbone of Chinese compute is like Johor, Malaysia.” (play from 7:09)

Johor sits right across the border from Singapore. Data center developers poured billions into the state to build massive server facilities. Inside those halls sit clusters of high-end processors. A Chinese engineering team in Shanghai connects via remote login, runs training jobs overnight, and downloads the model weights. The physical chip never touches mainland soil.

“In the vast majority of cases, this is legal,” (play from 32:09) Hawkins says. “This is what Megaspeed's business is. This is what a lot of Southeast Asian neoclouds do.” (play from 32:12)

Megaspeed established itself in Singapore to operate fleets of Nvidia processors. When regulators scrutinize intermediary companies, they look for illicit hardware diversion: physical smuggling of boxed cards across borders. Renting API access or remote server shells violates no current American export statute. The entire setup is open, contract-backed commerce.

The Commerce Department Starts Mapping Cloud Routes

Inside Washington, national security officials recognize the gap between physical enforcement and digital reality. If the goal of export controls is slowing down rival military and commercial models, letting labs rent equivalent capacity in Southeast Asia defeats the purpose.

Hawkins points out an unusual internal shift: “The enforcement unit of the part of the Commerce Department that handles chip export control, so the people who are like, 'Oh, you've violated, like you smuggled, we're going to go after you,' that unit started mapping where in the world remote access happens, which is very interesting, because remote access is not illegal.” (play from 34:53)

Tracking lawful server traffic signals an appetite to redefine what an export actually means. The Biden administration tried to formalize this before leaving office. It created a package known as the AI diffusion rule, designed to restrict remote compute access globally starting in May 2025. Yet the policy stalled when the Trump administration took office and declined to implement the enforcement mechanism.

Why Big Tech Wants the Loophole Open

Hardware makers and cloud vendors have lobbied aggressively against closing remote access. Banning overseas server rentals does not just stop Chinese developers from training models. It destroys revenue for American suppliers.

Hawkins notes the scale of the commercial blow: “This is sort of nightmare scenario for Nvidia if remote access were to be restricted. It's also a nightmare scenario for some hyperscalers like Oracle, that counts Chinese offtakers as a significant portion of its business.” (play from 35:42)

If Washington bans foreign rentals, Chinese buyers will not stop building software. They will move workloads to domestic alternatives like Huawei's Ascend processors. Huawei already sells accelerators aggressively across the Global South. Cutting off cloud access forces Chinese software developers to write directly for non-American chip architectures, permanently severing Silicon Valley's developer lock-in.

What to Do With This

Audit your cloud infrastructure dependencies before your next compute contract renewal. If your product relies on specialized foreign neoclouds or third-party GPU aggregators, identify which data centers house your physical instances. Shift mission-critical training pipelines to providers with clear compliance chains in jurisdictions with stable bilateral trade agreements, avoiding regional hubs currently under federal surveillance.