Chinese artificial intelligence developers are actively circumventing US export controls by routing workloads through high-end Nvidia graphics processors stationed across Southeast Asia.
Washington bans direct shipments of Nvidia’s flagship Blackwell and Rubin chips to China, citing military risks. The restrictions were designed to cripple Beijing’s access to advanced silicon. They did not work. Chinese startups—spearheaded by Moonshot AI, Z.ai, and DeepSeek—continue deploying advanced Western compute to train frontier models. The loophole is structural. US export rules govern where physical hardware lands, not who accesses the compute remotely across a digital network.
Nvidia sits at the nexus of the confrontation. The Silicon Valley chipmaker commands a $4.5 trillion market capitalization, running neck-and-neck with Apple as the world’s most valuable enterprise. Scrutiny on its distribution network spiked after Chief Executive Jensen Huang admitted to an affair with an executive at a Singapore-based distributor under US federal scrutiny for smuggling AI microchips into mainland China. Nvidia maintained it found no evidence of illicit diversion. Yet the technical pipelines linking Chinese software engineers to banned American silicon remain wide open.
A Chinese engineer does not need physical custody of a GPU to train a neural network. Server racks located in Malaysia or Thailand execute code just as quickly over high-speed fibre.
"US export controls restrict where a chip is shipped, not who uses it," said Stephen Wu, founder of investment firm Carthage Capital. An array of Nvidia processors can sit in Malaysia’s Johor region in strict legal compliance with trade authorities, Wu noted, while a client in Shenzhen logs in to run intensive training clusters thousands of miles away. White House adviser Michael Kratsios confirmed that Moonshot AI tapped high-tier Nvidia systems deployed inside Thailand. Smuggling networks compound the compliance breakdown. DeepSeek assembled its breakout model using thousands of processors dismantled overseas and reassembled inside Chinese borders, according to industry disclosures.
Washington attempted concessions earlier this year. Trade officials cleared Nvidia to sell downgraded H200 accelerators to Chinese customers, subject to strict licensing conditions and a mandatory 25 percent US government cut on sales. Beijing pushed back. Chinese state planners balked at the H200’s restricted specifications and heavy tariffs, urging state-backed institutions to shift demand toward homegrown hardware.
China's domestic semiconductor ecosystem is absorbing the pressure through sheer scale. Domestic chips process slower and demand significantly more electrical energy, but state subsidies neutralize the cost differential.
"Nvidia chips still make top-level AI work faster and easier, but they are no longer the only path," said Poe Zhao, founder of Hello China Tech. Zhao noted that while training on indigenous chips requires deeper engineering intervention, Beijing's energy subsidies allow domestic labs to deploy clusters of lower-efficiency processors in massive quantities. Huawei leads the domestic push, with DeepSeek preparing to train its upcoming architecture directly on Huawei accelerators. Huawei moved its next-generation AI processor launch forward by nine months, scheduling commercial readiness for the first quarter of 2027.
Federal enforcement agencies in Washington are drafting revised cloud-access rules to shut down third-party data centre leasing in Southeast Asia before Huawei's accelerated 2027 deployment window opens.