The Chip Never Crossed the Border. The Compute Did Anyway.
Washington's rules stop a Nvidia chip from shipping to China. They say nothing about a Chinese company renting time on that same chip from a data center next door — and for the better part of a year, that's reportedly exactly what's been happening.
Trace a compute request from Southeast Asia to China
Simplified model of the mechanism, not a live tracker: node positions and the "request" animation are illustrative. Path 1 (chip shipment) is always licensed and legal because the destination isn't China. Path 2 (remote login) is the loophole — toggle the regime, then send a request to see whether it completes.
The plain version
Imagine a customs agent stationed at the factory door, checking every truck that leaves with a shipment of advanced Nvidia AI chips, making sure none of them heads for China. That part of the system works: it's illegal to ship the physical chip there, and the rule is enforced.
But here's the workaround. Instead of shipping the chip to China, you ship it to a data center in Malaysia, Thailand, or Japan — completely legal, since the destination isn't China. Then a company in Beijing logs in over the internet and rents computing time on it, the same way anyone rents a server from Amazon or Google. No chip ever crosses a border. No truck for the customs agent to check.
Reportedly, that's exactly what's been happening. Chinese AI labs — including ByteDance, Alibaba, Tencent, and the startup Moonshot AI — have used rented compute in Southeast Asian and Japanese data centers to train frontier AI models on America's most advanced chips, without a single restricted GPU ever touching Chinese soil.
Washington built its rulebook around the wrong verb. It restricts shipping, when the thing that actually matters is using. A chip sitting in a rack in Johor does the exact same AI training work whether the login comes from San Francisco or Shenzhen — geography stopped being the gatekeeper the moment the cloud got involved.
There's a fix waiting in Congress. A bill called the Remote Access Security Act would legally treat "renting the chip remotely" the same as "shipping the chip" — requiring the same export license either way. It passed the House back in January. Half a year later, it's still stuck in the Senate, while the workaround keeps running.
The expert version
US export controls on advanced AI accelerators — Nvidia's A100/H100/H200/B200-class GPUs, gated by processing-performance and interconnect-bandwidth thresholds under the Export Administration Regulations — are administered by the Commerce Department's Bureau of Industry and Security through an End-Use/End-User and Entity List framework. The controls trigger on an "export," "reexport," or "transfer (in-country)" of a controlled item: legal terms of art built around the physical or technical transfer of a commodity, source code, or technology. The framework assumes the compute and the chip travel together — that using an H100 requires physical custody of one.
Cloud infrastructure breaks that assumption. A hyperscale facility in Johor, Bangkok, or Tokyo can lease GPU-hours over the public internet to a customer anywhere on Earth, and the silicon itself never leaves the rack. BIS has historically applied "deemed export" treatment mainly to narrower technology-transfer scenarios, such as handing controlled source code to a foreign national — not to pure compute-as-a-service. That gray zone widened further after Commerce rescinded the Biden-era AI Diffusion Rule and its cloud-provider know-your-customer due-diligence requirements in May 2025.
Reporting from CNBC and Asia Times, corroborated across multiple outlets, describes Chinese labs — ByteDance, Alibaba, Tencent, and Moonshot AI among them — leasing or installing compute capacity inside Southeast Asian and Japanese facilities, running training workloads there, then deploying the resulting model weights domestically. Only the trained model — software, not a controlled commodity — crosses back into China, and model weights sit outside the commodity-licensing regime entirely.
BIS is now compiling two separate watch lists — one for physical transshipment routes, one for remote-access arrangements — and drafting a rule to close the gap Commerce itself widened, per Tech Times. Legislatively, the Remote Access Security Act would amend the Export Control Reform Act to define remote or cloud access to a controlled item as an export event in its own right, subjecting GPU rental to the same licensing regime as physical shipment. It passed the House on a bipartisan vote in January 2026 and remains unpassed in the Senate, with the FY2027 NDAA floated as a possible legislative vehicle.
Why it matters for tech + supply chain: chip supply used to be the whole story in the AI race. Now the real chokepoint is a legal definition — and whoever exploits it fastest trains competitive models regardless of where the silicon physically sits.
Why it matters for tech + supply chain: it turns export compliance from a shipping-manifest problem into a jurisdiction-and-billing problem — cloud providers, not just chip makers, become the enforcement point, and "export" now has to be defined in software terms, not customs terms.