Malaysia Will Track Data Center Power Use to Catch Smuggled AI GPUs

Malaysia has approved 117 data center projects, 55 of them already operational, and its government now says some of that capacity is being used as a front: GPUs imported legally into Malaysian facilities, then shipped abroad in violation of U.S. export controls. Investment, Trade and Industry Minister Datuk Seri Johari Abdul Ghani confirmed the scheme on Thursday and unveiled draft legislation to stop it, reports Malay Mail.

The proposed law has two prongs. First, operators must disclose how many GPUs they import, their facilities' power consumption, physical locations, and the identities of both operators and hyperscale customers. Second, data centers will pay for every megawatt of capacity they reserve, whether or not they use it. A facility that secures 100 MW but draws 50 MW pays for the idle 50 MW too.

Why tie GPUs to electricity?

The logic is straightforward. If a data center imports thousands of accelerators but its power draw tells a different story, the discrepancy flags possible diversion. Malaysia wants AI hardware brought into the country to stay and work there, Johari said, and the Digital Ministry is now reviewing existing legislation to see whether local authorities can act against entities that re-export processors covered by U.S. bans.

The idle-capacity charge serves double duty. It discourages companies from hoarding grid capacity they won't need for another one to three years — a real constraint as cloud providers and AI infrastructure firms pile into Southeast Asia. And it raises the cost of pretending hardware runs in Malaysia when it has already left for China.

Johari acknowledged the smuggling problem plainly: some companies had rented space in Malaysian data centers specifically to import GPUs legally, then shipped them abroad. Malaysia's proliferation of AI facilities makes it an attractive transit point. It is easier, as the saying goes, to hide a needle in a pile of needles than in a haystack.

Can power meters actually count GPUs?

Not reliably, and that is the law's central weakness. Nameplate power ratings vary wildly across accelerator generations. Nvidia's H100 and H200 SXM GPUs are rated at 700W each; the B200 SXM carries a 1,000W rating. None of them necessarily consumes its rated figure under real workloads.

Without knowing server configurations, the draw of CPUs, memory, storage and networking gear, and a facility's power usage effectiveness (PUE), any estimate of deployed GPU counts from power data alone would be imprecise at best. There is also a timing problem: a considerable gap can separate a GPU's arrival from its actual deployment, so low consumption early in a project proves nothing.

Reliable verification would require physical inventory audits or mechanisms that track individual processors. Johari did not specify whether the law will mandate disclosure of GPU models and serial numbers, or whether authorities will conduct on-site inspections — the two tools that would close the loophole power data leaves open.

What happens next?

Malaysia is prepared to cooperate with U.S. counterparts and will provide information about facilities housing particular GPUs on request, Johari said. The Digital Ministry's legislative review is ongoing, and until it produces draft text with audit powers or serial-number reporting attached, power consumption remains a blunt instrument — one that raises smugglers' costs through idle-capacity charges, but cannot yet prove what sits inside the racks.