Ten days ago, federal prosecutors unsealed an indictment charging Supermicro co-founder Yih-Shyan "Wally" Liaw with conspiring to smuggle $2.5 billion in Nvidia GPUs to China. Dummy servers in a Southeast Asian warehouse. Serial number stickers peeled off with a hair dryer. Compliance auditors kept off-site while the real hardware crossed borders. The stock cratered 33%.
That was the crime. Now here is the proof that the product actually arrived.
Reuters reported this week that four Chinese universities, including two with direct ties to the People's Liberation Army, procured Super Micro servers loaded with restricted Nvidia A100 GPUs. The purchases were completed in 2025 and 2026. The chips have been banned for export to China since October 2022.
The two confirmed buyers are Beihang University in Beijing and the Harbin Institute of Technology in northeastern China. Both belong to a group known as the "Seven Sons of National Defense," seven elite universities affiliated with China's Ministry of Industry and Information Technology that devote at least half of their research budgets to military products. Three quarters of university graduates recruited by China's defense state-owned enterprises come from these schools. Hundreds of their scientists sit on PLA expert advisory committees, contributing to fighter jet and aircraft carrier programs.
Both Beihang and Harbin are on the U.S. Entity List, which makes it functionally impossible for American companies to secure export licenses to supply them. And yet.
Beihang University posted a March 16 procurement notice for a machine-learning workstation built on a Super Micro system configured with four Nvidia A100 GPUs. Harbin Institute of Technology posted a similar notice in July 2025 for a Super Micro system with eight A100s. Two additional universities, at least one with military connections, sought to make similar purchases through publicly available tender documents, though Reuters could not confirm whether those deals closed.
Nobody knows how the servers were sourced. Super Micro declined to comment on the procurement documents. Nvidia offered the standard line about working closely with customers and the U.S. government on compliance. The gap between that statement and the reality of A100s sitting inside PLA-linked research labs is about as wide as a gap gets.
Here is where the cooling industry should pay attention.
Each Nvidia A100 in SXM form factor runs at a thermal design power of 400 watts. The PCIe variants hit 250 to 300 watts. An eight-GPU system like the one Harbin purchased generates upward of 3.2 kilowatts from GPUs alone before you account for CPUs, memory, storage, networking, and power supply losses. A dense rack of these systems pushes past 30 kilowatts.
That heat has to go somewhere. And when the hardware operating beneath it exists outside any legitimate procurement channel, there is no OEM spec sheet guiding the thermal solution. No pre-qualified cold plate from CoolIT or Vertiv. No validated CDU from Motivair. The facility running these machines is improvising, and improvised thermal management at 400 watts per GPU is a recipe for throttled performance, premature failure, or both.
This is the part that procurement documents cannot capture. We wrote when the Liaw indictment dropped that a shadow market for AI compute generates unmeasured cooling demand. The university purchases confirm that demand is real and distributed across institutions with the resources and motivation to run these systems hard.
Supermicro has been here before. Between 2001 and 2003, the company illegally exported servers to Iran through a distributor in Dubai. That scheme produced roughly $455,000 in combined fines. The playbook is identical: third-party routing through a neighboring country, false documentation about end users, repeat purchases to normalize the relationship. Twenty years later, the intermediary is in Southeast Asia instead of the Gulf, and the dollar figure has multiplied by roughly five thousand.
The DOJ alleges that the unnamed Southeast Asian company became Supermicro's 11th largest customer globally, generating $99.7 million in fiscal 2024 revenue alone. Prosecutors say the group used false documents and convoluted transshipment schemes, staging dummy servers at warehouses that were supposed to hold the real inventory. Surveillance cameras filmed co-conspirator Ting-Wei "Willy" Sun unboxing fake servers and using a hair dryer to transfer serial number stickers between boxes. A Supermicro auditor was then sent photographs of the staged warehouse to sign off on compliance.
That is not a rogue employee. That is infrastructure built for scale.
Senators Jim Banks and Elizabeth Warren sent a bipartisan letter to Commerce Secretary Howard Lutnick urging the Bureau of Industry and Security to immediately pause and review all active export licenses covering advanced Nvidia chips and server systems destined for China and Southeast Asian intermediaries in Malaysia, Thailand, Vietnam, and Singapore. They explicitly challenged Jensen Huang's prior claims that chip diversion was not occurring, calling those statements contradicted by available reporting.
If BIS acts on that request, the downstream effects extend well beyond semiconductors. Export license freezes would slow server shipments. Slower server shipments mean delayed data center buildouts. Delayed buildouts mean delayed cooling infrastructure orders. Every CDU, every rear-door heat exchanger, every immersion tank scheduled against a capacity timeline tied to GPU delivery gets pushed right.
And on the other side of the controls, Chinese institutions that cannot source compliant hardware will continue finding it through channels nobody is measuring. The thermal load does not disappear because the procurement was illegal. Those GPUs still draw 400 watts. That heat still needs a path out of the chassis, out of the rack, and out of the building.
The Supermicro saga is now a two-part story. Part one was the indictment: $2.5 billion in smuggled hardware, a co-founder arrested, a stock in freefall. Part two is the proof of delivery: restricted GPUs confirmed inside blacklisted military research institutions, sourced through channels that remain completely opaque.
Cooling vendors selling into the legitimate AI infrastructure supply chain should be thinking about three things. First, the regulatory perimeter is expanding. Enforcement agencies will eventually look beyond controlled chips to items specially designed for their deployment, and that includes thermal management hardware integrated into GPU server platforms. Second, Supermicro holds roughly 10% of the global AI server market. If sanctions, debarment, or reputational damage shrinks that share, competitors like Dell, HPE, and Lenovo absorb the volume, and every one of those transitions reshuffles cooling vendor partnerships. Third, the shadow demand is real. If four universities felt comfortable posting public tender documents for banned hardware, the private procurement activity is orders of magnitude larger.
The A100 is not even current generation. Nvidia has shipped the H100, H200, and B200 since. Every successor draws more power and generates more heat. Whatever cooling gap exists in the shadow supply chain today will only widen as the smuggled hardware gets newer.