China smuggling of Nvidia GPUs may explain Jensen Huang’s departure

By Grace Kim

More than 1.5 gigawatts of export-restricted Nvidia AI accelerators may have reached China in recent years, according to SemiAnalysis’s Accelerator Model, which tracks accelerator production and destinations.

The research firm presented the estimate in its latest newsletter. It excludes processors Nvidia developed for export compliance, including H20 and H800 chips. SemiAnalysis did not disclose the GPU count or split shipments by product generation. The firm also claims the U.S. Department of Commerce reached a similar conclusion about the scale of unauthorized shipments.

SemiAnalysis’s prescription is specific: “Restrictions should focus on the amount and capabilities of hardware available to Chinese entities rather than completely prohibit any access.”

How much computing power does 1.5 GW represent?

The research firm does not explain how it converts accelerator quantities into power. Its likely method multiplies package-level nameplate power by GPU count, although that calculation remains an assumption.

If all hardware consisted of Nvidia H200 SXM processors rated at up to 700 watts, the estimate implies about 2.142 million units. If all hardware consisted of B200 SXM processors rated at up to 1,000 watts, it implies 1.5 million units. A mix of H200 and B200 would fall between those figures.

xAI’s Colossus One had 220,000 to 230,000 Hopper H100 and H200 GPUs. Colossus 2 targets 1.2 million Blackwell B200 and B300 GPUs. SemiAnalysis compares the alleged China inventory with xAI’s entire existing Colossus operation. It says the diverted capacity is comparable or larger.

What do the smuggling cases show?

Both cases remain before the courts.

Prosecutors charged Supermicro co-founder Wally Liaw, Taiwan sales executive Steven Chang and contractor Willy Sun in March 2026. The indictment alleges that the defendants sold about $2.5 billion of servers to a company in Southeast Asia. It says $510 million of that equipment was diverted to China and obtained by Alibaba.

Prosecutors say the defendants used substitute servers and transferred identification labels to conceal actual destinations during audits.

Authorities also accused nine people connected to Nvidia, Albatron, Chief Telecom, Flying Tiger and Supermicro in Taiwan. Prosecutors allege the group arranged unauthorized exports of 130 Nvidia B300-based servers. Seventy-four of those servers eventually reached China.

The suspects used shell companies, forged documents and declared Japan as the destination. Taiwanese authorities discovered the operation during a drug-smuggling investigation.

Why might Huang’s departure matter now?

One separate report alleged that Alice Huang, identified as an alleged Chinese intelligence asset and a Megaspeed executive, facilitated smuggling of $7.5 billion of Nvidia AI hardware into China. It placed the total at about 240,000 GPUs, mostly H100 and H200 models.

The claim emerged amid reports of Jensen Huang’s extramarital affair with Alice Huang. SemiAnalysis cited no direct evidence that intelligence agencies influenced Nvidia’s board or that Huang participated in unauthorized exports.

Analysts speculate that some members of the U.S. intelligence community sought Huang’s removal because he opposed aggressive restrictions on China’s access to American AI hardware. SemiAnalysis suggests the scale of violations made Huang’s personal conduct a major issue at that moment.

Their argument remains circumstantial. The analysts believe Huang’s departure removes an influential opponent of stricter controls. They also expect it to accelerate limits on Chinese cloud access to American AI hardware.

Would tighter controls accelerate Chinese chip development?

The U.S. House passed the Remote Access Security Act (RASA) on January 12. The bill would extend export-control authority to remote access to advanced AI hardware hosted outside China, including data centers in Europe and the Middle East. SemiAnalysis reports that Nvidia and Oracle actively oppose the proposal.

The research firm favors Chinese developers renting capacity from trusted providers outside China under U.S. oversight.

Continued access supports revenue for Nvidia, AMD and cloud service providers. It lets U.S. authorities monitor Chinese tenants in foreign data centers and terminate access for national-security reasons. It also reduces Chinese developers’ incentives to adopt domestic accelerators that could later compete with Nvidia and AMD abroad.

Chinese laboratories have developed the open-source tools DeepEP, DeepGEMM and MSA. The tools improve AI workload efficiency on Nvidia hardware and benefit developers worldwide. American companies also depend on Chinese engineers, component suppliers and manufacturers.

Broad restrictions could reduce open-source optimizations, weaken international cooperation and accelerate software tailored to domestic chips.

Access to AMD or Nvidia accelerators also helps Chinese laboratories build advanced models that may rival leading U.S. systems. SemiAnalysis acknowledges national-security risks, including potential use in weapons of mass destruction. It supports limits, not a complete cutoff for Chinese developers.

Huang has consistently warned that broad curbs would accelerate domestic accelerator adoption. SemiAnalysis also warns that an all-inclusive ban could weaken American high-tech companies and international software collaboration.

Under the analysts’ proposal, trusted cloud capacity, rather than physical export totals, will shape access. The next commercial test is whether tighter remote-access controls redirect demand toward Chinese accelerators or reduce business for Nvidia, AMD and cloud service providers.