On November 12, 2026, Jensen Huang ended his more than three-decade career at Nvidia by resigning. Four days earlier, he publicly acknowledged an extramarital affair. According to the latest newsletter from semiconductor research firm SemiAnalysis, another factor that drove the board to act was a vast AI chip smuggling network — an estimated more than 1.5 gigawatts (GW) of US export-controlled Nvidia hardware flowing into China, equivalent to at most roughly 1.5 million GPUs.
SemiAnalysis tracks global AI chip production and flows using its own "accelerator model." Its count excludes Nvidia's H20, H800 and other downgraded versions designed for compliance, and counts only genuinely restricted high-performance hardware. If all of it were H200 SXM (700W rated power per card), the total would be about 2.14 million units; if all were B200 SXM (1,000W per card), roughly 1.5 million; under actual mixed deployments, the number falls between the two. The US Department of Commerce has independently reached a conclusion of a similar order of magnitude.
xAI's Colossus One data center holds roughly 220,000 to 230,000 Hopper-series H100/H200 GPUs, and its planned Colossus 2 targets 1.2 million Blackwell B200/B300 chips. The total restricted compute that flowed into China could match or exceed the scale of xAI's Colossus cluster to date.
How Smuggling Works: From the Whitelist to Transit Routing
Taiwan's Keelung District Prosecutors Office indicted nine defendants on August 24, 2026, accusing them of illegally reselling Supermicro AI servers equipped with Nvidia B300 chips. The case began in February 2025, when server sales company Feihu Technology obtained qualification on Nvidia's authorized distributor whitelist and secured purchasing rights for 130 servers. After the original buyer defaulted, the hardware was resold to Chinese buyers and routed through Japan and Indonesia; 74 units entered China and 56 were seized before departure. Those involved included a delivery manager in Nvidia's Taiwan office, a sales manager at Supermicro's Taiwan branch, and the CEO of Supermicro distributor Albatron Technology. The 74 servers generated more than $21.2 million in illegal proceeds for the main operators.
SemiAnalysis's 1.5 GW estimate depicts a systemic, multi-channel macro-level flow spanning several years. The mechanism mirrors the Taiwan case: using legitimate distributor whitelist status to obtain supply, masking the final destination through third-country transit, and evading scrutiny with shell companies and forged documents. The export control loophole lies in the penetration of the authorization chain and in tracking mechanisms that lag behind actual flows.
Megaspeed: Where Intelligence and Business Intersect
Alice Huang (no relation to Jensen Huang), one of Huang's extramarital partners, founded Megaspeed International in 2023, a data center company registered in Singapore. Over the past two to three years, Megaspeed became one of Nvidia's largest overseas customers, purchasing a cumulative roughly $4.8 billion to $5 billion of Nvidia chips and related hardware. US intelligence agencies believe Alice Huang has ties to China's Ministry of State Security and have opened an investigation.
Starting in February 2025, Singapore authorities investigated whether Megaspeed colluded with US-based Supermicro to illegally transship servers carrying Nvidia chips to China. Alice Huang used industry connections and shell companies to obtain access to roughly 240,000 restricted chips in total. During this period, Shanghai-based company Shanghai Century Litong became an Nvidia priority partner.
According to a TechCrunch timeline, by late 2025 the Trump administration had received briefings on Alice Huang's identity but did not act immediately. Jensen Huang had previously successfully lobbied the Trump administration to ease some restrictions on chip sales to China.
Where Each Stakeholder Stands
The immediate impact on Nvidia is a leadership vacuum. CFO Colette Kress was appointed interim CEO, and Harvey C. Jones became the first chairman of the board. The board has launched an independent investigation. Huang agreed to testify before Congress before Thanksgiving. The company faces a period of uncertainty in its technology roadmap, customer relationships and internal governance, coming at a critical window as Blackwell architecture capacity ramps.
For China's AI industry, the 1.5 GW compute increment is equivalent to the core hardware required for one or more Colossus-class hyperscale data centers. This accumulation of compute spans several years, indicating that large-scale compute acquisition had already occurred before export controls could systematically block it.
The Taiwan indictment reveals that Nvidia's distributor whitelist system can in practice be bought, bypassed and subcontracted. Supermicro has once again been linked to a smuggling case, and its compliance system will face stricter scrutiny.
The Fundamental Predicament of Export Controls
The case exposes the core contradiction of AI chip export controls: the object of control is tangible hardware, while the circulation network is an intangible chain of commercial relationships. Physically tracking chips such as the H100, H200 and B200 is extremely difficult; they can be disassembled and repackaged to circulate in forms that are harder to identify. Imposing whitelist authorization on intermediaries essentially outsources compliance obligations to the node in the supply chain that is hardest to regulate.
The deeper predicament is that chip export controls are a supply-side tool, while demand-side incentives are strong. Reselling 74 Nvidia B300 servers yields more than $21.2 million, against a risk of up to five years in prison. Until the scarcity of AI compute changes materially, this arbitrage space will keep attracting new participants.
What May Happen Next
The core questions a congressional investigation will press are when the government learned of the matter, how much it knew, and why it did not act sooner. According to reports, the Trump administration had received relevant briefings by late 2025, after which Jensen Huang continued to push for some restrictions to be loosened.
The export control mechanism faces pressure for systemic reconstruction. If the 1.5 GW order of magnitude estimated by SemiAnalysis is confirmed by an independent Commerce Department investigation, it would mean the control framework has systematically failed over the past several years. Possible directions include introducing physical tracking requirements for AI chips or imposing stricter due diligence obligations on intermediaries.
Nvidia's trajectory will depend on the scope and conclusions of the independent investigation. If the investigation defines the problem as individual conduct rather than systemic compliance failure, the impact at the company level may be manageable; if it finds that senior executives knew and failed to disclose, the legal and regulatory costs will be more severe. Signals to watch include the timeline for disclosing the results of Nvidia's board investigation, the Commerce Department's official response to the SemiAnalysis estimate, and the content of Huang's testimony at the congressional hearing.
The semiconductor industry has long treated export controls as a business risk that can be "managed" by compliance teams rather than a security problem requiring redesign at the supply chain architecture level. The estimate of 1.5 million GPUs will make that framing hard to sustain.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接