The GPU Black Market that Washington Can't Shut Down

The U.S. banned AI chip exports. Billions still reached China — moved by shell companies, Southeast Asian middlemen, and operators gaming the system.

Supermicro smuggled servers seized by BIS inspectors. Source: U.S. Department of Justice. Illustration edited with Google Gemini.

By Bernie Margulies·Apr 20, 2026

We first heard about it the way everyone hears about it — over drinks, from someone who runs a neocloud. We'll call him James. He was describing his company's GPU allocation from NVIDIA, the deals they were closing, the facilities they were building. And then, almost as an aside, as though he were telling us where he parks at the airport: "Most of these chips end up in China." He said it the way you'd say the sky is blue. Not a confession. Not a whisper. Just the way things are.

James was not the last person to tell us this. Over the course of several months, we spoke with multiple neocloud operators, GPU brokers, data center developers, and distributors across the US and Asia. The conversations had a strange uniformity. Everyone knew. Everyone had a version of the same story. Nobody seemed particularly worried about telling it.

There are warehouses in Hong Kong full of Western and Chinese electronics. These are places where you can buy a knockoff iPhone charger, a custom circuit board, or, if you know whom to ask, an export-controlled NVIDIA H200 GPU for approximately 213,000 HKD. That's roughly $29,700 — near the US list price — a figure that tells you something important: supply is so abundant that smuggling hasn't even inflated the market. You don't even need to go there personally. Search eBay for H100 or H200 GPUs and most sellers ship from China. The H100 has been export-banned to China since October 2022; the H200 ban was lifted only in December 2025. Neither timeline explains how either chip is plentiful in Shenzhen.

A Gamers Nexus investigation in 2025 spent roughly three weeks filming across Hong Kong, Shenzhen, and Taipei, and one of their sources — a middleman identified only as "Mr. 5" — put the situation plainly. "Of course they know," he said of the vendors, the brokers, the entire process. "But 'open one eye, close one eye.' They know you can buy it." Asked whether his company would stop if NVIDIA wanted it to, his answer was simple: "No." The Chinese market, as he and others in the industry describe it, is too big and too profitable to walk away from.^{[1]Gamers Nexus, "The Nvidia AI GPU Black Market: Investigating Smuggling, Corruption, and Governments" (2025)https://gamersnexus.net/gpus-deep-dive-news/nvidia-ai-gpu-black-market-investigating-smuggling-corruption-governments}

The United States treats these chips as instruments of national security concern demanding strict export controls. Since October 2022, the Bureau of Industry and Security has issued four rounds of restrictions, each one tighter than the last.^{[2]Bureau of Industry and Security, Interim Final Rule on Advanced Computing and Semiconductor Manufacturing Items (October 7, 2022)https://www.federalregister.gov/documents/2022/10/13/2022-21658/implementation-of-additional-export-controls-certain-advanced-computing-and-semiconductor} The policy goal is straightforward: slow China's ability to train frontier AI models and build AI-enabled weapons systems. But the enforcement has been half-baked.

In March 2026, federal prosecutors arrested Supermicro co-founder Yih-Shyan "Wally" Liaw for allegedly directing the diversion of $2.5 billion in NVIDIA servers to China.^{[3]CNN, "Co-founder of tech company charged with diverting $2.5 billion in Nvidia AI chips to China" (March 2026)https://www.cnn.com/2026/03/19/politics/super-micro-computer-founder-charged-ai-chips-china} Multiple documented networks have moved transactions in the hundreds of millions of dollars, operating for months at a time.^{[4]CNAS, "Countering AI Chip Smuggling Has Become a National Security Priority" (2025)https://www.cnas.org/publications/reports/countering-ai-chip-smuggling-has-become-a-national-security-priority} ChinaTalk estimates China now has access to approximately 2.8 million H100-equivalents through a combination of legal purchases, smuggling, domestic production, and remote access to overseas data centers.^{[5]ChinaTalk, "How Many Chips Does China Have?" estimate of 2.8 million H100-equivalents (2025)https://www.chinatalk.media/p/how-many-chips-does-china-have}

Everyone in the industry knows. The Chinese idiom heard most often — 睁一只眼, 闭一只眼 — translates as "open one eye, close one eye," similar to turning a blind eye.^{[1]Gamers Nexus, "The Nvidia AI GPU Black Market: Investigating Smuggling, Corruption, and Governments" (2025)https://gamersnexus.net/gpus-deep-dive-news/nvidia-ai-gpu-black-market-investigating-smuggling-corruption-governments}

Why these chips matter

Training a frontier AI model is not a single calculation. It is thousands of GPU chips running in parallel for weeks or months, burning through electricity at a rate that would power a small town, communicating at speeds measured in terabytes per second.

The chips that make this possible — NVIDIA's H100, H200, and B200 — are manufactured exclusively by TSMC in Taiwan using extreme ultraviolet lithography machines made by ASML in the Netherlands. No other company on earth can produce chips at this level. The entire global AI supply chain runs through these three companies in three countries.

That concentration creates a chokepoint, and Washington has used it. The US restricts what TSMC can fabricate for Chinese customers. It restricts what ASML can sell to Chinese fabs. It restricts what NVIDIA can ship to Chinese buyers. Three levers, pulled simultaneously, on three companies who have had little choice but to comply.

The military applications are not hypothetical. A CSET Georgetown analysis of Chinese military procurement records identified 97 unique high-end AI chips ordered by the PLA. Nearly all identifiable chips were designed by US companies: NVIDIA, Xilinx, Intel, Microsemi.^{[6]CSET Georgetown, "Silicon Twist" analysis of PLA procurement records finding US-designed chips (2023)https://cset.georgetown.edu/publication/silicon-twist/}

This is the dual-use problem at its sharpest. The same GPU that trains a medical imaging model also runs a hedge fund's trading algorithms — or trains autonomous weapons of death. There is no technical distinction between civilian and military use of AI models.

The escalation

What follows is a chase that has played out across four years — Washington tightening rules, industry finding gaps, Washington tightening again — with each round revealing how difficult it is to control a general-purpose computing component.

Past export controls targeted finished weapons systems — things with obvious military applications and limited commercial use. But GPUs can be used for all kinds of commercial purposes. Furthermore, CoCom — the Cold War multilateral export control regime — regulated a few hundred categories of goods traded between multiple participating governments. But GPUs are for businesses, not governments.^{[7]Wikipedia, "Coordinating Committee for Multilateral Export Controls" (CoCom)https://en.wikipedia.org/wiki/Coordinating_Committee_for_Multilateral_Export_Controls}That mismatch is why today's chip controls are structured as US action, and not multilateral (multiple countries in alignment).

US AI chip export control escalation

October 2022: the first restrictions

It started here. The U.S. Bureau of Industry and Security (BIS) drew a line: any chip exceeding 4,800 TOPS or 600 GB/s now required a license for export to China.^{[2]Bureau of Industry and Security, Interim Final Rule on Advanced Computing and Semiconductor Manufacturing Items (October 7, 2022)https://www.federalregister.gov/documents/2022/10/13/2022-21658/implementation-of-additional-export-controls-certain-advanced-computing-and-semiconductor} The rule effectively banned NVIDIA's A100 and H100 overnight. BIS went further, extending the Foreign Direct Product Rule — which claims US jurisdiction over any chip made using American technology anywhere in the world — to semiconductors. Since virtually every advanced chip uses American EDA software or fabrication tools at some stage, the FDPR gave Washington global reach. It was the broadest unilateral technology restriction since the Cold War.

October 2023: closing the workaround

NVIDIA responded to the first rules the way NVIDIA responds to most problems: by engineering around them. The company designed the A800 and H800 specifically for the Chinese market, reducing interconnect bandwidth to slide beneath the control thresholds while preserving most of the training capability customers actually cared about. BIS concluded these chips provided "nearly comparable AI model training capability" and expanded the restrictions, shifting to a "total processing performance" metric that captured the workaround chips and extending controls to approximately 40 additional countries.^{[8]Bureau of Industry and Security, Updated Export Controls on Advanced Computing Items (October 17, 2023)https://www.federalregister.gov/documents/2023/10/25/2023-23055/implementation-of-additional-export-controls-certain-advanced-computing-items-supercomputer-and}

December 2024: targeting the tools

The third round went upstream. BIS added 24 types of semiconductor manufacturing equipment, three categories of software tools, and high-bandwidth memory to the controlled items list. It put 140 entities on the Entity List. This was no longer about restricting finished chips — it was about restricting the tools needed to build them.^{[9]Bureau of Industry and Security, Expanded Controls on Semiconductor Manufacturing Equipment, HBM, and 140 Entity List Additions (December 2024)https://www.bis.gov/press-release/commerce-strengthens-export-controls-restrict-chinas-capability-produce-advanced-semiconductors-military}

January 2025: dead on arrival

Five days before leaving office, the Biden administration published its most ambitious framework yet — the AI Diffusion Rule — dividing the world into three tiers: unlimited access for close allies, capped access for countries like India, the UAE, and Singapore, and a full ban for China, Russia, and North Korea.^{[10]Bureau of Industry and Security, Framework for Artificial Intelligence Diffusion, Interim Final Rule (January 15, 2025); rescinded May 13, 2025https://www.bis.gov/press-release/department-commerce-announces-rescission-biden-era-artificial-intelligence-diffusion-rule-strengthens} On May 13, 2025 — two days before the rule would have taken effect — the Trump administration rescinded it. The Commerce Department said the rule "would have stifled American innovation."^{[10]Bureau of Industry and Security, Framework for Artificial Intelligence Diffusion, Interim Final Rule (January 15, 2025); rescinded May 13, 2025https://www.bis.gov/press-release/department-commerce-announces-rescission-biden-era-artificial-intelligence-diffusion-rule-strengthens}

April 2025: the last legal chip

NVIDIA still had one chip it could legally sell to China: the H20, designed to fall below every performance threshold. Then reports surfaced that DeepSeek had used H20s to train its R1 reasoning model. BIS informed NVIDIA that the H20 would require an export license "indefinitely." NVIDIA disclosed $5.5 billion in expected charges.^{[11]NVIDIA SEC filing disclosing $5.5 billion in expected charges from H20 export license requirement (April 2025)https://www.npr.org/2025/04/16/nx-s1-5366665/nvidia-china-h20-chips-exports}

December 2025: the reversal

On December 8, 2025, Trump announced the US would allow NVIDIA to sell H200 chips to approved Chinese customers, with 25% of the revenue going to the US Treasury. "I have informed President Xi, of China, that the United States will allow NVIDIA to ship its H200 products to approved customers in China," Trump posted.^{[12]CNN, "Trump greenlights exports of Nvidia H200 chips to China" (December 8, 2025)https://www.cnn.com/2025/12/08/tech/nvidia-h200-chips-china-trump-export}

Only the H200 was approved. Blackwell and Rubin chips remain restricted. The H100 stays under the October 2022 controls. Commerce said similar arrangements for AMD and Intel were being finalized.^{[12]CNN, "Trump greenlights exports of Nvidia H200 chips to China" (December 8, 2025)https://www.cnn.com/2025/12/08/tech/nvidia-h200-chips-china-trump-export}

For NVIDIA, part of the $50 billion China market reopened overnight. For BIS, the same H200s its inspectors were chasing through Southeast Asian warehouses were now legally exportable with paperwork and a 25% tax.

Where things stand as of early 2026

The main controls remain in force. The AI Diffusion Rule is dead. BIS has revised its review posture from "presumption of denial" to "case-by-case review" — a shift that, in practice, means more ambiguity, not less. NVIDIA CEO Jensen Huang stood at Computex in May 2025 and said what everyone in the room already believed: "I think, all in all, the export control was a failure."^{[13]Fortune, "Jensen Huang says export control was a failure" (May 2025)https://fortune.com/2025/05/22/nvidia-ceo-jensen-huang-failure-us-restrictions-chips-semiconductors-china-ai-artificial-intelligence/} The $50 billion China market is, in his words, "effectively closed to U.S. industry" — but closed to US industry formally, does not mean Chinese buyers wouldn't find a way. NVIDIA's China market share dropped from 95% to 50%. That other 45% didn't disappear. It went underground, or offshore, or both.

So where did it go?

The smugglers

The cases that follow are the ones that resulted in indictments or public investigations. Industry participants widely believe they represent a fraction of total activity — that for every Supermicro or Megaspeed that gets caught, dozens of smaller operations move chips without detection. The documented evasion methods range from sophisticated corporate structures spanning four countries to individuals carrying GPUs through airport customs claiming "personal use as a gaming PC." The assumption is that most diversion runs through Southeast Asia. A Wire China investigation found that much of it goes directly through the United States — chips purchased by US-registered entities, shipped to US addresses, and routed out from there.

Documented smuggling cases by value of chips involved

Sources: DOJ indictments, Tom's Hardware, CNN. Amounts represent alleged value of chips involved, not necessarily profits.

The hair dryer

In December 2025, a BIS inspector visited a warehouse in Southeast Asia as part of a routine audit. What he found was not routine. In one area, thousands of non-working "dummy servers" had been staged to create the impression of a legitimate operation. In another, someone was using a hair dryer to peel serial-number stickers off NVIDIA servers.

The warehouse belonged to a pass-through company that federal prosecutors allege was controlled by Supermicro co-founder Yih-Shyan "Wally" Liaw.^{[3]CNN, "Co-founder of tech company charged with diverting $2.5 billion in Nvidia AI chips to China" (March 2026)https://www.cnn.com/2026/03/19/politics/super-micro-computer-founder-charged-ai-chips-china} The operation, according to the indictment, worked like this: servers containing NVIDIA H200 and B200 GPUs were assembled in the United States, shipped to Taiwan, forwarded to Southeast Asia, repackaged in unmarked boxes, and diverted to mainland China. The Southeast Asian pass-through company became Supermicro's 11th-largest customer globally — $99.7 million in fiscal 2024 alone.

The pace accelerated when it mattered most. When the Trump administration announced new restrictions effective May 13, 2025, Liaw allegedly texted his associates: "We need to speed these up before May 13!" Over $500 million in servers shipped to China in just three weeks between late April and mid-May 2025. The total alleged diversion: $2.5 billion. Supermicro stock fell 33% the day the charges were announced.

The fictitious brand

Operation Gatekeeper — the DOJ's name, not ours — centered on Hao Global LLC, a Houston shell company that placed a $55.6 million order with Lenovo for 800 H100 and 1,600 H200 GPUs, claiming the end user was a Thailand-based AI company.^{[14]DOJ, "U.S. Authorities Shut Down Major China-Linked AI Tech Smuggling Network" (Operation Gatekeeper, December 2025)https://www.justice.gov/opa/pr/us-authorities-shut-down-major-china-linked-ai-tech-smuggling-network} The end user did not exist. Workers in a New Jersey facility removed NVIDIA labels and barcodes from every chip and replaced them with "SANDKYAN" labels — a completely fictitious brand, invented for the sole purpose of disguising what the hardware was — and reclassified the GPUs as "adapters" on export paperwork. The operation moved $160 million through New York to Singapore, Hong Kong, Canada, and ultimately China.

The real estate company

Janford Realtor LLC was registered in Tampa, Florida. It never conducted a single real estate transaction. It existed for one purpose: to receive wire transfers from Chinese entities and purchase GPUs.^{[15]DOJ, "U.S. Citizens and Chinese Nationals Arrested for Exporting AI Technology" (Janford Realtor LLC case, November 2025)https://www.justice.gov/opa/pr/us-citizens-and-chinese-nationals-arrested-exporting-artificial-intelligence-technology} The company received $4 million in wires from a Chinese bank account, bought NVIDIA A100 and H100 GPUs, and shipped 400 of them to China through Malaysia and Thailand. Law enforcement intercepted a third shipment — 10 HPE supercomputers stuffed with H100 GPUs and 50 H200 GPUs — before it left the country. The four defendants include a US citizen born in Hong Kong, a Chinese national on an F-1 student visa, and a US citizen from Huntsville, Alabama. Each faces up to 20 years.

The premier partner

Megaspeed International described itself as a "premier business partner of Nvidia." It was not wrong, exactly. Formed in 2023 as an offshore entity of 7Road — a state-backed Chinese cloud company — Megaspeed acquired at least $4.6 billion in NVIDIA hardware and more than 136,000 GPUs, representing nearly 2% of NVIDIA's worldwide data center sales.^{[16]Tom's Hardware, "Nvidia's biggest Southeast Asia customer exposes the limits of US AI export controls" (Megaspeed/7Road investigation, 2025)https://www.tomshardware.com/pc-components/gpus/nvidias-biggest-sea-customer-exposes-the-limits-of-us-ai-export-controls}

An on-the-ground investigation found a near-empty Singapore office. A small Malaysian storefront that lacked engineering staff. Ownership structures linked to shell entities in Shanghai. Megaspeed abandoned $3.2 billion in additional planned purchases and ceased to exist. Bain Capital's Bridge Data Centres quietly removed Megaspeed from its Malaysian facility.

The subsidiary loophole

Aivres Systems is a California company — incorporated in America, purchasing American products, operating on American soil. It is also a subsidiary of Inspur, a major Chinese tech firm that sits on the Entity List for supplying the Chinese military with supercomputing hardware.^{[16]Tom's Hardware, "Nvidia's biggest Southeast Asia customer exposes the limits of US AI export controls" (Megaspeed/7Road investigation, 2025)https://www.tomshardware.com/pc-components/gpus/nvidias-biggest-sea-customer-exposes-the-limits-of-us-ai-export-controls} Because Aivres operates as a US entity, it could lawfully buy NVIDIA products and resell them. It became the primary supplier for Megaspeed's $4.6 billion procurement operation and brokered a separate $100 million deal for 2,300 Blackwell GPUs for INF Tech, a Shanghai AI startup, routing the transaction through an Indonesian telecom company.

The rename

Sugon, a Chinese supercomputer manufacturer linked to military surveillance programs, was added to the Entity List in 2019. The fix was simple. Sugon's executives established Nettrix, which promptly partnered with NVIDIA, Intel, and Microsoft, selling servers equipped with American chips to Entity Listed organizations.^{[17]Asia Times, "Sugon spin-off helps China evade US chip bans" (August 2024)https://asiatimes.com/2024/08/sugon-spin-off-helps-china-evade-us-chip-bans/} Sugon and Nettrix share a physical office complex. They share overlapping executives. The Entity List only targets specific company names and addresses, so registering a new name at a neighboring address was enough to bypass it entirely. SemiAnalysis put it simply: "It's trivial to set up a friendly, non-restricted entity next door."^{[18]SemiAnalysis, "Fab Whack-A-Mole: Chinese Companies Are Evading U.S. Sanctions" (October 2024)https://semianalysis.com/2024/10/28/fab-whack-a-mole-chinese-companies/}

Two million chips through TSMC

Huawei — the company whose name appears at the center of nearly every US-China tech restriction — obtained approximately two million Ascend 910 AI chiplets from TSMC through a chain of intermediaries that included Sophgo, a design partner of crypto mining firm Bitmain.^{[19]Tom's Hardware, "Huawei reportedly acquired two million Ascend 910 AI chips from TSMC through shell companies" (2025)https://www.tomshardware.com/tech-industry/artificial-intelligence/huawei-reportedly-acquired-two-million-ascend-910-ai-chips-from-tsmc-last-year-through-shell-companies} TSMC did not know the chips were destined for Huawei. An independent teardown firm, TechInsights, identified the design after the fact. By then, millions of chips had already shipped. TSMC halted all remaining shipments and now faces a potential $1 billion fine from the Commerce Department.

How chips reach China: documented evasion methods

The middlemen

When we asked neocloud operators how the chips actually move, the answer was always simpler than the indictments suggest. "You just use a US entity," said one operator we'll call Kevin, who runs an APAC-focused neocloud. "US address, US bank account, US paperwork. Then you funnel everything through Hong Kong or Malaysia." He described it the way you'd describe setting up a Stripe account — procedural, routine, boring. Several operators told us NVIDIA itself was very friendly to APAC neoclouds at GTC and at private side events, whose end customers are, at best, ambiguous.

Chinese companies renting American compute

The export controls restricted where chips could be shipped. They did not restrict who could access compute remotely. A data center in Johor, Malaysia filled with NVIDIA GPUs, built by a Chinese-linked operator, staffed by Chinese engineers, serving Chinese customers — technically compliant, as long as the hardware stayed in Malaysia.^{[20]Asia Society Policy Institute, "Malaysia's Gamble: Turning Data Centres Into Industrial Power" (2025)https://asiasociety.org/policy-institute/malaysias-gamble-turning-data-centres-industrial-power}

So if you cannot ship the chip to China, rent it where it sits. ByteDance has been renting NVIDIA's most advanced chips from Oracle for AI computing on US soil — a $7 billion cloud-based GPU strategy that routes around the controls entirely.^{[21]Data Center Dynamics, "Chinese companies exploiting sanctions loophole by accessing GPUs in the US" (ByteDance-Oracle, 2025)https://www.datacenterdynamics.com/en/news/chinese-companies-exploiting-sanctions-loophole-by-accessing-gpus-in-the-us-report/} Alibaba and Tencent held similar discussions with NVIDIA. Tencent — designated by the Department of Defense as a "Chinese Military Company" — signed contracts worth over $1.2 billion with Japanese neocloud Datasection for 15,000 Blackwell processors in Japan, plus an $800 million deal for a data center in Sydney. Two smaller US cloud companies declined to rent NVIDIA chips to Chinese entities because it "seemed to go against the spirit" of the sanctions. Others did not decline.

The House passed the Remote Access Security Act 369-22 in January 2026, extending export controls to cloud computing for the first time.^{[22]U.S. House of Representatives, Remote Access Security Act (H.R. 2683), passed 369-22 (January 12, 2026)https://www.congress.gov/bill/119th-congress/house-bill/2683} Before that bill, the controls had no answer for remote access.

Johor's boom

Johor built data-center capacity by operator, mid-2025

Source: Data Center Dynamics, "Johor: Past, Present and Future" (2025).

Johor state had approved 42 data center projects totaling approximately $39 billion in investment as of mid-2025.^{[23]DCD, "Johor: Past, Present and Future" (data center capacity analysis, 2025)https://www.datacenterdynamics.com/en/analysis/the-past-present-and-future-of-johor/} Chinese-linked operators control the majority of that capacity. Bridge Data Centres — owned by Bain Capital, formerly Chindata Group — holds 32%. DayOne — formerly GDS International, China's largest carrier-neutral data center operator — holds 29%.^{[23]DCD, "Johor: Past, Present and Future" (data center capacity analysis, 2025)https://www.datacenterdynamics.com/en/analysis/the-past-present-and-future-of-johor/}

ByteDance is the anchor tenant at Bridge's MY06 facility, a 110 MW hyperscale data center in Sedenak Tech Park. In March 2026, ByteDance signed a deal with Aolani Cloud for approximately 36,000 Blackwell B200 GPUs in Malaysia — an estimated $2.5 billion.^{[24]Tom's Hardware, "ByteDance to access 36,000 Blackwell GPU cluster through Malaysia cloud operator" (March 2026)https://www.tomshardware.com/pc-components/gpus/chinas-bytedance-to-access-36-000-blackwell-gpu-cluster-through-malaysia-cloud-operator-nvidia-confirms-no-objections-deal-is-in-line-with-us-export-controls} The company had already been leasing H100s from Aolani since February 2025, with a separate $2.1 billion facility expansion planned.

The build-out extends well beyond ByteDance. Alibaba Cloud collaborated with the Malaysia Internet Exchange to create a 120 MW data hub in Johor's Iskandar region, housing 50,000 NVIDIA A100 GPUs for Tencent and ByteDance. YTL Green Data Center Park in Kulai — a 605 MW facility backed by $3.5 billion in total investment — has 70% of its capacity leased to Chinese AI firms. NVIDIA and YTL partnered on $4.3 billion in AI data centers in Malaysia. The scale of Chinese capital flowing into Malaysian data infrastructure is not subtle.

The Aperia case and Malaysia's response

In February 2025, the scene shifted to Singapore, where Aperia Cloud Services CEO Alan Wei and COO Aaron Woon were charged with fraud for misleading Dell and Supermicro about the end users of servers containing NVIDIA chips.^{[25]Fortune, "Singapore charges three with fraud in cases reportedly linked to smuggling of Nvidia chips to China" (February 2025)https://fortune.com/asia/2025/02/28/singapore-charges-fraud-nvidia-chip-smuggling/} The chips were allegedly diverted to DeepSeek. A fourth defendant — Aperia CFO Jenny Lim — was charged in April 2026. In total, nine people were arrested across raids at 22 locations.^{[26]Tom's Hardware, "Singapore police bust major ring smuggling Nvidia GPUs to China-based DeepSeek" (February 2025)https://www.tomshardware.com/tech-industry/artificial-intelligence/singapore-police-bust-major-ring-smuggling-nvidia-gpus-to-china-based-deepseek-report}

Malaysia tightened its own controls in July 2025, requiring a Strategic Trade Permit for all exports, transshipments, or transits of US-origin AI chips, with 30-day advance notification for suspected diversion.^{[27]TechWire Asia, "Malaysia tightens oversight of data centres amid AI chip controls" (September 2025)https://techwireasia.com/2025/09/malaysia-tightens-oversight-of-data-centres-amid-ai-chip-controls/} Trade Minister Tengku Zafrul formed a task force to ensure "servers end up in data centres that they're supposed to and not suddenly move to another ship." Johor rejected approximately 30% of new data center applications in 2024 — a signal that even the beneficiaries of the boom were beginning to worry about what it looked like.

NVIDIA's Singapore gap

NVIDIA does not fully disclose revenue by end-user geography — it reports billing location, not delivery destination. That is why Singapore appears as 28% of revenue while receiving 1% of chips.^{[28]Tom's Hardware, "NVIDIA Singapore GPU sales are 28% of revenue but only 1% are delivered to the country" (2026)https://www.tomshardware.com/tech-industry/deepseek-gpu-smuggling-probe-shows-nvidias-singapore-gpu-sales-are-28-percent-of-its-revenue-but-only-1-percent-are-delivered-to-the-country-report}Reported China exposure dropped from 26% to roughly 13% after the October 2022 controls.

NVIDIA has acknowledged this in SEC filings, noting that "customers use Singapore to centralize invoicing while our products are almost always shipped elsewhere." The question is where "elsewhere" ends up. Senators Banks and Warren wrote to Commerce Secretary Lutnick: "If Nvidia could not detect and prohibit non-China customers from allegedly illegally diverting GPUs bound for Southeast Asia to China, it is extremely implausible that it can meet existing end user/use control requirements."^{[29]Senate Banking Committee, Senators Banks and Warren letter to Commerce Secretary Lutnick on NVIDIA export licenses (March 2026)https://www.banking.senate.gov/newsroom/minority/senators-banks-and-warren-urge-commerce-secretary-lutnick-to-suspend-or-reconsider-nvidia-export-licenses-amid-supermicro-chip-diversion-indictment}

The enforcers

The technology gap

When a member of Congress asked BIS Undersecretary Alan Estevez whether the bureau had modern IT systems, his answer was memorable. "The answer to that is an emphatic no. We're using antiquated systems fielded in the mid-2000s using 1990s technology."^{[30]Alan Estevez, BIS Undersecretary testimony to House Foreign Affairs Committee (March 2024)https://www.congress.gov/118/meeting/house/117003/witnesses/HHRG-118-FA00-Wstate-EstevezA-20240321.pdf} The license adjudication system dates from 2006. The enforcement system dates from 2008. Analysts — the people tasked with determining whether a shell company in Malaysia is a front for Chinese military procurement — "perform their work primarily using Google searches and Microsoft Excel." CSIS estimates modern tools could deliver 5-10x greater analyst productivity.

The math

BIS has fewer than 600 total employees and an annual budget of $191 million.^{[31]CSIS, "Mismatch of Strategy and Budgets in AI Chip Export Controls" (December 2024)https://www.csis.org/analysis/mismatch-strategy-and-budgets-ai-chip-export-controls} The profits from just three documented smuggling cases exceed double the bureau's entire annual enforcement budget.^{[4]CNAS, "Countering AI Chip Smuggling Has Become a National Security Priority" (2025)https://www.cnas.org/publications/reports/countering-ai-chip-smuggling-has-become-a-national-security-priority} CSIS called the budget "approximately equivalent to two F-35 fighters."^{[31]CSIS, "Mismatch of Strategy and Budgets in AI Chip Export Controls" (December 2024)https://www.csis.org/analysis/mismatch-strategy-and-budgets-ai-chip-export-controls}

The workload is staggering. Nearly 38,000 license applications in FY2023, up 53% from about 25,000 a decade earlier — processed by roughly 150 sworn special agents.^{[32]GAO, "Export Controls: Commerce Should Improve Workforce Planning and Information Sharing" (2025)https://www.gao.gov/products/gao-25-107431} A single export control officer covers Australia, Singapore, and Malaysia. Eleven officers are stationed at US embassies worldwide — eleven, for the entire planet. In 2024, BIS completed approximately 1,440 end-use checks across 60 countries. A shell company can be created online in hours for a few thousand dollars. BIS takes years to investigate one.^{[4]CNAS, "Countering AI Chip Smuggling Has Become a National Security Priority" (2025)https://www.cnas.org/publications/reports/countering-ai-chip-smuggling-has-become-a-national-security-priority}

CSIS Director Gregory Allen wrote what might serve as the epitaph for the current enforcement regime: "China is betting that its network of smugglers and shell companies can find the leaks in the BIS export control enforcement barrier. As long as Congress continues to neglect BIS by providing grossly inadequate resources, China has a reasonable expectation of success."^{[31]CSIS, "Mismatch of Strategy and Budgets in AI Chip Export Controls" (December 2024)https://www.csis.org/analysis/mismatch-strategy-and-budgets-ai-chip-export-controls}

NVIDIA's incentive problem

NVIDIA is contractually required to perform end-use diligence — verify the buyer, confirm the destination, ensure compliance. But NVIDIA also holds roughly 80% market share in AI training hardware, and every chip that reaches a Chinese end user is revenue the company would lose if enforcement were airtight. The financial incentive to not look too hard is not a conspiracy theory. It is a structural feature of a system that asks the dominant seller to police its own demand.

The neocloud operators we spoke with described the dynamic from the other side of the table. One — we'll call him Ray — told us he fabricates utilization data to justify requesting larger GPU allocations from NVIDIA. His data suggests his clusters are running near capacity on legitimate workloads. They are not. The chips get rented out to Chinese customers.

Jensen Huang has publicly called the controls "a failure" and lobbied for relaxation.^{[13]Fortune, "Jensen Huang says export control was a failure" (May 2025)https://fortune.com/2025/05/22/nvidia-ceo-jensen-huang-failure-us-restrictions-chips-semiconductors-china-ai-artificial-intelligence/} His argument is that it'd be far better to let China build on top of American chips, than let them build up their own ecosystem. He framed it as a matter of national pride, NVIDIA is an American company that has the right to win in China too.

But Cerebras CEO Andrew Feldman went further, accusing NVIDIA of having "armed China single-handedly." Senators Banks and Warren questioned whether NVIDIA can credibly self-police when its largest growth market depends on the supply chains the controls are supposed to restrict.^{[29]Senate Banking Committee, Senators Banks and Warren letter to Commerce Secretary Lutnick on NVIDIA export licenses (March 2026)https://www.banking.senate.gov/newsroom/minority/senators-banks-and-warren-urge-commerce-secretary-lutnick-to-suspend-or-reconsider-nvidia-export-licenses-amid-supermicro-chip-diversion-indictment}

How many chips got through

How China accesses AI compute

Estimated share of ~2.8M H100-equivalents (ChinaTalk, 2025). Proportions are directional estimates based on ChinaTalk, CNAS, and SemiAnalysis reporting. Exact breakdown is unknown.

CNAS puts the median estimate at approximately 140,000 chips smuggled to China in 2024, with a range stretching from 17,500 to 780,000 — a spread that itself reflects how little anyone actually knows.^{[4]CNAS, "Countering AI Chip Smuggling Has Become a National Security Priority" (2025)https://www.cnas.org/publications/reports/countering-ai-chip-smuggling-has-become-a-national-security-priority} ChinaTalk's broader estimate — incorporating legal purchases, smuggling, domestic production, and remote access — reaches 2.8 million H100-equivalents.^{[5]ChinaTalk, "How Many Chips Does China Have?" estimate of 2.8 million H100-equivalents (2025)https://www.chinatalk.media/p/how-many-chips-does-china-have} Of 22 notable Chinese AI models developed through 2025, only two were trained exclusively on Chinese-designed chips. At Huaqiangbei — the same market where Mr. 5 explained how the game works — researchers found 132 domestic listings for export-controlled chips, with aggregate inventory of roughly 100,000 GPUs.^{[4]CNAS, "Countering AI Chip Smuggling Has Become a National Security Priority" (2025)https://www.cnas.org/publications/reports/countering-ai-chip-smuggling-has-become-a-national-security-priority}

Legislative response

The Chip Security Act passed the House Foreign Affairs Committee 42-0 in March 2026, requiring exported AI chips to include location verification, periodic server check-ins, and software-based authentication before export.^{[33]Congress.gov, H.R.3447 "Chip Security Act" (119th Congress, 2025-2026)https://www.congress.gov/bill/119th-congress/house-bill/3447/text} The Stop Stealing Our Chips Act (December 2025) creates a whistleblower incentive: 10-30% of resulting penalties for actionable tips, with anti-retaliation protections. BIS median annual penalties since 1996: only $13 million (inflation-adjusted). The White House requested $313 million for BIS in fiscal year 2026, a 64% increase.^{[32]GAO, "Export Controls: Commerce Should Improve Workforce Planning and Information Sharing" (2025)https://www.gao.gov/products/gao-25-107431}

The talent war

The smuggling story is about chips. But there is a parallel front where the commodity being moved is not silicon — it is the people who know how to make it. Over 3,000 Taiwanese chip engineers have moved to mainland China, roughly 10% of Taiwan's 40,000 semiconductor R&D workforce.^{[34]Nikkei Asia, "Taiwan loses 3,000 chip engineers to Made in China 2025"https://asia.nikkei.com/Business/China-tech/Taiwan-loses-3-000-chip-engineers-to-Made-in-China-2025} These are not junior hires. China wants process engineers with specific knowledge of 7nm, 5nm, 3nm, and 2nm fabrication — knowledge that cannot be replicated by buying equipment alone.

The offer

Huawei contacts TSMC engineers approximately every three months, offering triple their current salary.^{[35]Tom's Hardware, "China intensifies efforts to poach semiconductor talent from Taiwan" (2025)https://www.tomshardware.com/tech-industry/semiconductors/china-intensifies-efforts-to-poach-semiconductor-talent-from-taiwan-claims-report-international-restrictions-motivate-illicit-efforts-to-obtain-talent-and-equipment} SMIC set up a front company in Hsinchu — TSMC's headquarters city — disguised as a Samoa-based entity to circumvent Taiwan's laws prohibiting unauthorized Chinese business operations.^{[36]CNBC, "Taiwan alleges China chipmaker SMIC illegally poached tech talent" (March 2025)https://www.cnbc.com/2025/03/28/taiwan-alleges-china-chipmaker-smic-illegally-poached-tech-talent.html} In March 2025, Taiwan's investigation bureau raided 34 locations across six cities, questioned 90 individuals, and investigated 11 Chinese enterprises for illegal talent poaching, with SMIC as the headline target.^{[37]Caixin Global, "China Hires Over 100 TSMC Engineers in Push for Chip Leadership" (August 2020)https://www.caixinglobal.com/2020-08-12/china-hires-over-100-tsmc-engineers-in-push-for-chip-leadership-101591851.html} A separate, earlier prosecution targeted Bitmain, which used two front companies — IC Link and WiseCore Technology — to recruit at least 100 engineers from TSMC, MediaTek, and other Taiwanese firms over roughly three years beginning in 2018.

The financial incentives are staggering. Liang Mong-song, former TSMC Senior R&D Director credited with nearly 500 TSMC patents, joined SMIC as Co-CEO in 2017. His annual pay jumped to $1.53 million — up 450% in a single year. He received a housing unit worth $3.4 million and roughly one million SMIC shares. Total 2020 compensation: approximately $4.9 million.

The remote poaching model

Bitmain pioneered something more subtle than relocation: engineers remain physically in Taiwan while working for Chinese companies remotely. IC designers can do their work without leaving the island. Bitmain employed Taiwanese engineers at WiseCore, had TSMC manufacture the chips, and ASE package them — all without a single engineer crossing a border. The knowledge moved. The people stayed.

Taiwan's crackdown

In 2022, Taiwan amended the National Security Act to add charges for economic espionage related to "national core key technologies" in semiconductors. Penalties: up to 12 years imprisonment and fines of $1 million to $3.5 million.

The first prosecution came in August 2025. A former TSMC engineer named Chen Li-ming, who had joined Tokyo Electron's marketing department, used past connections to obtain 2nm process trade secrets from current TSMC employees.^{[38]Focus Taiwan, "3 TSMC employees indicted on trade secret theft charges" (August 2025)https://focustaiwan.tw/society/202508270024} He photographed over 1,000 images of confidential content on TSMC computer screens. Prosecutors are seeking 14 years. Two current TSMC engineers who supplied information face 7-9 years. Tokyo Electron Taiwan faces NT$120 million ($3.82 million) in fines. Taiwan's investigation bureau has examined 100 similar cases since 2020.

The other side

China is not only evading the controls. It is building around them. Reuters described the effort as China's version of the Manhattan Project — a state-directed crash program to replicate what the West spent decades developing.^{[39]Reuters, "How China built its 'Manhattan Project' to rival the West in AI chips" (December 17, 2025)https://www.reuters.com/world/china/how-china-built-its-manhattan-project-rival-west-ai-chips-2025-12-17/}

In December 2025, a research facility in Shenzhen validated a functional prototype of a domestic EUV lithography machine — the kind of equipment that only ASML in the Netherlands currently produces.^{[39]Reuters, "How China built its 'Manhattan Project' to rival the West in AI chips" (December 17, 2025)https://www.reuters.com/world/china/how-china-built-its-manhattan-project-rival-west-ai-chips-2025-12-17/} The machine was built by a team that included former ASML engineers who reverse-engineered the company's technology. Huawei is coordinating thousands of engineers across companies and state research institutes, with recruits working under false identities inside secure facilities. The secrecy is intentional.

In May 2024, China announced a $47.5 billion state investment fund for semiconductor development. By November 2025, state-funded data centers were ordered to discontinue foreign AI chip purchases entirely. Huawei produced an estimated 200,000 AI chips in 2025, up from near-zero domestic production — still a fraction of what China accesses through legal and illegal channels, but a trajectory that moves in one direction.

And China has its own chokepoint. In July 2023, Beijing restricted exports of gallium and germanium, two materials critical to semiconductor manufacturing. China produces roughly 90% of the world's gallium and 60% of its germanium. By December 2024, the restriction escalated to an outright export ban on these materials to the United States — the first country-specific ban. SemiAnalysis assessed the state of play: "There is clear evidence after two years under the controls that the [Chinese] companies are not suffering, they are thriving."^{[18]SemiAnalysis, "Fab Whack-A-Mole: Chinese Companies Are Evading U.S. Sanctions" (October 2024)https://semianalysis.com/2024/10/28/fab-whack-a-mole-chinese-companies/}

The global fallout

The export controls set off a chain reaction that neither Washington nor Beijing fully anticipated. Countries that once would have bought cloud capacity from AWS, Azure, or Google are now building their own compute infrastructure — not because they want to, but because depending on American export policy for access to AI feels, to many governments, like a risk they can no longer take.

This has benefited NVIDIA. Its sovereign AI revenue exceeded $30 billion for fiscal year 2026, more than tripling year-over-year while overall revenue grew 65% — sovereign AI expanding roughly three times faster than the company overall, and reshaping which countries and companies are building compute.^{[40]NVIDIA Q4 FY2026 Earnings, sovereign AI revenue exceeding $30 billion for full fiscal yearhttps://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-fourth-quarter-and-fiscal-2026}

Sovereign AI GPU commitments

Sources: NVIDIA newsroom, government press releases. Figures represent announced commitments; actual deployments may differ.

The Gulf

Saudi Arabia's Humain plans to deploy up to 600,000 NVIDIA GPUs over three years, roughly $18-24 billion in GPU spend. The first 18,000 GB300 chips were announced in November 2025. xAI is building its first major deployment outside the US at Humain's 500 MW flagship facility.^{[41]Data Center Dynamics, "Saudi Arabian AI venture HUMAIN buys 18,000 NVIDIA GB300 chips" (2026)https://www.datacenterdynamics.com/en/news/saudi-arabian-ai-venture-humain-buys-18000-nvidia-gb300-chips-several-hundred-thousand-more-on-the-way/}

The UAE's Stargate project broke ground in March 2026: $30 billion for a 5 GW Abu Dhabi AI campus anchored by a 1 GW Stargate cluster, with G42 allocated roughly 100,000 advanced NVIDIA chips annually under the 2025 US-UAE agreement. Commerce stalled parts of the deal over G42's historical ties to Huawei. G42 severed its China investments in 2024 in exchange for $1.5 billion from Microsoft.^{[42]Middle East Institute, "From Crude to Compute: Building the GCC AI Stack" (2025)https://www.mei.edu/publications/crude-compute-building-gcc-ai-stack}

US policymakers worry that Gulf sovereign AI programs could become the next conduit for Chinese access to compute. Both Saudi Arabia and the UAE maintain technology relationships with Beijing. The original UAE deal envisioned up to 500,000 H100 chips annually — a volume that raised questions about whether all that capacity would serve domestic needs. Export licenses for cutting-edge NVIDIA chips to the G42 remain in limbo. G42 was an investor in Cerebras until December 2025, when it exited following a US national security review.

Asia

Japan committed 1 trillion yen ($6.34 billion) to a Sovereign AI package, with combined public and private AI investment exceeding $135 billion through 2030.^{[43]NVIDIA Blog, "NVIDIA and SoftBank Accelerate Japan's Journey to Global AI Powerhouse" (2025)https://nvidianews.nvidia.com/news/nvidia-and-softbank-accelerate-japans-journey-to-global-ai-powerhouse} South Korea committed over 250,000 NVIDIA GPUs across sovereign clouds as part of a $735 billion initiative.

India is moving at a different scale. Reliance pledged $110 billion at the India AI Impact Summit 2026, Microsoft $17.5 billion, AWS $15.2 billion.^{[44]Microsoft News, "Microsoft invests US$17.5 billion in India" (December 2025)https://news.microsoft.com/source/asia/2025/12/09/microsoft-invests-us17-5-billion-in-india-to-drive-ai-diffusion-at-population-scale/} The government targets 100,000 sovereign GPUs by late 2026.

Europe

The UK committed up to £11 billion (about $15 billion) to sovereign AI, including 120,000 NVIDIA Blackwell GPUs.^{[45]NVIDIA Newsroom, "NVIDIA and United Kingdom Build Nation's AI Infrastructure" (2025)https://nvidianews.nvidia.com/news/nvidia-and-united-kingdom-build-nations-ai-infrastructure-and-ecosystem-to-fuel-innovation-economic-growth-and-jobs} Deutsche Telekom launched one of Europe's largest AI factories in Munich with nearly 10,000 Blackwell GPUs.^{[46]Deutsche Telekom, "AI sovereignty for Germany and Europe" with NVIDIA AI factory launch (2026)https://www.telekom.com/en/media/media-information/archive/ai-sovereignty-for-germany-and-europe-1098708} France committed €109 billion, with Mistral AI deploying 18,000 Grace Blackwell systems. The EU's InvestAI initiative targets €200 billion across five gigafactories.

The mismatch

The Chip Security Act — passed 42-0 in committee — would require location verification and server check-ins for exported chips.^{[33]Congress.gov, H.R.3447 "Chip Security Act" (119th Congress, 2025-2026)https://www.congress.gov/bill/119th-congress/house-bill/3447/text} The Stop Stealing Our Chips Act creates whistleblower incentives at 10-30% of resulting penalties. The White House requested $313 million for BIS in fiscal year 2026, a 64% increase.^{[32]GAO, "Export Controls: Commerce Should Improve Workforce Planning and Information Sharing" (2025)https://www.gao.gov/products/gao-25-107431} These are real responses to a real problem.

The enforcement gap is a bug that can't be patched with legislation. The structural tension the policies do not resolve is this: the company responsible for policing chip diversion is the same company that benefits most from looking the other way. The bureau tasked with enforcement has 150 agents, 2006-era IT, and a tiny budget.

BIS median annual penalties since 1996: $13 million, inflation-adjusted. The profits from a single smuggling operation can exceed that in a week. Shell companies are created in hours. Investigations take years.

China's counter-moves add a second dimension of uncertainty. Gallium and germanium export bans could disrupt semiconductor supply chains if tensions escalate.

The phrase heard most often, from Silicon Valley to Johor to Singapore, remains the same: open one eye, close one eye.

References

Residual Value Insurance Solutions for GPUs

Coverage creates a minimum value ("floor") for what your GPUs are worth at a future date. If they sell below the floor, the policy pays you the difference.

Learn how it works →

ShareLinkedIn X Facebook