Why the Nvidia–Megaspeed story matters for U.S. national security

Why the Nvidia–Megaspeed story matters for U.S. national security

Imagine a crisis over Taiwan. Both sides have satellites, drones, and radar feeding them more images and signals than any army of analysts could ever review. The side with better AI can sort through all of it faster: spotting ships as they leave port, tracking where forces are gathering, and flagging which targets matter most, hours or even days before the other side catches up. In a crisis, that kind of head start can decide who moves first and who is left reacting.

The same race is already underway in cyberspace, and it doesn’t wait for a war. AI can scan software for weak spots far faster than human experts can. Attackers use it to find ways in, and defenders use it to find and patch the holes first. U.S. officials have warned that Chinese government hackers have already broken into American telecom networks and planted footholds in systems that run power and water. The side with stronger AI is better placed to find those openings first, whether to exploit them or to close them.

That is the practical reason Washington cares who can use Nvidia’s most powerful chips. Computing power is what builds and runs those AI systems, and the best chips today are designed by an American company. The same equipment can help a hospital read scans or help a military make faster decisions. The security stakes depend on who gets the computing power, how much they get, and what they use it for. Those are the questions at the center of the Megaspeed story.

The short version

On November 12, Nvidia announced that Jensen Huang had resigned as president, CEO, and board member, effective immediately. Chief financial officer Colette Kress became interim CEO while the board searches for a permanent successor, and longtime director Harvey Jones became the company’s first-ever board chair. Huang had led Nvidia since he co-founded it in 1993. His departure ends one chapter, but the central security questions are still open. Who got access to the chips, and what did officials do after they were warned?

Megaspeed, a small Singapore company with roots in a Chinese state-linked gaming firm, became one of Nvidia’s biggest customers in Southeast Asia and bought billions of dollars of advanced AI hardware starting in 2024. U.S. officials suspected that some of that hardware was helping Chinese companies get around American export controls. A senior Megaspeed executive, Alice Huang (no relation), had a romantic relationship with Jensen Huang, and U.S. intelligence concluded that China’s spy agency had recruited her, partly to push him to lobby Washington for looser chip rules.

That much is about Nvidia. The rest of the story is about Washington. According to the reporting, President Trump was briefed on the intelligence in late 2025 and again in May 2026. A Commerce Department investigation into Megaspeed was moved to the Justice Department after Nvidia’s top lawyer contacted Commerce officials. Four federal agencies pushed to blacklist Megaspeed, and Commerce has not done it. Trump reportedly ordered a crackdown on illegal chip diversion three times this year, and his Commerce secretary, Howard Lutnick, reportedly stalled it each time. Meanwhile, the administration opened the door to more legal chip sales to China, brought Jensen Huang along on a presidential trip to Beijing, and gave him the National Medal of Science. 

Why these chips are worth restricting

An AI chip does many calculations at once. Put thousands of advanced chips together, connect them with fast networks, and they can work as a single system. A team can use that system to train AI, which means teaching it to recognize patterns from large amounts of data. It can also use the system to run trained AI across vast amounts of new information. Think of the difference between a tool that can review one satellite image and a service that can review a steady stream of images from many locations, day and night.

Speed matters at both stages. Developing a reliable system takes repeated experiments. A team trains it, tests it, finds mistakes, and tries again. Faster chips let the team complete more of those cycles in a year. Once the system is deployed, more computing power lets it handle more work at once. The chips also need power, cooling, fast connections, good data, and people who know how to use them. A large, working cluster is the valuable thing.

The U.S. Commerce Department has named military planning, logistics, radar, and signals intelligence among its concerns. In plain English, that includes helping a military sort through intercepted communications, spot changes in large volumes of sensor data, or plan how to move and resupply its forces. The same technology can help intelligence agencies and police expand surveillance. It also has plenty of civilian uses, and the public record has not established any military use in the Megaspeed case.

Why does it matter if the end user is a “civilian” Chinese company? Because in China the line between commercial and military work is thin by design. Beijing’s official policy of “military-civil fusion” directs private companies to support the armed forces, and Chinese law requires organizations to cooperate with state intelligence work when asked. That helps explain why, on June 8, 2026, the Pentagon added Alibaba and Baidu to its list of companies it says support China’s military, joining Tencent, which was added in 2025. All three deny military ties. These are the same companies at the center of the Megaspeed story and of the administration’s licensed chip sales.

People on the front lines of this competition say computing power is the bottleneck. In May 2025, a vice president at Tencent’s cloud unit called the shortage of AI chips and computing resources the company’s most severe problem and warned that tighter U.S. controls would widen the gap. Greg Allen of the Center for Strategic and International Studies has testified that without the export controls, the first million-chip AI cluster might well have been built in China instead of the United States.

Can China make chips as powerful as Nvidia’s best?

China makes its own AI chips, including Huawei’s Ascend line. So far, the evidence indicates that China has not shown it can manufacture chips comparable to Nvidia’s best at similar scale. Nvidia designs its chips and relies on Taiwan’s TSMC to manufacture its leading processors. China’s biggest contract chipmaker, SMIC, has been producing advanced chips, but its mass-production technology still trails TSMC’s. On September 17, 2026, Huawei’s rotating chairman, Eric Xu, said the company could not produce enough AI computing equipment to meet demand inside China and was limiting sales abroad. Huawei is trying to make up for weaker individual chips by wiring huge numbers of them together, which takes more electricity, more space, and more engineering.

Peter Zeihan makes a distinction in his November 12 discussion that helps explain why this matters. A finished chip gives a buyer computing power today. A chip design could help a rival understand how the chip works, but manufacturing it takes specialized equipment, materials, and the ability to make large numbers of working chips reliably. Zeihan raises possible access to Nvidia’s designs as a concern. His discussion offers no evidence that Beijing obtained them.

Chinese developers can still make progress with domestic chips, more efficient software, or larger groups of less powerful chips. So why bother restricting access to Nvidia’s best? Because the goal is to raise the cost and slow the work. If a team needs more machines, more electricity, or more time to run the same experiments, it has fewer chances to improve its systems and may take longer to put them to use. That delay gives the United States and its allies time to improve their own capabilities and defenses. Controls have costs, too. They can reduce U.S. chipmakers’ sales and give Chinese firms a stronger reason to build around domestic chips. The judgment Washington has to make is whether selling a particular chip, in a particular quantity, to a particular customer is worth giving up some of that time.

Jacob Shapiro’s November 10 video spells out the strongest objection to export controls, and he says he thinks Huang’s argument is the right one. If Chinese companies can buy American chips and build their software around them, they have less reason to switch to Chinese alternatives. Cut off that supply and Chinese chipmakers gain customers, investment, and a reason to improve faster. Shapiro points to Huawei, which the controls did not kill and arguably pushed to innovate, and to the first Trump administration’s failed effort to block Huawei’s 5G equipment. But Shapiro also concedes the key point on the other side: you can’t stop the spread of technology, but you can slow it down. Slowing it down is exactly what the controls are for. Both effects can be true. Restrictions can slow some projects today and speed up China’s effort to rely on its own suppliers over the longer run. The policy question is how much time the U.S. gains and whether that gain outweighs the cost. Whether Megaspeed complied with the rules that actually applied is a separate question.

If this is an AI race, why give China access to our fastest chips?

We’re told the United States can’t afford to slow down AI development or deployment because China might get ahead. If speed matters that much here, it matters in China too. Giving Chinese developers access to our most powerful chips can help them run more experiments, improve their models faster, and deploy them on a larger scale. That is the same advantage that supposedly makes it so urgent for us to build faster at home.

There is a case for licensing sales of some less powerful chips. Keeping Chinese companies on American hardware and software could give them less reason to switch to domestic suppliers. But that argument needs to explain where the line is and why. Which chips, in what quantities, for which customers, and with what safeguards? A case for limited sales does little to explain why we should shrug if restricted users get access to the most powerful chips through an intermediary.

If the same people invoke the race with China to dismiss calls for caution at home and then wave away the risk of giving China access to our best chips, I think that’s hypocritical. They treat computing power as decisive when it helps make the case for moving faster here, then treat it as incidental when China gets access to it. If getting there first really matters, helping a competitor get there sooner deserves a serious explanation.

The administration’s own officials have undercut the idea that these decisions were purely about security. In July 2025, after the U.S. reversed course and allowed Nvidia to resume selling its China-specific H20 chip, Treasury Secretary Scott Bessent described the earlier restriction as a “negotiating chip” used in trade talks with Beijing. Commerce Undersecretary Jeffrey Kessler had told Congress weeks earlier, under oath, that controls are based on national security and foreign policy, not bargaining leverage. Bessent’s comment matters because export controls only work if allies and companies believe they are about security. Once they look like trade concessions, every rule becomes negotiable.

How Megaspeed fits in

Megaspeed is a Singapore-based company that rents out AI computing power. It was created in 2023, when 7Road, a Chinese gaming and cloud company with ties to state-backed investors, spun off its overseas operation in Singapore and renamed it Megaspeed. As Shapiro points out, that happened only months after the Biden administration’s sweeping October 2022 restrictions on chip exports to China. Senator Tom Cotton has described Megaspeed as a spin-off of a “Chinese government-owned” company. According to Fortune, a Singaporean firm called Swiftdata acquired Megaspeed in November 2023, and Swiftdata’s owner, Singaporean businessman Tan Yong Pong, became sole director. In 2024 he set up Speedmatrix, a Malaysian subsidiary that received shipments of Nvidia hardware. The Times described Megaspeed as having essentially a storefront office and little local presence in Singapore, even as it became one of Nvidia’s biggest customers in Southeast Asia.

Alice Huang, also known as Huang Le, is a Chinese citizen who worked as a television reporter in China before becoming a top executive at Megaspeed. According to the Times, she had earlier been a director at a Shanghai fund linked to state-backed firms that invested in 7Road. Fortune reports that she stepped down as managing director when Swiftdata took over in late 2023. Nvidia’s general counsel later said, according to Bloomberg, that she never held that role on paper. U.S. intelligence concluded that China’s Ministry of State Security, its main spy agency, had recruited her. That is a reported intelligence assessment. What information she obtained, what she passed on, and whether any of it came from Jensen Huang are still open questions.

The amount of hardware involved has grown with each new report. In October 2025, the Times reported that Megaspeed had bought roughly $2 billion of Nvidia products. In December 2025, Bloomberg estimated from customs data that Megaspeed had imported at least $4.6 billion of equipment containing about 136,000 Nvidia chips. The Times has since reported that Megaspeed’s Malaysian affiliate bought at least $4.8 billion of hardware containing Nvidia chips between June 2024 and February 2026, which runs past the date Nvidia says it cut Megaspeed off. The Information puts the figure moved through Megaspeed and Speedmatrix at $7.5 billion at least. These numbers measure different things over different periods. All of them describe a very large customer for a company that started in 2023.

There are two ways a customer in China could benefit from hardware sold to a company in Southeast Asia. The first is physical diversion: the first buyer receives the chips, and the equipment later goes to a destination or customer that needs U.S. permission. The second is remote rental. The equipment stays in a foreign data center while the customer in China runs its work over the internet. Renting a large block of computing time can deliver much of the benefit of owning the chips for a specific project.

The legal questions differ. A physical transfer can trigger export or re-export rules. Remote rental is mostly legal today, because Commerce does not treat cloud computing services as exports. The Remote Access Security Act would close that gap. The House passed it 369 to 22 on January 12, 2026, and it has been sitting in the Senate ever since. In the meantime, Commerce issued guidance on May 31, 2026, saying that exporting advanced chips to any company headquartered in China, or whose ultimate parent is there, requires a license no matter what country that company operates in. That rule could matter a great deal if investigators conclude that a Chinese company effectively controls or finances a foreign data center.

Fortune, citing the Times, reports that Speedmatrix received more than 200 shipments of restricted Nvidia technology from June 2024 through June 2025. Fortune also cites The Information for an allegation that the chips ultimately powered data centers built in Shanghai and Shenzhen by Shanghai Century Litong, a Chinese company in which Alice Huang’s younger sister is a founding shareholder, and which sells computing capacity to Alibaba and Tencent. Fortune says Megaspeed reportedly did both things described above: it ran data centers in Johor, Malaysia, and, according to the reports, also moved chips into China. Shipment counts establish the scale of deliveries to the intermediary. To prove an unlawful transfer, investigators still need to trace the hardware to its final destination and determine what licenses applied. Megaspeed denies wrongdoing. Nvidia has said it found no evidence of illicit transfers. Singapore says it is investigating Megaspeed for suspected breaches of its domestic laws. It has separately disputed reports that it investigated Alice Huang over the espionage allegations.

This is part of a wider pattern. On November 12, Reuters reported that Chinese AI firms were rushing to rent Nvidia-powered capacity abroad because they feared tighter U.S. rules on remote access. Bloomberg reported on November 10, citing filings with Beijing regulators, that the Chinese AI company Moonshot signed a contract in August to rent 32 servers built around Nvidia’s B300, one of the most powerful AI chips on the market and one the U.S. bars from export to China. In early November, the FT reported that Oracle had canceled a $7 billion deal to give Tencent access to 100,000 AI chips in Southeast Asian data centers after U.S. officials raised concerns. Reuters also reported that Megaspeed said it had cooperated with U.S. inspections and hired a U.S. law firm to inventory its chips and review its compliance.

Who owns a company, who pays for its equipment, and who uses its computing power are three different questions. Nvidia said Megaspeed had no Chinese shareholders. In a May 13, 2026 report, the short-selling firm Culper Research cited a Malaysian corporate filing showing that in June 2024, two months before Speedmatrix started operating, it pledged its business as collateral to Apex Enterprise Solutions, a Singapore company whose parent is Alibaba. Culper also cited Apex’s Singapore financial statements showing more than $4 billion in prepayments, and it reported that Aivres, the California-based server maker formerly known as Inspur Systems, shipped more than $4 billion of products to Megaspeed and Speedmatrix. (Aivres’s former parent, Inspur, is on the U.S. Entity List.) Culper disclosed that it was short Nvidia and stood to profit if the shares fell, so its claims deserve scrutiny. But the core finding no longer rests on Culper alone. On November 9, the FT reported, citing company documents, that Alibaba lent money to Speedmatrix through Apex in June 2024. The Times has separately reported that Megaspeed rents its data-center capacity to Alibaba.

Put those pieces together and you get a striking picture. The Pentagon has formally listed Alibaba as a company that supports China’s military. Alibaba allegedly helped finance the Megaspeed subsidiary that bought billions of dollars of Nvidia hardware. Megaspeed rents computing power to Alibaba. And by May 2026, Commerce had cleared Alibaba to buy Nvidia’s H200 chips directly. The public record leaves open where every chip went and whether any export rule was broken. It already shows one arm of the U.S. government treating a company as a military risk while another treats it as an approved customer.

This is why I would put access to computing power first. If a restricted organization got substantial access, it gained something of practical value even if nobody handed over a chip design or a classified document. The relationship creates a second set of risks. Private access to a CEO can reveal which customers are getting equipment and when. It can also create openings to ask for introductions or special treatment, or give someone leverage if they threaten to expose the relationship. Those are reasons to examine communications and decisions. Public reporting has not established that Jensen Huang was blackmailed, shared secrets, or directed anyone to approve an improper shipment.

There is also a counterintelligence question worth asking. In his video, Shapiro passes along a point from Alex Kantrowitz of the Big Technology podcast. Jensen Huang was close to Trump, and his conversations with senior U.S. officials could have value to a foreign intelligence service even if they contained no classified information. A private conversation about what Washington was considering on chip sales, for example, could help Beijing anticipate a decision. Kantrowitz’s hypothetical: what if Huang left his phone out and someone read his texts with the president, or planted something on the device? That is a possibility worth investigating, not evidence that it happened. Nothing public shows that anyone read Huang’s messages or compromised a device.

What Beijing reportedly wanted

The Times reported that a classified memo, prepared this spring under the supervision of Marco Rubio (who serves as both secretary of state and national security adviser), concluded that the Ministry of State Security had recruited Alice Huang for two purposes: to gather information about American AI chips, and to push Jensen Huang to lobby Washington for looser export controls. The memo was based on electronic surveillance of Alice Huang. Intelligence officials asked for the identity of an American in the intercepted communications who appeared to be intimately involved with her, and that process, known as unmasking, named Jensen Huang.

That second goal is what turns a personal scandal into a policy scandal. Jensen Huang was the most prominent and persistent advocate for selling advanced chips to China, and he largely got his way. The U.S. reversed its H20 restrictions in July 2025, struck a deal in August 2025 that let the government take a cut of China sales, and in December 2025 announced it would allow H200 sales to approved Chinese buyers in exchange for 25 percent of the revenue. There are many reasons an American CEO might lobby for access to his biggest foreign market, and nothing public shows that Alice Huang changed his views or his arguments. As Shapiro notes, Huang’s position was self-serving, though Shapiro thinks the argument itself is right. But if U.S. intelligence believed Beijing was trying to shape that lobbying through a relationship, officials who knew had good reason to give his advice extra scrutiny. The record suggests they did the opposite.

What we know about the relationship

Jensen Huang and Alice Huang were photographed at a group gathering in Taipei in June 2024, the same month, Culper notes, that Speedmatrix pledged its business to Alibaba’s Apex. The Washington Post reported that they were seen together again in Taipei in May 2025. For more than a week after the Times first reported on the intelligence in late October, Nvidia described his contacts with her as ordinary customer meetings, with a spokesman saying none of his interactions with her were unusual or unexpected. On November 8, he acknowledged an affair in a letter to employees, without naming her. The same day, Nvidia’s board opened an independent investigation, saying it had no reason at that point to believe U.S. national security or Nvidia secrets had been compromised. The public reporting gives us dates when the two were seen together. It does not establish when the relationship became romantic, how long it lasted, or when or whether it ended. Days later, reports emerged of a second affair, this one with an Nvidia executive who reported to him, and the board met to consider removing him before he resigned.

Which Nvidia chips can be sold to China?

The rules depend on the chip, the buyer, and the date. The U.S. looks at how fast a chip can calculate, how quickly it can read data from memory, and, under some rules, how fast chips can share data with each other. Those features matter when thousands of chips work together on the same AI model.

Blackwell B200 and B300, and the coming Rubin generation. These are Nvidia’s top-tier data-center chips, designed to do more AI work in less time. They sit above the H200 tier and fall outside the January 2026 policy that opened a path for H200 sales. Exports to China face a presumption of denial, which means the U.S. normally turns down license requests. Huang lobbied for permission to sell a version of Blackwell to China. The administration has not approved it.

H100 and H200. Both are powerful data-center chips from the earlier Hopper generation. The H200 holds more data in memory and moves it faster than the H100. Both required U.S. licenses for China during the reported 2024–2025 shipment period. Under a rule that took effect January 15, 2026, the U.S. began considering H200-class and less advanced chips case by case, subject to conditions. Buyers must show security procedures and promise no military use, the chips must pass through the U.S. for testing, a 25 percent tariff applies, and sales to China are capped at half of U.S. volume. Nvidia says the government began granting it licenses in February 2026 to ship small amounts of H200s to specific China-based customers. In May, Reuters reported that Commerce had cleared about 10 Chinese companies, including Alibaba, Tencent, ByteDance, and JD.com, to buy up to 75,000 H200s each, but that Beijing had discouraged the purchases and no deliveries had happened yet. Shipments began over the summer. In its filing for the quarter that ended July 26, 2026, Nvidia said it had shipped only a fraction of what the licenses allowed. Those sales were less than 1 percent of its $89 billion in data-center revenue for the quarter, and Nvidia took a $400 million charge on unsold H200 inventory. The FT reported that ByteDance and Tencent each received about 10,000 of the chips. Researchers at the Institute for AI Policy and Strategy have pointed out that Commerce has no practical way to enforce the conditions once the chips are inside China.

H20. Nvidia designed this less powerful data-center chip for China under earlier rules. In April 2025, the U.S. began requiring a license for it too, then reversed course in July. Beijing then discouraged Chinese companies from buying it, and Nvidia says it was unable to sell its H20 inventory. Some gaming and workstation chips fall outside these controls entirely. A chip designed for the Chinese market can still become restricted when the rules change.

What this means for Megaspeed. Permission for a specific buyer to receive H200 chips in 2026 tells us little about whether an earlier shipment through Megaspeed was lawful. That depends on which chips were shipped, where they went, who received them, and what rules and licenses applied at the time.

How carefully are licensed buyers checked? In early November, the Times reported that Commerce had approved about two dozen Chinese companies to buy or handle H200 chips. One of them, New Matrix, was 17 months old, its registered office near Beijing appeared vacant, and its owners had ties to companies linked to Wise Road, a Chinese investment fund the U.S. blacklisted in 2024 for helping Beijing acquire sensitive chip businesses. Another licensee, Maginfra, also had a deserted office, and analysts tied it to Inspur, a blacklisted server maker. Two others had owners tied to firms that serve China’s military. Analysts’ judgment that New Matrix may be a front is an assessment, not an official finding, and the Times did not report that these firms received or diverted chips. But the findings bear directly on whether the licensing system can tell a legitimate buyer from a front, which is the same question Megaspeed raises.

What would putting Megaspeed on the Entity List do?

The Entity List names companies and other parties the U.S. government considers a national security or foreign policy concern. If Megaspeed were added, every supplier, not just Nvidia, would need U.S. permission to sell or transfer the covered items to it. Nvidia says it already stopped its own direct sales in June 2025. A federal listing would also reach server makers and resellers, which matters because much Nvidia hardware reaches customers inside servers built by other companies.

The list only works if it keeps up with new front companies. The FT reported that Commerce has not added a single company in more than a year, the longest gap in 18 years. Separately, an interagency panel approved adding the Chinese AI lab DeepSeek, the memory chipmaker CXMT, and more than 100 other firms, and those listings have also stalled, reportedly to avoid upsetting Beijing. When the list stops growing, procurement networks simply move to new intermediaries.

Who knew what, and when?

Three related efforts run through this timeline. U.S. Commerce investigated whether Megaspeed helped route restricted Nvidia chips to China, before the case moved to the Justice Department, which is now running a criminal inquiry. Singapore investigated possible false declarations connected to Megaspeed. And U.S. counterintelligence officials examined whether Alice Huang worked for Chinese intelligence and whether her access to Jensen Huang posed a security risk. When one agency learned something does not tell us when another agency, or the president, learned it.

The Singapore revenue that disappeared

As recently as mid-2025, Singapore was officially one of Nvidia’s biggest markets. In the quarter that ended in July 2025, the month after Nvidia says it cut off Megaspeed, Singapore-billed revenue was $10.2 billion, or 22 percent of the total, up 81 percent from a year earlier. Two years before that, it had been 8 percent. The next quarter, Nvidia changed its reporting method, and Singapore vanished. Later restated figures indicate that most of the Singapore-billed revenue was reclassified as U.S.-related.

Nvidia’s explanation at the time was that customers used Singapore as an invoicing hub while the products were almost always shipped elsewhere, and that more than 99 percent of the export-controlled data-center revenue billed to Singapore came from orders by U.S.-based customers. That claim deserves to be stated, and it may be largely true. But “U.S.-based customer” describes who placed the order, not where the chips ended up. A U.S.-based server maker can buy chips, build them into servers, and sell those servers to a data center in Malaysia. Culper reports that Aivres, a California-based server maker, shipped more than $4 billion of equipment to Megaspeed and Speedmatrix.

When Nvidia changed methods, it said the new approach gave a better picture, because billing addresses often reflected centralized invoicing. As the Wall Street Journal’s Jonathan Weil points out, a customer’s headquarters doesn’t show where chips physically go either, so it is unclear why one administrative address was better than another. The Singapore figure covered every customer billed there, so it says nothing directly about how many chips went to Megaspeed or China, and Nvidia has declined to say where Megaspeed’s sales were counted. The practical effect is clear, though. Investors lost the ability to see whether Singapore sales dropped after the Megaspeed cutoff, just as government scrutiny was intensifying. Many buyers also get Nvidia chips inside servers made by U.S. companies such as Dell and Supermicro, and whether those sales were booked as U.S. revenue, wherever the chips ended up, is another question Nvidia hasn’t answered.

When should officials have acted?

The first opportunity was in 2025, while the shipment investigation was active. Commerce had every reason to build a complete inventory, compare it with records from the companies that sold servers containing Nvidia chips, and work with authorities in the countries where the equipment was supposed to be. Bloomberg’s analysis of trade records found that Nvidia’s initial count missed about 50,000 chips sold through hardware partners. Nvidia later told Bloomberg it had checked additional warehouses that held the missing equipment, without saying how many chips it found. Two of the data-center operators Nvidia cited reportedly denied hosting Megaspeed hardware. That gap is a reason not to take a company’s inventory on trust. It also shows the limits of a company’s own sales block. Other firms can sell complete servers, which is how Megaspeed’s affiliate could keep buying Nvidia hardware into 2026. If Commerce stepped back when the case moved to Justice, it should be able to explain who took over, what evidence went with it, and whether the checks continued.

Then there is the Teter episode. In late summer 2025, Nvidia general counsel Tim Teter called Kessler, the Commerce official who oversees export controls, and asked whether the U.S. had played a role in Alice Huang’s detention in Singapore. Nvidia says Teter did not know about the affair. In the weeks afterward, Bloomberg reports, Commerce enforcement staff were told to route all Nvidia-related requests through Kessler or get his approval. Teter then emailed investigators, saying Kessler had given him permission to contact them. When one declined to meet, the staff forwarded the emails to Kessler, who told them to handle the case by the book. The case was later moved to the Justice Department, and Bloomberg reports that in the following months Nvidia gave Commerce an incomplete account of its Megaspeed business. A company under scrutiny asking a senior official about an investigation is not illegal by itself, and Kessler did tell his staff to follow the rules. But an order to route all Nvidia-related matters through a political appointee, issued weeks after that appointee took a call from Nvidia’s top lawyer, is exactly the kind of decision that needs a documented reason. So does moving the case out of Commerce.

The second opportunity was when the intelligence warning reached senior officials. If Trump was warned in late 2025, the question is what he asked his staff to do then. As far as the public record shows, the answer is nothing. Jensen Huang did not get a defensive briefing, the standard step of warning an American that a foreign intelligence service may be targeting him. That omission matters in two directions. If Huang didn’t know, a warning could have protected him and his company. If officials worried that he might be compromised, they had reason to limit what he heard. Instead, he kept his access. By May 2026, if Rubio warned Trump as reported and Lutnick was sent the memo, both had reason to ask whether the intelligence changed the risks of pending licenses, existing approvals, or bringing Huang into sensitive discussions on the China trip. A proportionate response could have included a counterintelligence briefing for Nvidia’s leadership, tighter limits on sensitive conversations during the trip, or a targeted review of licenses and the proposed listing. Officials could have had legitimate reasons to protect sources or let investigators work quietly. That makes a dated record of what they actually did especially important.

In The Bulwark’s November 12 discussion, Bill Kristol, who served as Vice President Dan Quayle’s chief of staff, said the Times story left him “mind boggled,” because a president or vice president briefed on a memo like that would not treat it as business as usual. His focus was on what Huang’s late addition to the China trip signaled. Huang was initially left off the delegation, and the Times says it is unclear whether that was connected to the memo. After press reports noted his absence, Trump personally arranged for him to board in Alaska and hailed him as “the Great Jensen Huang.” Kristol’s point is that bringing someone on a presidential trip sends a message, including to Xi Jinping. It publicly signaled confidence in Huang at a moment when he was pushing for more chip sales to China, and while the U.S. was approving H200 buyers. That decision deserves its own examination. Did staff recommend leaving him off? What meetings did he attend, and what was he told? Kristol says Trump overruled “what must have been” a National Security Council recommendation. That is his inference, and the public reporting has not confirmed that a formal recommendation existed.

In October, Trump gave Huang the National Medal of Science. Whatever Huang’s scientific achievements, the president awarded the honor nearly a year after being told that a suspected Chinese agent was close to him, and five months after a second warning.

What Singapore says Washington knew in 2025

On November 12, Singapore’s government said it had held Alice Huang’s passport from June 2025 to February 2026, and had asked the Justice Department and FBI in July 2025 to help it get cooperation from Supermicro, an American server maker and close Nvidia partner. Despite repeated requests, Singapore said it had received limited information or assistance. The Times cited two people who said Supermicro supplied Megaspeed. Supermicro says it never sold products to Megaspeed and has complied with U.S. export rules.

The overlap is the troubling part. Singapore was still holding Alice Huang’s passport for months after Trump was reportedly briefed that she was a Chinese intelligence asset. A close partner had the suspected agent in hand, was asking Washington for help, and says little came back. The public record does not tell us why. There may have been a reason to stay quiet, such as protecting a source or an ongoing operation. But a close partner publicly complaining about U.S. cooperation in a spy case is rare, and it suggests the gap was real.

The suspected honeypot and the White House response

A “honeypot” is an intelligence operation that uses a romantic or sexual relationship to gain access, information, or influence. Both Zeihan and Shapiro note that this is an unusual method for China. Its intelligence services are better known for spreading a very wide net, collecting huge amounts of scattered information from many people and piecing it together later. Zeihan calls a targeted honeypot more of a classic Russian technique. If the allegations hold up, this would be an unusually focused operation against one of the most important people in the global technology industry. Investigators have reason to ask whether Alice Huang used the relationship to learn about Nvidia’s plans or U.S. policy, to influence decisions about Megaspeed, or to shape Huang’s lobbying. How the relationship began, whether it affected any chip order or license, and what information may have reached Beijing are still unresolved.

The sharpest question for Trump is what he did after being warned twice. Huang still joined the delegation. Asked this month, Trump said, “No, I didn’t hear about it then,” while allowing that he might have been shown intelligence he didn’t recall. A White House official said the Times’s sources “have no idea what they’re talking about,” without explaining further. Asked whether he wished he had acted on the intelligence, Trump called Huang “an honorable guy” who “probably has an excuse.” According to the Times, people familiar with the matter said that when Trump was first briefed he was wary of disrupting Nvidia’s chip sales or complicating trade with China. As of November 12, Commerce had not confirmed whether Lutnick personally saw the memo Axios says was sent to him.

The reported warnings make it reasonable to ask whether either man set aside a known risk. If the relationship gave Beijing access to sensitive U.S. deliberations, or helped restricted users get advanced computing power, either would be a significant national security breach. Even before the full damage is known, a credible warning about the CEO of the world’s most strategically important tech company called for a serious response. The public record gives Congress solid grounds to examine how the administration handled it.

Does it matter who leaked it?

Shapiro argues that the most important question is who made all this public. His answer is someone inside the U.S. national security establishment who opposed Trump’s softer line on China and wanted to discredit the most prominent voice against export controls. China would have wanted to keep a compromising relationship secret and use it as leverage, not expose it, so he thinks Beijing is an unlikely source. He doubts that Huang himself, or a rival company, was behind it. Kristol sees the leaks differently: as officials getting their own record on paper, a way of saying they tried to stop this. Both may be partly right, and the motive is a fair question. Leaks are selective, and people who leak usually have a point of view.

But the motive of a leaker does not make the underlying facts false. And a large part of this story no longer depends on anonymous sources. Jensen Huang has admitted the affair. Nvidia’s board opened an investigation, and Huang resigned. Singapore’s government has confirmed on the record that it held Alice Huang’s passport for eight months and that the U.S. gave it little help. Senator Cotton’s 2025 letter to Lutnick is public. Nvidia’s own SEC filings show the Singapore revenue and the reporting change. The Alibaba financing is documented in corporate filings. The Pentagon’s listing of Alibaba is official. If anything, the leaks raise one more question for investigators: if officials inside the government were alarmed enough to go to reporters, what did they try inside the government first, and who overruled them?

Why did Lutnick resist tighter enforcement?

Axios reported on November 11 that Trump approved action against suspected chip diversion three times this year, in May, in August, and on November 5. Its sources said Lutnick and Kessler resisted moving ahead each time. In the November 5 Situation Room meeting, Bessent confronted Lutnick about his department’s resistance. Lutnick appeared to agree but said he needed to check with Trump. The group moved to the Oval Office, where Trump, after consulting Elon Musk, reportedly said, “Let’s get it done.” Axios says Lutnick proposed a new delay once everyone was outside. Lutnick reportedly warned that the measures could set off a trade war with China, and Commerce has denied the account of the confrontation.

Taken alone, a cabinet secretary worried about trade retaliation is having a policy disagreement. What makes this more serious is the pattern around it: an Entity List frozen for more than a year, stalled listings for DeepSeek and more than 100 other firms, Megaspeed left off despite requests from four agencies, H200 licenses for buyers with empty offices and blacklist ties, and an enforcement chief who took a call from Nvidia’s lawyer. Each piece may have an explanation. Together they show an export-control office that repeatedly sided with selling over enforcing, even when the president, by Axios’s account, told it to act. On November 12, Reuters reported that senior White House officials were discussing whether Lutnick should be pushed out. Axios reported the same day that the White House was still publicly standing by him.

Kristol finds it inconceivable that Lutnick would have resisted without talking to Trump, given how long and how closely the two men have been allied. That reading would put the responsibility back on the president. The Axios account cuts the other way, with Trump repeatedly authorizing action and Lutnick slowing it down. Both can’t be fully true, and which one is closer to the truth is among the most important things for investigators to settle. If Trump authorized a crackdown and his Commerce secretary stalled it three times, that is a serious failure of command. If Trump publicly authorized it while privately signaling otherwise, that is worse.

Those enforcement decisions concern suspected diversion and should be examined separately from the administration’s policy of allowing some H200 sales under license. Congress should ask: what did Trump authorize on each occasion, who was supposed to carry it out, and what held it up?

Could Lutnick or his family have benefited? Reuters reported that Lutnick agreed to give up the economic benefits of his interests in Cantor Fitzgerald, BGC, and Newmark as of May 16, 2025, and that the Cantor ownership transfer included trusts for his adult children. His sons Brandon and Kyle now run Cantor. A July 2025 ethics waiver still barred Lutnick from taking part in specific matters in which any of those firms was a party or represented one. The Lutnick and Trump families also have business ties of their own. In June, the Times reported that after Lutnick and Trump negotiated a deal that could bring up to $1.6 billion in federal financing for a Kazakhstan tungsten project, investors linked to a firm partly owned by Donald Trump Jr. and Eric Trump joined partners to take a 20 percent stake in a related entity, and Cantor helped raise $210 million for one of the deal’s lead investors. That deal has nothing to do with Nvidia, and the Trump sons say they were passive investors. But it helps explain why Lutnick and Trump might be reluctant to turn on each other, a point Andrew Egger raised on The Bulwark. The available reporting does not show whether Cantor or any of Lutnick’s former firms had a financial stake or client role in an Nvidia- or Megaspeed-related transaction. The records that could answer that are deal-specific financing or advisory agreements, any ethics reviews or recusals, and Commerce’s written reasons for delaying or declining action. Concern about trade retaliation could explain a policy disagreement. A deal-specific financial tie would raise a different question: whose interests the decision served.

The money around the decisions

Nvidia’s ballroom contribution and Trump’s Nvidia stock trades are documented. The Washington Post reported that Nvidia committed $25 million to Trump’s White House ballroom project in October 2025, nearly three months after the summer 2025 deal that let Nvidia resume China sales with the government taking a cut of the revenue. Twenty-five million dollars was the largest amount any donor gave, tied with Apple and two others. The money went through a trust, and the White House did not disclose amounts at the time. Nvidia called the donation independent of its business interests. Watchdog groups say the secrecy makes it impossible to tell whether donors with business before the government are buying access.

Trump’s financial disclosure shows that his portfolio bought up to $1.6 million and sold up to $1 million of Nvidia stock in August 2026. The disclosure became public on October 8, the day he gave Huang the National Medal of Science. The filing reports ranges, so his exact position is unclear. The White House told Reuters that outside managers make investment decisions independently and that the portfolio tracks recognized indexes such as the Schwab 1000. A president’s financial exposure to a company affected by his administration’s decisions is relevant to any conflict-of-interest question. Nothing public proves that the holdings or the ballroom contribution influenced any decision.

In his November 12 discussion, Zeihan suggests that money and personal access may have influenced Trump and Lutnick. The material reviewed for this report does not show that a ballroom contribution bought access or a license, that Lutnick took a bribe, or that either man stood to gain financially from a Megaspeed decision. The distinction matters, because the evidence of ignored warnings and stalled enforcement is much stronger than the evidence of a corrupt bargain.

Negligence, recklessness, or something worse?

These words have specific meanings, and the evidence supports them to different degrees.

Negligence means failing to take reasonable care against a known risk. The public record makes a strong case here. Officials knew by late 2024 that Commerce suspected Megaspeed. They knew by late 2025 that intelligence tied a Megaspeed executive to Chinese intelligence and to Nvidia’s CEO. Senator Cotton warned Lutnick in writing. Four agencies asked for a listing. Singapore asked for help. As far as the public can see, the response was to move the case, leave the company off the blacklist, skip the defensive briefing, and keep loosening chip sales.

Recklessness means knowingly disregarding a serious risk. The case for recklessness rests on the May 2026 decisions. By then the president had reportedly been warned twice, a classified memo was circulating, and the administration still put Huang on the trip to Beijing, the same week Reuters reported that Commerce had approved Chinese H200 buyers. Whether that crosses the line depends on what officials actually knew and believed, which only the internal records can show.

Self-dealing means officials putting their own interests ahead of their duties. Here the evidence is thinnest. There are documented donations, stock trades, and business ties between the president’s and the commerce secretary’s families, and there is an undeniable pattern of decisions that favored Nvidia’s sales. But no public evidence connects any specific Nvidia or Megaspeed decision to personal gain. I’d treat this as the question investigators most need to answer, and not as something established.

The strongest criticism I think the evidence supports today is that Commerce and the White House owe a full account of what they did at each decision point. If officials had credible evidence of ongoing diversion, the authority to address it, and no sound reason to wait, the case for a preventable failure gets stronger. A referral to Justice that preserved the evidence and kept an active investigation going would change how the same timeline reads. A political decision to put trade negotiations ahead of enforcement would raise a different question than a decision based on thin evidence.

A criminal case would need evidence of a specific offense and the state of mind the relevant law requires. For an illegal export, investigators would need the chip model, the shipment and license records, the actual buyer and destination, and evidence of what the people arranging the deal knew. A claim of deliberate interference would need evidence of who directed what, why, and which legal duty or proceeding was affected. A company’s contact with regulators, an official’s delay, or the timing of a change in financial reporting can justify scrutiny. Each still leaves real questions about intent and legality.

What to watch next

Nvidia reports quarterly results on November 17, Kress’s first call with investors as interim CEO. The House Republican chairs had requested a closed-door briefing with Huang on November 18, and it is unclear whether that will go ahead now that he has resigned. The bigger change comes in January, when Democrats take control of both the House and the Senate after their November 3 midterm victories, along with the power to subpoena documents and witnesses. Members of both parties have already pressed Lutnick for answers about what he and Kessler knew. Kristol calls this a classic case for a special counsel, because the administration can’t credibly investigate itself, but he doubts this Justice Department will appoint one. The Justice Department’s criminal inquiry into Megaspeed is ongoing, as is Singapore’s. The Remote Access Security Act is still waiting in the Senate. And the most telling near-term signal will be whether Commerce finally adds Megaspeed to the Entity List, and whether Lutnick is still in charge when it decides.

The records that could settle much of this include the dated intelligence briefings and who received them, the instructions that followed, the Commerce-to-Justice referral and case files, the license files for approved Chinese buyers, and the shipping and data-center records showing where the chips went and who used them. Until those come out, my judgment is that the national security risk is real and the evidence of actual damage is incomplete. If restricted Chinese customers used those chips to speed up their work, the U.S. may already have lost some of the time its controls were meant to buy. That is the consequence worth investigating, even after Jensen Huang’s departure.

Sources

Links go to the original source wherever one was available; some require a subscription. Where the original could not be linked, the link goes to a public summary or republication and is marked as such. 

Government and company records

• U.S. Commerce Department, Industry Guidance to Prevent Diversion of Advanced Computing Integrated Circuits, May 13, 2025
• Federal Register, Revision to License Review Policy for Advanced Computing Commodities, January 15, 2026
• U.S. Congress, Remote Access Security Act, H.R. 2683: as passed by the House, January 12, 2026; as referred to the Senate, January 13, 2026
• Nvidia Form 10-Q, quarter ended April 27, 2025 (Singapore invoicing and “over 99%” disclosure); quarter ended July 27, 2025 (Singapore revenue); quarter ended October 26, 2025 (change to headquarters-based geographic reporting)
• Nvidia Form 10-Q, quarter ended April 26, 2026 (H200 licenses beginning February 2026); quarter ended July 26, 2026 (first H200 shipments, $0.4 billion charge); CNBC on the Q2 fiscal 2027 results
• Nvidia, NVIDIA Announces CEO Transition, November 12, 2026; Form 8-K
• Trump’s periodic transaction report, signed September 17, 2026 (August 2026 Nvidia trades)
• Senator Tom Cotton, letter to Secretary Lutnick, October 13, 2025; press release, October 14, 2025
• Senators Elizabeth Warren and Chris Coons, records preservation letter to the White House, Commerce and State, November 11, 2026
• Senator Michael Bennet, Bennet presses Trump administration on failure to stop Megaspeed’s apparent AI tech diversion to China, November 11, 2026
• House Foreign Affairs Committee Democrats, letter to Commerce on AI chip exports, November 2026
• House Select Committee on the CCP Democrats, letter to Secretary Lutnick on export controls as a bargaining chip, July 27, 2025
• U.S. Department of Defense, Section 1260H List of Chinese Military Companies, June 8, 2026 — WilmerHale summary; TechCrunch
• U.S. Commerce Department, guidance on exports to China-headquartered companies, May 31, 2026
• Reps. Brian Mast, John Moolenaar and Rick Crawford, letter to Jensen Huang requesting a closed-door briefing, November 4, 2026
• The White House, Lutnick limited ethics waiver, July 8, 2025
• U.S. Cybersecurity and Infrastructure Security Agency, PRC State-Sponsored Actors Compromise and Maintain Persistent Access to U.S. Critical Infrastructure, February 7, 2024
• FBI and CISA, Joint Statement on People’s Republic of China Targeting of Commercial Telecommunications Infrastructure, November 13, 2024

The New York Times

• A Mystery C.E.O. and Billions in Sales: Is China Buying Banned Nvidia Chips?, October 9, 2025 (listed October 10 in Techmeme’s contemporaneous index)
• Nvidia Chief Seen as Target of Chinese Spy Plot, October 29, 2026; see also Ars Technica’s coverage, October 29, 2026
• Report on China’s spy agency, Alice Huang and Jensen Huang, November 6, 2026
• Nvidia board investigation, November 8, 2026; Rubio Warned Officials About Nvidia Chief’s Entanglement With Chinese Spy, November 8, 2026 — republished text
• Commerce’s Megaspeed inquiry and Nvidia’s intervention, November 9, 2026
• Rubio’s warning to Trump on the China trip, November 10, 2026
• H200 licensees with empty offices and blacklist ties, early November 2026 — Jerusalem Post summary
• Trump and Lutnick sons tied to Kazakhstan tungsten deal, June 2026 — Boston Globe republication
• Paul Mozur, Singapore Says U.S. Not Helpful on Investigation Into Suspected Spy’s Company (shared, unlocked link), November 12, 2026

Other news reporting

• Bloomberg, Nvidia Intervened in US Probe of Megaspeed Chip Smuggling, November 9, 2026 (including the inventory discrepancy); excerpt via Daring Fireball; Bloomberg, Megaspeed import analysis, December 2025
• Bloomberg, Moonshot AI rents Nvidia B300 servers, November 10, 2026 — summary
• Financial Times, U.S. Commerce Department delays Megaspeed’s inclusion on trade blacklist, November 10, 2026; Alibaba–Apex loan to Speedmatrix, November 9, 2026; Oracle–Tencent deal canceled, November 3, 2026; H200 shipments to ByteDance and Tencent, August 2026 — TIKR summary
• The Next Web, US holds off blacklisting DeepSeek and 100-plus Chinese firms
• CNN, Despite denial, Trump was briefed on spy concerns related to Nvidia CEO during trip to China (republished by CP24), November 10, 2026
• Axios, Lutnick was sent a classified warning, November 10, 2026
• Axios, Lutnick stalled Trump’s Nvidia–China chip curbs after Oval talks, November 11, 2026
• Axios, Bessent–Lutnick exchange over Megaspeed and Commerce’s denial, November 11, 2026
• Axios, Trump sticks by Commerce Secretary Lutnick despite Nvidia inaction, November 12, 2026
• Reuters, Senator seeks Commerce Department investigation into Megaspeed, October 14, 2025
• Reuters, U.S. clears H200 chip sales to 10 China firms, May 14, 2026
• Reuters, Huawei says AI chip demand outstrips supply (republished by The Edge), September 17, 2026
• Reuters, Trump discloses Nvidia stock trades as he prepares to honor CEO Huang (republished by GV Wire), October 8, 2026
• Reuters, Lutnick stalled Trump crackdown on diversion of Nvidia chips to China, Axios reports, November 11, 2026
• Reuters, Chinese AI firms rush to secure Nvidia chip capacity as spy scandal stirs fear of U.S. action, November 12, 2026
• Reuters, Nvidia customer Megaspeed says it cooperated with U.S. inspections, November 12, 2026
• Reuters, Nvidia CEO Jensen Huang to step down, November 12, 2026; Fortune on the CEO transition
• Reuters/The Globe and Mail, Lutnick’s White House future in doubt over AI chip shipments to China, November 12, 2026
• Reuters, Lutnick transfers Cantor Fitzgerald stakes to children, May 19, 2025
• Export Compliance Daily, U.S. to lift export curbs on Nvidia’s H20 chips to China (Bessent “negotiating chip”), July 16, 2025
• Semafor, Commerce to open up exports of Nvidia H200 chips to China (Greg Allen testimony; Tencent’s Wang Qi), December 8, 2025
• Investor’s Business Daily, U.S., China Are In AI Infrastructure Booms. But They’re Running Different Races, October 30, 2026
• Associated Press, Nvidia investigates whether CEO was targeted by Chinese spy, November 8–9, 2026
• NPR, Democrats win full control of Congress, November 4, 2026; NBC News
• Fortune, How Megaspeed built a network in Southeast Asia to move Nvidia chips to Chinese companies, November 11, 2026
• The Wall Street Journal, Jonathan Weil, Nvidia’s Disappearing Singapore Sales Deepen Mystery Around Megaspeed, November 11, 2026
• The Washington Post, Secret documents reveal donors behind $330 million for Trump’s ballroom, October 15, 2026
• David A. Graham, The Atlantic, Trump’s Stance on Jensen Huang Is a Tell, November 11, 2026

Research, analysis and commentary

• Culper Research, Nvidia (NVDA): The China Problem, May 13, 2026. Culper disclosed a short position in Nvidia.
• Institute for AI Policy and Strategy, New BIS Licensing Policy for H200s: Tough Guidelines, Weak Enforcement, January 2026
• Jacob Shapiro, Why the Deep State Leaked the Nvidia CEO’s Spy Scandal, November 10, 2026 (includes a point attributed to Alex Kantrowitz’s Big Technology podcast)
• The Bulwark, Morning Shots Live with Andrew Egger and Bill Kristol, November 12, 2026
• Peter Zeihan, Nvidia’s CEO Gets Caught in the Honeypot, November 12, 2026