黄仁勋卸任英伟达CEO后，公司走向与出口管制变局

Jensen Huang is no longer the CEO of Nvidia. Nvidia’s board was investigating Jensen’s conduct and weighing his future at the company he founded, and clearly the conduct issue is real; it is ethically problematic to have relationships with a customer who is a Chinese CCP asset and with multiple subordinates.

Additionally, we believe there is more at play with his resignatiom after only four days. We believe it is directly tied to the vast smuggling of GPUs. Our Accelerator Model tracks every AI chip’s production and destination, and our data shows more than 1.5GW of Nvidia chips, which are restricted, still ultimately ended up in China. This figure doesn’t even include chips that Nvidia created to circumvent US rules, a story we originally broke.

We believe the US intelligence community is a large part of the push to get Jensen out of Nvidia. Specifically, we believe it is likely that RASA or similar legislation, which has already passed the house, will move forward. AI and chip restrictions on China are going to ramp now much faster with the biggest defender Jensen, gone. We are strong believers in some restrictions, but poorly shaped ones are damaging to US interests.

This sort of legislation will have consequences, and we don’t think politicians are thinking them through.

Here are the clear benefits of the current regime where Chinese labs are able to access GPUs remotely, and NOT in China:

Profits flow to American companies.
The US can monitor what Chinese labs are doing.
The US can cut off their access to the GPUs through politics or by force almost instantly, if they choose.
American companies such as NVIDIA, AMD, Broadcom, Google, Amazon, Microsoft, Meta, Supermicro, and Dell all depend on Chinese companies in their supply chain.
American companies such as NVIDIA and AMD employ Chinese engineers (who live in China, to be clear) to develop hardware and low-level software for their GPU systems. Without these engineers, we are certain that the user experience for these GPUs would be materially worse than it is today.
American companies, especially the entire startup ecosystem benefit from the open source research that Chinese companies, students and independent researchers publish using these GPUs today.
China is (even slightly) less motivated to use their domestic accelerators.
Innovations from China flow globally through open innovations of wide expert parallelism, OSS kernel libraries: MSA, DeepEP, DeepGEMM, etc.

With that said we do believe in some AI restrictions to China, but those are more shaped like the MATCH Act then RASA. Biden style diffusion rules make the most sense.

Maybe biggest of all: open source is under attack. It’s an open secret that Nvidia works directly with the Chinese labs on open source and helping them optimize GPUs and models. These labs utilize Blackwell GPUs, which Nvidia knows. The US open source ecosystem and many startups rely on Chinese AI models. This is significantly impaired with Jensen gone and the China hawks winning.

But the question that matters for shareholders (and the AI supply chain) is a different one: what is Nvidia without Jensen?

Our view is that the company was never built for him not to be there. He was at the helm for 33 years, and even though the board pressured him out (or he “resigned”) Nvidia would still be materially better off with him there instead of gone. The company was uniquely built to suit the man himself. The company without the man might be an entirely different beast. SpaceX and Tesla don't function without Elon, Nvidia doesn't function without Jensen.

One King, 60 Direct Reports

Nvidia’s corporate structure is odd for a company of its size and scale. Jensen has roughly 60 direct reports, according to his public statements. He has said that CEOs should have the most direct reports because those people require the least amount of management and that he doesn’t have one-on-one meetings with them.

But few of them can do things without Jensen. The company is run like a kingdom with one king, with everyone else, regardless of seniority, a peon doing his bidding. Jensen has said that it doesn’t matter who reports to who as their title isn’t related to how important their job is, and it’s one of the reasons why Nvidia is the “smallest large company in the word” and has so much fewer employees – about 42,000 – despite its big market cap. At Nvidia, there is less politics as a result. Jensen has said this flat network empowers employees and allows the company to adapt more quickly because its technology is also moving so quickly.

This is one of the reasons Nvidia moves fast. Once Jensen set a direction, everyone moved on it together. Former employees have told us that product development and engineering often happen until the last minute so the company can change its mind if it needs to. For example, the GB200 design came together a lot closer to the product cycle than probably anyone would have liked, one former executive told us. Jensen has said he wishes he could keep the company in startup mode forever.

The cost of that structure, however, is that Nvidia doesn’t have a deep bench of successors. Look at Nvidia’s website: only five others, two of whom are legal and finance executives, are featured prominently among his reports. Most of his direct reports aren’t trained for the spotlight and are actively discouraged from seeking it. Few leave to become CEOs of other major companies, with Rene Haas at Arm the rare exception. Few at Nvidia get the chance to shine externally because Jensen makes all the big decisions. But much of Nvidia’s executive team doesn’t mind because heavy is the head that wears the crown, and many are happy to stay until retirement age because they’re well paid, treated well, and believe in the mission and their leader.

No One Wants the Job

Colette Kress has been named interim CEO, but Nvidia is looking for a permanent CEO. That cannot be someone with finance background. The obvious internal names being floated to succeed Jensen on the technical side are Ian Buck and Jonah Alben. Buck, who created CUDA out of his doctoral research at Stanford, joined Nvidia in 2004 and now runs the datacenter business as vice president of hyperscale and high-performance computing. He is often placed in charge of things while Jensen travels and frequently speaks at conferences.

Alben, who has been at Nvidia since 1997, is senior vice president of GPU engineering and oversees the design of the company’s GPUs. But realistically, no one person can run it all. We don’t think Buck is capable of running the technical hardware engineering side of the house. Alben and others on his team have technical autonomy, but even he doesn’t control the entire chip side, where other senior leaders who handle many major technical and business divisions also report directly to Jensen. Both are amazing but likely not able to exert the supply chain mastery Jensen has.

There is also the question of whether anyone wants to do the job of CEO. An executive told us earlier this year that “Ian would rather die than succeed Jensen and wouldn’t want that job. Jonah is more respected and people say he could, but again, no one would really want that job.” Circumstances, though, have obviously changed. Kress is the interim CEO, but wrangling more than 60 VPs to a cohesive vision of the company takes a certain kind of person, and otherwise the firm’s direction risks dissolving into noise.

Founder Bets Do Not Survive Professional Managers

Jensen is one of the last old-school tech founders, of the same generation as Elon Musk and a bit older than Mark Zuckerberg. The other companies that came up alongside Nvidia have all transitioned to professional managers, including AMD, Google, Microsoft, Apple and Amazon.

Nvidia’s success relies on big bets that can take years, if not a decade, to pay off. A founder with the full support of the board can take those risks without worrying about a revolt. CUDA lost money for years while Wall Street analysts kept questioning it, but Jensen held steady and the bet paid off enormously. As one former employee put it: “Jensen put the whole company behind programmable shaders, CUDA and tensor cores despite a lot of opposition to spending transistors on them when use was limited.”

Omniverse and autonomous driving are the new bets within Nvidia. Internally, Jensen calls them “zero billion dollar bets,” former employees tell us. They make no money today, but he sees them as just as big as CUDA. He also built the Shield TV consumer entertainment box and the DGX Cloud service on the idea that Nvidia needs to understand its customers to design better products, even if those two businesses made little or no money and were ultimately cast aside.

An Nvidia successor who isn’t the founder, no matter how long they have been at the company, answers to the board and shareholders in a different way. Long-term bets and convictions of this kind become much less likely, and one of the biggest issues is Jensen’s legendary kingmaking and dealmaking, which is going to be hard to replicate.

Customers Bought Jensen, Not Just GPUs

Jensen was the chief decider on chip allocation to the hyperscalers and everyone else. Customers are always jockeying for allocation and fighting for Nvidia’s attention, and more and more of Nvidia’s deals require extensive Nvidia support. As we wrote in July, Nvidia has a choice over whether to allocate more capacity to hyperscalers, neoclouds or neoclouds on its backstop program, where it credit-supports the cluster and takes a share of the cloud revenue. This central bank of Nvidia was backed by the full faith and credit of Jensen. The next CEO is unlikely to be able to make the bold bets Jensen did.

Nvidia likes to decide based on what’s feasible, but much of it comes down to what makes strategic sense. It doesn’t always make strategic sense to skip the hyperscalers, or to work with them and treat them preferentially. Many of these customers now need reassurances that they will continue to be supported by Nvidia through financial guarantees or allocation. Nvidia disclosed $530B of gross off-balance-sheet guarantees as of its fiscal second quarter, up from $184B the prior quarter, including $36B of take-or-pay floors under neocloud capacity, and it holds a $99B equity book.

Many customers also want a deal with Jensen for reasons that go beyond the numbers. Neoclouds around the world raised their money because Jensen backed them, not because they compared chips and chose Nvidia. Nvidia’s backstops let neoclouds and neolabs raise money without the hyperscalers: it sells GPUs at full margin on day one, then floors the buyer’s revenue, guarantees the buyer’s landlord or signs the lease itself, so that operators without an investment-grade rating can borrow at investment-grade pricing. If the Jensen put is not there, there might be a large contraction in this ecosystem.

The Supply Chain Stranglehold Loosens

Nvidia has a stranglehold on its suppliers. If competitors had access to supply, they would ship much higher volume. They can’t get it because Jensen has personally locked up memory in South Korea and chips in Taiwan. As we wrote in The Great AI Silicon Shortage, Nvidia has locked in most of the supply for logic wafers, memory and various other necessary components, and Jensen’s 2025 trip to Korea was to secure memory supply, before other executives quite realized how in shortage memory was. Our industry clients thought our forecasts for 2026 were too aggressive, but then they realized we were right. Nvidia’s supply and capacity commitments jumped from $119B to $279B in a single quarter, primarily for memory.

That stranglehold is personal. Jensen was the one who won these deals, CEO to CEO, with the major supply chain companies around the world. In Korea, we estimate SK Hynix’s 2027 HBM4 pricing could range from roughly $3.0 to $3.3 per Gb for Nvidia versus $3.7 to $4.1 per Gb for other Tier 1 customers including AMD, Broadcom, Amazon and Google, and we believe there is an implicit understanding that Nvidia gives SK Group priority GPU allocation in return. Without him, these firms will no longer prioritize Nvidia to the same degree, and capacity will be easier for others to acquire. We think it’s likely that in 2028 more than 70% of total capacity could be non-Nvidia.

What’s more, beyond strangling the supply, Jensen’s vision is what pushed the supply chain in Taiwan to invest harder and believe more in this buildout than anyone else. Jensen’s legendary supply chain dinners and commitments are what pushed suppliers in Taiwan to deliver on technologically and logistically aggressive plans.

The AI Industry Loses Its Open Source Spokesman

Without Jensen, the AI industry loses its chief spokesman. He was everywhere, speaking on AI as its chief proponent. He was the face of AI, and no one can replace that.

Jensen also set Nvidia’s China policy and led its push for open source AI models, which give countries, enterprises and neoclouds a reason to buy GPUs beyond the handful of frontier labs, who are now building their own chips.

As we wrote in September, no one needs open source AI more than Nvidia: it is on track for a world where it has only two real customers, Anthropic and OpenAI, or seven including the hyperscalers, and all seven except SpaceX are trying to build their own XPUs to cannibalize Nvidia’s margins. AI models and applications are the complements to its GPUs.

It is an open secret that Chinese frontier labs use Blackwell GPUs & NVIDIA DevTech collaborates closely with the Chinese frontier labs & learns a lot from their end users, which results in innovations to the CUDA ecosystem like wide expert parallelism, DeepGEMM, DeepEP etc. These then spread to the western open source community and startups. These startups are massively negatively impacted by the departure.

Without Jensen being the spokesperson and offering financial and engineering help for open models, Chinese labs may not be able to open source kernel libraries like MiniMax MSA, DeepGEMM, DeepEP, Moonshot’s FlashKDA, which are clearly targeted at NVIDIA Blackwell GPUs. Many production engines used widely across the world like SGLang & vLLM use these innovations from Chinese labs. Chinese labs may have to go back to straight up denying that they have any NVIDIA GPUs and keep their NVIDIA software internal only instead of open sourcing Blackwell software.

This is fundamentally bad that innovations from China won’t be shared across the NVIDIA ecosystem & the general ML ecosystem anymore. It is positive for closed AI labs but bad for startups.

A manager lead company could depriortize underperforming Nemotron, With many models going closed source at the frontier labs in the U.S., Nvidia was single-handedly propping up the other American open source models with its efforts by investing in firms like Reflection, Thinking Machines, Poolside, and the Hugging Face acquisition.

Nvidia was the one pushing continuous engagement with China, on the argument that China should use and rely on American technology. As we wrote last year, after the U.S. government banned the H20 from going into China and then indicated Nvidia would be granted a license to export it, the U.S. needs to balance having China on the U.S. AI stack and slowing the development of its own, while also limiting the quality and volumes shipped into China. Now, what if he was pushing it from a compromised place? That is up to the investigation to figure out. Clearly the China hawks are winning, so we have to ask a bit of an odd question. Why now?

Why Now Does Jensen’s Personal Life Come to Light?

We aren’t ones for conspiracy theories, but a few events around the same time seem to create a narrative that is hard to ignore. The intelligence community had much to gain. The second New York Times story on the topic explicitly states (five times) that the basis for its reporting is the U.S. intelligence community. Trump, Rubio and Lutnick have known about all of this since before their trip to China in May. What’s more is that there is reporting that Lutnick stalled the Chinese chip crackdown. Something has shifted in that relationship, and now the wind is blowing in the other direction.

We actually don’t believe Alice Huang is a spy in the traditional sense. She is a well connected business woman. She like any other well connected business person in China, has contact with the CCP.

Our Accelerator Model has tracked over 1.5GW of Nvidia GPUs that shouldn’t be in China, The Department of Commerce now also finally believes that there is more than 1.5GW of NVIDIA chips in China that shouldn’t be there. Now, with the midterms swinging the Democrat’s way, someone has decided it is time to put the hammer down both to stop the willing circumvention of the existing export controls and to expand them to cut off China’s access to frontier chips once and for all.

To be clear, the Trump administration is split on the issue, but Trump was giving Lutnick cover. Trump is no longer providing that cover. What’s more, there is a critical bill called the Remote Access Security Act (“RASA”) that has passed the House that has yet to receive a vote in the Senate, and that vote is happening right now.

RASA Up for a Vote

RASA passed the House 369-22 on Jan. 12. It passed on a motion to suspend the rules, which requires a two-thirds majority. The House Foreign Affairs Committee approved it 51-0 in April 2025. Though it was referred to the Senate Banking Committee, it hasn’t yet received a vote. We know that NVIDIA and Oracle have been lobbying against it, hard.

Of course, RASA isn’t the only way that the U.S. government can stop Chinese companies from accessing GPUs remotely. But RASA is the bill that is the hot topic of discussion in Washington right now. It is clearly popular, having already passed the House, and the substantive portions of the bill seem downright obvious to people. Why allow Chinese companies to access GPUs remotely that they aren’t allowed to import?

For what it’s worth, we are aware of neoclouds in Asia, Australia, the Middle East, Nordics, Europe, Canada and even the U.S. that are actively selling capacity to Chinese companies. All of them would be impacted by this bill. This is another global export control bill, and one with far-reaching implications. Jensen, and likely Lutnick, did not want this to happen.

So let’s go against the popular narrative and explore the possibility that despite all of the recent headlines, passing RASA (or something like it) will have unintended consequences. To follow the old adage about American politics, we believe that if all the politicians agree on something, it’s probably a bad idea.

As we stated earlier, here are the clear benefits of the current regime where Chinese labs are able to access GPUs remotely:

Profits flow to American companies.
The US can monitor what Chinese labs are doing.
The US can cut off their access to the GPUs through politics or by force almost instantly, if they choose.
American companies such as NVIDIA, AMD, Broadcom, Google, Amazon, Microsoft, Meta, Supermicro, and Dell all depend on Chinese companies in their supply chain.
American companies such as NVIDIA and AMD employ Chinese engineers (who live in China, to be clear) to develop hardware and low-level software for their GPU systems. Without these engineers, we are certain that the user experience for these GPUs would be materially worse than it is today.
American companies benefit from the open source research that Chinese companies, students and independent researchers publish using these GPUs today.
China is (even slightly) less motivated to use their domestic accelerators.
Innovations from China flow globally through open innovations of wide expert parallelism, OSS kernel libraries: MSA, DeepEP, DeepGEMM, etc.

To be clear: in a world where Chinese labs are completely cut off from accessing NVIDIA GPUs, they would all be even more motivated to move off of NVIDIA GPUs and onto domestic accelerators.

Which is already happening today.

In February, Z.ai claimed inference support for GLM-5 on seven domestic accelerators
In April, DeepSeek published their V4 paper and later open-sourced an inference megakernel for the V4 Pro model on Huawei Ascend NPUs.
In June, Meituan claimed LongCat 2.0, which claimed that it was pretrained entirely on Huawei Ascend 910B NPUs.

Source: GLM-5 Paper from Z.ai

Politicians continue to miss how intertwined the China <> Taiwan <> US relationship actually is. Many Chinese nationals work at the frontier labs. Many Chinese engineers are employed by NVIDIA and AMD, writing low level code to bring up new chips. Many Americans learn from Chinese open source contributions. Many of the raw materials needed to produce these chips come from China. Many of the L1-6 ODM board and component suppliers are Chinese.

Escalating tensions will guarantee that less chips are produced, and the pace of AI progress slows for everyone.

We believe the US should limit or control how much comprue is delivered to Chinese labs, and banning export to China while allowing remote access by trusted clouds is a the best method to do this.

Biden style diffusion rules make the most sense.

Export Control Violations and Enforcement

For all those who say that there is no way GPUs can be smuggled into China and advertised for rental on rednote because they’re too heavy to fit in a backpack, we disagree.

Source: ads on Rednote for an 800 GPU cluster of RTX 5090, and another for a 1,024 GPU minimum purchase in China.

“We’ve studied all their deployments of our GPUs and counted every one. There’s no evidence of any AI chip diversion. These are massive systems. The Grace Blackwell system is nearly two tons, and so you’re not going to be putting that in your pocket or your backpack anytime soon. The important thing is that the countries and the companies that we sell to recognize that diversion is not allowed and everybody would like to continue to buy Nvidia technology.” - Jensen Huang at Computex in May 2025

Since then, we have at least two high profile counter examples: $2.5B worth of servers being smuggled by 3x people at Supermicro, and $65M worth being smuggled by 9x people at NVIDIA, Supermicro, Albatron, Flying Tiger, and Chief Telecom (a Taiwanese distributor, reseller, and datacenter operator).

The first $2.5B story is about Wally Liaw, a co-founder of SuperMicro and their Head of Sales. Wally was indicted on March 17, 2026 (US v. Liaw, 1:26-cr-00100, SDNY) along with Steven Chang (Head of Sales for Taiwan) and Willy Sun (local Taiwan contractor). The indictment states that $2.5B worth of servers were sold to a company in Southeast Asia that then $510M were smuggled into China. It also claims that Wally, Steven and Willy staged dummy servers to pass a company audit in Aug 2025, and a Commerce Department audit in Dec 2025.

The indictment literally includes a picture of Willy Sun and his friend using a hair dryer to move labels and serial number stickers from real servers to dummy servers so they could pass an audit.

Source: https://storage.courtlistener.com/recap/gov.uscourts.nysd.660027/gov.uscourts.nysd.660027.2.0.pdf

Source: https://storage.courtlistener.com/recap/gov.uscourts.nysd.660027/gov.uscourts.nysd.660027.2.0.pdf

Since the indictment in March:

Wally has pleaded not guilty and goes to trial on March 29 next year (it was set to start on Nov 2, but was delayed after Supermicro was subpoenaed by SDNY, as they stated in their 10-K, filed Aug 31)
Willy has pleaded guilty to four counts, including staging the dummy servers, posing as a legal assistant named Michael, lying to the Commerce Department officer in Dec 2025 during their audit, and deleting an encrypted messaging app from his phone the day he was arrested (Mar 19). He pleaded guilty on Oct 8, and sentencing is set for Sep 8 next year.
Steven is currently considered a fugitive.

Source: Supermicro’s 10-K from August

Moving on to the second case in Keelung, Taiwan. 9 people have been indicted (so far) for smuggling 130 B300 servers, with 74 reaching China earlier this year:

Chang Teng-lung, Senior Manager NVIDIA Taiwan (key figure)
Lin and Wang, Sales Managers at Supermicro Taiwan
Lu, CEO of Albatron, a Supermicro distributor
Chen, President of Flying Tiger, a server reseller
Tsai, Project Manager at Chief Telecom, a datacenter operator
Yu, head of a customs clearance company
Wang #2, Head of a server reseller company
An unnamed 9th defendant who siphoned money from Albatron

As the story goes, the Taiwan Coast Guard came across the servers during a drug-smuggling investigation in May and detained three people. They then searched 12 more locations, and seized 56 servers, luxury cars, and NT$9M (about $275k USD) in cash.

Following this, the people from Supermicro, Albatron and Chief Telecom were indicted in June, they got Chang from NVIDIA in July, and by August they had all 9. The first hearing for Chang was held on Oct 16.

The defendants are accused of using shell companies, faking the audit, forging documents (including a mockup of a Supermicro website for export permits), and declaring the exports for Japan, where they would be shipped on to China.

For what it’s worth, current process on orders above eight servers triggers an onsite inspection by Supermicro. The coast guard’s inspection apparently found the servers by chance, while pursuing an unrelated drug-smuggling case. This inspection happened after Flying Tiger had applied and received a license to purchase 130 servers from Supermicro in Sep 2025 with themselves named as the end user, and passed an onsite inspection at a Chief Telecom datacenter. The inspection was supposed to confirm they had enough rack space and power to fit the servers, but prosecutors claim it’s obvious that the site doesn’t meet these requirements. They didn’t even have a lease signed.

Here is a picture from the indictment showing Supermicro servers on pallets with crime scene tape around them, cash, and a porsche

Source: Keelung indictment

Both of these cases are being litigated.

It’s Likely Worse Than We Thought

So, if we had to say “why now,” it’s likely to remove opposition to an extremely hawkish China export control bill: after a large shift in Congress and heading into a lame duck session, it’s best to get your opposition out of the way to get the bill passed. That is our best conspiracy theory take on the situation, and it lines up with the ripples from Washington. So now let’s return to our original question.