Nvidia After Jensen Huang: Ben Thompson on the F1 Car Without Its Driver

Ben Thompson and Andrew Sharp spend the 300th episode of Sharp Tech on what Nvidia and the wider AI industry look like after Jensen Huang’s exit as CEO, and on why Thompson thinks Washington’s win on chip export controls is a long-term loss for American semiconductors. The episode was recorded on November 12, 2026, the day Nvidia named CFO Colette Kress interim CEO after two weeks of reporting on Huang’s affairs and on a customer, Megaspeed, suspected of moving chips to China. Thompson has interviewed Huang repeatedly for Stratechery, and the hosts treat the day as a funeral. You can watch the full episode on YouTube.

TLDW

Thompson argues that Nvidia was built like a Formula 1 car around one driver: every specialist reported up to Huang, which made the company fast and left it with weak governance and no succession plan. He credits Huang with more than CUDA and GPUs, saying Huang personally sold Asia’s conservative supply chain on building for AI and bankrolled the neoclouds and open models that kept Nvidia from depending on a few giant buyers. The hosts read the scandal as Washington’s national security establishment finally getting a man who had treated the chip ban as illegitimate, and Thompson restates his case that the ban hands China the reason to build its own ecosystem. The last twenty minutes are lighter: Thompson’s 3D printer, his basement hardware store, and an AI agent that finally clicked a Mac permission dialog on its own.

Thoughts

Thompson’s case against the chip ban is stronger than the caricature of it, and it is worth separating from the man who lobbied for it. His claim is that for decades the United States sold other countries technology two or three generations behind the leading edge on purpose. The buyer gets something far better than what it can make, so nobody inside that country has a reason to nurse a weak domestic startup, and the money flows back to fund the next American generation. A total ban reverses the incentive. It does not just allow a Chinese alternative to CUDA, it requires one, because the developers who would rather use Nvidia no longer can. The part that deserves more attention is his point about where ecosystems come from. Governments can fund a project, but developers pull an ecosystem into existence by choosing it, and Microsoft’s billions could not conjure one for Windows Phone while Android and the iPhone were on the shelf. If that is right, the ban is the one policy that supplies the missing ingredient.

The line that will outlast the news cycle is Thompson’s reading of the fall as a story where the virtues and the flaws are the same traits. CUDA was a bet that nearly sank the company. So was pushing through a generation of hot, slow chips to get programmability. A founder who has been right against consensus that many times stops being able to tell a bad bet from an unpopular one, and Thompson says Huang had many off-ramps on China and took none because he thought he was too big to fail. What I would add is that the hosts never quite reconcile two of their own positions. Thompson thinks the policy is wrong, and he also says plainly that laws are laws and that he suspects Huang treated the ban as something to appear to follow. Both can be true, but it means the best argument against export controls has just lost its most prominent advocate in the worst possible way, and Sharp is right to wear black for the argument as much as for the man.

The near-term risk in this episode is financial, not technical. Thompson describes the AI buildout as having moved from spending free cash flow, to borrowing, to raising equity, with real revenue needed on the other side and a bridge needed to get there. Huang was that bridge: backstopping companies, supporting customers, acting as what Thompson calls the bank for the bubble. An interim chief who is also the CFO has every professional reason to stop doing that, and credit was already tightening for the neoclouds before any of this. Nvidia itself will be fine for a while. The companies that existed because Huang wanted more buyers than Amazon, Google and Microsoft are the ones to watch, and so are the open model developers who had exactly one deep-pocketed patron whose incentives matched theirs.

The organizational argument explains why this is worse for Nvidia than a normal succession. Thompson’s picture of company growth runs from walking, to a bicycle, to a car, to a bullet train: each step is faster and more efficient and can go fewer places. Nvidia never became the train. It stayed a functional organization of specialists with a single generalist at the top, which is how it turned on a dime and also why it had no succession plan and a board that never built one. His Apple comparison is the useful one. Steve Jobs’s illness gave Apple years of warning, and Tim Cook used them to make the company an entity in its own right. Nvidia got two weeks. Whoever takes the job permanently inherits a machine designed for someone else’s reflexes, and the first instinct of any successor will be to add the structure that slows it down.

The closing segment looks like a hobby update and is quietly the most forward-looking part. Thompson says he has not opened his 3D printer’s software once, or his label printer’s. He describes the part to an AI model, it writes the 3D model, slices it, sends it to the printer and watches the camera. He then revises a claim he made earlier about everyone having software-sized holes in their life: what he had was a missing system of record, and the app he built to fill it mostly goes unused now because he just asks the agent. That is a small data point about one enthusiast, but it points the same direction as the Nvidia story. The scarce thing is no longer the software. It is the compute underneath it, which is why who controls the chips, and who gets to buy them, became a fight worth ending a career over.

Key Takeaways

Sharp opens with the governance puzzle: Nvidia stock was down about 15% in five days on the prospect of Huang leaving, so most shareholders plainly wanted him to stay. Thompson’s answer is that boards have their own incentives, including personal liability once they know of misconduct.
Thompson calls a relationship with a subordinate a corporate red line in America today, and says there may be more the board knew that is not yet public.
Nvidia’s founding insight, in Thompson’s telling, was the programmable shader: graphics chips that could run programs instead of fixed commands. CUDA grew out of that, and the bet nearly killed the company when the first chips to support it ran hot and lost to ATI.
AlexNet, an image recognition system built on Nvidia chips with CUDA, was the wake-up call that the hardware mattered beyond games. ChatGPT was the shift that made it stick.
Thompson lists three moats: CUDA, selling whole systems instead of chips (made possible by buying Mellanox for networking), and the supply chain. He says AMD has made faster chips and still loses because it does not build systems.
Huang’s trips through Seoul and Taipei were about securing memory and components, and about persuading closely held Asian suppliers to invest. Thompson says the world would have far fewer chips today without that salesmanship.
Huang said on stage that his keynotes were the most efficient way to tell suppliers where Nvidia was going so they could prepare.
Nvidia funded and favored neoclouds because a seller with only a handful of huge buyers loses pricing power. Thompson compares it to AMD gaining on Intel inside the hyperscalers, who will do the work to qualify a second supplier.
For AI startups and open models, Thompson calls Huang’s exit a disaster: the one entity with enormous resources whose incentives lined up with theirs has lost the person who drove those bets.
Thompson doubts the story of Alice Huang as a trained spy. His view is that anyone moving billions of dollars of restricted hardware is known to the Chinese state, and that the spy framing mostly served to start the endgame.
Both hosts reject the idea that Huang lobbied to sell to China because of the relationship. Sharp notes Megaspeed was about 1% of Nvidia’s revenue, and Thompson says Huang wanted China sales because of his moat and his business.
The feud with the national security world began, Thompson says, when Nvidia answered the first export rule within a month with a lower-bandwidth chip that complied. Officials saw circumvention. Huang saw himself following the law.
Thompson treats chips reaching the Chinese military as an obvious cost of selling into China at all, and says he owns that trade-off when he opposes a total ban.
He ties support for the ban to a belief that only the next two or three years matter because superintelligence is close, and asks what happens to the chip industry if that belief is wrong.
Both expect much stronger chip controls now, more investigations, and a higher chance of a money crunch across AI infrastructure.

Chapters

00:00 Nvidia After Jensen Huang

Sharp recaps the week: a New York Times report citing five sources at Nvidia on an affair with an executive at the company, a Wall Street Journal report that the board was preparing to remove Huang, and the official change on Thursday morning. He asks why a board would do something shareholders did not want. Thompson walks through the principal-agent problem and notes that a board can be misaligned with shareholders too. He also says Nvidia’s governance was not admirable, a point he later partly defends.

05:29 The Bets That Built Nvidia and How AI Changed It

Thompson calls Huang a singular figure in tech history and credits The Nvidia Way by Tae Kim for the F1 analogy. He traces the line from the first failed graphics card to programmable shaders, CUDA, AlexNet and ChatGPT. He remembers the years when Huang was throwing CUDA demos at the wall as the most fun time to watch the company, and says the first stadium-sized GTC was peak fame and the least interesting Huang, because the roadmap had become the business. Success also gave customers real reasons to fight CUDA.

17:14 Jensen’s Role in the AI Buildout

The supply chain is the third moat. Thompson describes Asian suppliers as conservative, closely held businesses that do not want to overinvest, and says Huang is the one who got them building. He then names what he sees as the root of the trouble: Huang became so focused on preserving CUDA as a moat that it led him into fights over chip allocation, neocloud financing and China. Thompson notes that Nvidia chips still lead on tokens per watt, which matters when electricity is the constraint.

26:50 Nvidia’s F1 Car Without Its Driver

Thompson says dealing with Nvidia felt like dealing with a startup, charming and seat of the pants, at a company worth about four and a half trillion dollars. He compares it to Apple under Steve Jobs, a functional organization where specialists report up and the only generalist is the CEO, and recalls Steve Ballmer’s attempt to reorganize Microsoft the same way. Huang’s 60 direct reports are what that looks like. It works because Nvidia really sells one product, a GPU that software turns into many.

38:52 What Jensen’s Absence Means for Startups and Open Models

Sharp worries there will be fewer bets like Nvidia’s spending on open source AI. Thompson agrees it is a disaster for startups, who built on Nvidia because developers and capacity were everywhere. He then previews the China argument: the people who could create an alternative to CUDA are engineers and companies who use AI, not the Chinese government, and all of them would prefer Nvidia as long as its chips are available.

53:00 Alice Huang and the Spy Allegations

During the recording Sharp gets word of a Reuters report on another relationship with an Nvidia employee. Thompson does not want to speculate on numbers and says he is not shocked. On Alice Huang of Megaspeed, he argues that in China every company ultimately answers to the state, so the line between spy and well-connected trader is blurry. He asks why, if she was such a threat, it was not dealt with months ago, and says the salacious framing raised the pressure for what followed.

1:02:24 Washington Gets Its Revenge

Thompson says Nvidia had no presence in Washington when the first rule arrived, and that its quick, compliant workaround chip set the tone: officials saw a bad actor and Huang saw idiots. Sharp compares it to Anthropic trying to dictate terms to the government and says the leaks look coordinated to delegitimize Huang’s policy agenda. Thompson adds a second offense, that Huang had the president’s ear and got the H20 restriction reversed over the objections of the national security establishment.

1:08:58 The Case Against Chip Bans

Thompson says the burden of proof belongs to the people who abandoned fifty years of practice on technology transfer. Selling older generations abroad prevents competitors from forming and pays for research at home. He argues that in the fullness of time China will have its own CUDA and its own manufacturing and will sell both worldwide. He also questions the premise that the United States could hold superintelligence alone without China responding, and points to how fragile TSMC’s lines are.

1:17:41 Risks to the AI Boom, Hubris and Downfall

Sharp declares the sell-to-China position finished. Thompson calls it a disaster for the chip industry in the very long run and bad for AI financing now, since Huang was backstopping companies through the gap between investment and revenue. Sharp says he likes Huang’s optimism and still finds it mystifying that an American company kept putting itself in this position. Thompson says Huang’s country was Nvidia, that he suspects there is fire behind the smoke, and that the story is an old one about a flawed great figure.

1:28:05 300 Episodes: 3D Printing, Home Inventory and a Permission Dialog

Thompson describes his Bambu Lab H2C printer, about $1,500, and a plan to never drive to the hardware store again. An AI agent drew up the list of fasteners, filled the shopping carts with hundreds of small packs, designed rails to hold 96 loose bins, and ran a label workflow with a spare keyboard turned into a single “next” button. The episode ends with his delight that an agent with outside-in access to its own Mac mini finally clicked OK on a permission dialog by itself.

Notable Quotes

“It’s not a chip game anymore. It’s a systems game and AMD is not building systems.”
Ben Thompson, on Nvidia’s second moat

“He was the salesman in the supply chain that got Asia believing in AI so they could supply everything that’s going on today.”
Ben Thompson, on Jensen Huang’s role in the buildout

“Ecosystems are pulled into existence by the developers who use them and want them and leverage them.”
Ben Thompson, on why a government cannot order up a CUDA alternative

“I think Jensen, he never fully appreciated how powerful his enemies in Washington were.”
Andrew Sharp, on how the scandal unfolded

“It’s a beautiful system. The chip ban throws it all away.”
Ben Thompson, on selling older generations of technology abroad

“What makes this into a myth, what makes this into a movie is all the characteristics that made Nvidia what it was are the characteristics that led to his downfall.”
Ben Thompson, on Huang’s big bets

“The degree to which AI closes the loop on idea to done is so tight it’s unbelievable.”
Ben Thompson, on building things with AI agents

Watch the full conversation here.