Sam Altman and Brad Gerstner faced Semafor’s editors days after Jensen Huang resigned as Nvidia CEO, and both used the moment to argue about who should control AI. In his interview with Reed Albergotti at Semafor’s Silicon Valley and the World on November 16, 2026, the OpenAI CEO covered Huang, the double counting of chip demand, the voter backlash against data centers, recursive self-improvement, memory prices and AI for science. In a separate session with Ben Smith, Altimeter Capital’s Brad Gerstner explained his letter to Nvidia’s board, his $5 billion request in the Anthropic IPO, and why he thinks Trump Accounts will reshape the social contract. Both videos are embedded here.
The background: Huang stepped down on November 12, and Nvidia’s CFO Colette Kress became interim CEO while the board searches for a permanent successor. Semafor reports that he left after reports of several extramarital affairs, including with two subordinates and a woman suspected of being a Chinese intelligence asset. It also reports that Huang is weighing a comeback the board is not considering. Those are reported allegations, and the investigations are ongoing.
TLDW
Altman declined to say whether Huang should return, calling him a flawed hero the AI revolution is hard to imagine without. He spent most of his time on a different problem: the public suspects AI’s builders are not “on team human”, and he says OpenAI will hold back on recursive self-improvement to keep people in control. Gerstner said his letter to Nvidia’s board asked for a slow, careful process, not Huang’s immediate return, and argued that selling chips to Southeast Asian clouds was legal while selling to a sham company would not be. He also explained why Altimeter cut its AI exposure but asked for $5 billion of the Anthropic IPO, and why universal capital accounts for children are his answer to the age of AI.
Thoughts
Altman’s most consequential line was about recursive self-improvement, not Nvidia. Asked whether OpenAI is already there, he said the company builds models faster than it could without its own technology, but is “definitely not” pressing a button and letting a model design its successor without oversight, and that it “could be pushing more towards the RSI spectrum” than it is. Then he called holding back “a tactic”, set by a principle he calls unwavering: people must stay in control of the future. That is a frontier lab saying out loud that it is choosing a slower capability curve than it could reach, and naming the rule that decides it. Whether the rule survives a competitor who does push the button is the real question. But it gives outsiders something concrete to hold OpenAI to, which is more than “we take safety seriously”. It also links to his answer on the data center backlash: he thinks voters are not against server farms so much as against the sense that AI’s builders want a “handover” to machines.
Gerstner did the most useful work on the Nvidia story by pulling apart three questions that the coverage keeps merging. The first is Huang’s personal behavior, which he called “pretty bad” and which the board has to weigh. The second is whether Huang misled his board, which Gerstner, a corporate and securities lawyer by training, said he would take “very seriously.” The third is China. Here he drew a sharp line. Shipping chips to Singapore, Malaysia or Indonesia, where Chinese firms rent them through the cloud, was legal under the rules of the past two years, and the tighter licensing rule that Treasury and Commerce had just rushed through proves the old rules allowed it. Selling chips to a sham company like Megaspeed, knowing they were headed to China, would be illegal, and if that happened, he said, Huang should not run Nvidia. Altman, whose own board fired him three years ago, stayed on the emotional ground: “Heroes have flaws,” and the AI revolution is hard to imagine without Huang. Taken together, the two set out the bar a comeback would have to clear. The personal scandal alone might not settle it. The export question would.
Both men are bullish on AI and cautious about the money, and their answers fit together better than the headlines suggest. Altman said he has never seen revenue growth with margins this good in any tech revolution of his career, and that this is why he thinks the debt behind the buildout is sound, while allowing for “foam” and a buildout that could outrun what the economy can absorb. His earlier remark about double counting fits: many new chip makers each assume OpenAI and Anthropic will buy their whole output, which cannot all be true. Gerstner gave the investor’s version. He cut Altimeter’s public AI and semiconductor exposure from large to small after warning in May that the sector was 10 to 20 percent overvalued, yet he put in for $5 billion of the Anthropic IPO, the largest request in his firm’s history. His reason is that AI is moving from an infrastructure phase to an intelligence and applications phase. So the smart money is not leaving AI. It is moving from the companies that sell picks and shovels to the labs.
The memory answer was the most revealing part of Altman’s interview, because it explains a shortage everyone is now paying for. He walked back the notorious RAM comment (he wants people to own lots of memory) and then explained why there is so little of it. Memory makers have spent decades living a business where, as someone described it to him, “you invest $20 billion and you make $19 billion.” After that many booms and busts, nobody wanted to commit capex on the scale he was demanding when he was mocked for the reported $7 trillion figure. Now DRAM prices are up something like 5x to 7x, by his estimate. The supply chain section points the same way. OpenAI has been held up at different times by memory, rack assembly, data center space, and once by a shortage of PCB laminate. AI’s ceiling is set less by any one company and more by the weakest link in an old industrial chain.
Read side by side, the two interviews meet on the public’s fear of AI. Altman’s picture of the bad bargain is vivid: cures and cheap goods in exchange for losing all agency, and living “like a Labrador.” His answer is a principle about control. Gerstner’s answer is about ownership. Trump Accounts, which grew out of his Invest America push, give every newborn $1,000 in the S&P 500, and companies, philanthropists like Michael and Susan Dell, and now Elon Musk with a pledged SpaceX share for every child are adding to them. Gerstner calls the accounts a “universal capital fabric” that is functionally like Social Security and lets the country “amend the social contract” as AI changes work. That is a more concrete idea than anything from the labs, because it gives ordinary people a stake in the gains rather than a promise. It does not answer Altman’s question of who decides. A child who owns a few shares of the winners still has no say in how the models are built.
Key Takeaways
- Altman would not say whether Jensen Huang should return to Nvidia, saying he did not know enough, but called Huang a long-time hero and said it is hard to imagine the AI revolution without him.
- Altman said the chip industry is double counting demand: many new chip vendors are each telling the supply chain they will sell most of their volume to OpenAI and Anthropic.
- On the data center referendums, Altman said the deeper fear is about agency and the balance of power, and that the industry must commit to “people staying in control of the future.”
- OpenAI uses its own models to build new ones faster, but Altman said it is deliberately not pushing as far toward recursive self-improvement as it could, because that would not be safe yet.
- OpenAI now works with embedded outside evaluators, including METR, and Altman said outside opinion on safety will matter more over time.
- Altman agreed in principle that users should be able to see, edit and export the memory built up by Dots, OpenAI’s always-on agent, but said he had no product answer yet.
- On memory chips, Altman said DRAM prices have risen roughly 5x to 7x, and blamed a memory industry so scarred by booms and busts that it would not commit to the capex he called for years ago.
- Altman expects the coming generations of models to surprise people as engines of discovery, and said mathematicians, physicists, chemists and materials scientists now report finding things they could not have found otherwise.
- Gerstner said his Friday letter to Nvidia’s board, co-written with Gavin Baker, called for a careful process that keeps Huang’s return possible, not for his immediate reinstatement.
- Gerstner argued that cloud sales to Chinese firms through Southeast Asia were legal under past rules, but that knowingly selling to a sham company headed for China would disqualify Huang.
- Altimeter cut its public AI and semiconductor exposure to small this year, yet requested $5 billion in the Anthropic IPO, which Gerstner said is 10 to 20 times oversubscribed.
- Gerstner is comfortable with Anthropic’s unusual governance (a single share voted by seven co-founders) because it shields the founders from short-term activists and political pressure.
- Trump Accounts give each newborn $1,000 in the S&P 500, and Gerstner said they have 85 to 90 percent support across both parties, with expansion ideas that reach everyone under 40.
Detailed Summary
Altman on Jensen Huang and the Nvidia Guarantee
Albergotti opened with Saturday Night Live’s sketch on the scandal and the letter of support circulating for Huang, which he said was copied from the one Altman’s supporters drafted when OpenAI’s board removed him. Altman said he did not know enough to say whether Huang should come back, but that he has strong empathy for the situation and “extreme gratitude” that Nvidia bet early on OpenAI. Asked whether Nvidia’s guarantee of about $105 billion for OpenAI’s Ohio data center is worth less now, he said people still trust Nvidia’s guarantees, and that demand for AI is strong enough that the industry will be fine.
IPOs, Investor Demand and Double Counting
With Anthropic going public first, Altman said OpenAI will list “at some point” but that being public has real downsides at a time of high-stakes safety decisions. He said the world wants exposure to AI and that, if anything, investor demand has outrun supply. He explained his reported double counting remark to Samsung executives: when there was one chip supplier, wafer forecasts tracked real demand, but now many chip makers each expect to sell most of their output to the same few labs. He added that the supply chain has bottlenecked on memory, rack assembly, data center capacity, and once on PCB laminate.
The Data Center Backlash and Team Human
After nearly all of the data center referendums passed in the election, Albergotti asked how the industry had told its story so badly. Altman said a pitch that trades cures and cheap goods for losing all agency is “a terrible sales pitch”, and that data centers are the physical stand-in for that fear. He said people can tell that not everyone in AI wants humans in charge, and that the industry must reject a “handover” to AI, gradual or sudden. He said he still wants a party closer to America’s culture of innovation, and that he last spoke with Trump at the China state dinner.
Espionage, Pragmatic Centrism and Recursive Self-Improvement
On spying, Altman said OpenAI has long lived with people seeking an edge, from a camera in a conference room to engineers being taken out for drinks to get them to share pre-training secrets. He described the company’s stance as “pragmatic centrists”: neither building AI in secret nor treating it like any other technology. On recursive self-improvement, he said OpenAI builds models faster using its own tools, but is not letting models improve themselves without oversight. He said he knew of no measure like METR’s report that Anthropic doubled its capability speed, and that OpenAI now works with embedded evaluators, including METR.
Dots, Memory Lock-In and the RAM Comment
Albergotti, a Dots user, asked whether people could export the memory the agent builds up. Altman agreed in principle and compared it to the biggest lock-in in tech, the iPhone camera roll and iMessage history, but said an assistant’s understanding of a life is much harder to export than photos. On RAM, he clarified that he wants people to own plenty of memory. His point was that, in a shortage, shared memory running all the time is a more efficient use than memory sitting idle in home machines. He said memory makers are now ramping production, years later than he wanted.
Debt, the Buildout and AI for Science
Altman said data centers will get built, perhaps further from communities, and that the way the industry engages with towns has to change. On debt, he pointed to unusually strong revenue and margins, and said the market is underrating the next wave: AI for science, chip design and routine work inside organizations. OpenAI announced more free compute for academic researchers that morning. Altman said the next model generations will surprise people as engines of discovery, and that he hopes people will look back at 2026 as a rough time to have lived.
Gerstner’s Letter to the Nvidia Board
Ben Smith opened with reports that Nvidia’s board would not bring Huang back. Gerstner said he could not tell whether that was one director or the whole board, and that his letter asked the board to take the time it needs, not to reinstate Huang at once. He called Nvidia vital to national and economic security and said AI is 75 percent of US GDP growth this year. He said he wants founders running companies, “but not at any cost”, and that he texted Huang only to tell him to take the time he needs with his family. He did not sign the online petition calling for Huang’s immediate return.
China, Export Rules and Megaspeed
Gerstner described an administration split between China hawks, with Treasury Secretary Scott Bessent among them, and an engagement camp that includes Howard Lutnick, with Trump, a self-described deal junkie, leaning toward engagement. He argued that sending chips to Southeast Asian clouds that Chinese firms rent was legal, and that the new licensing rule shows it was. Knowingly selling to a sham company like Megaspeed so that chips reach China would break the law and, if proven, would rule Huang out. He said Trump’s comment about threatening an AI executive with jail was not about Huang.
Cutting AI Exposure and Betting on Anthropic
Gerstner said he warned on May 28 that AI could pull back 20 percent and had taken Altimeter’s public exposure “from large to small”, though its long-term venture and growth funds were unchanged. Anthropic is the exception. He said OpenAI and Anthropic are the two main ways to bet on intelligence and that Altimeter owns both. He accepts Anthropic’s founder-controlled governance because the labs face heavy political pressure, and he called its fight with the Department of War over “all lawful purposes” the messy sausage-making of a democracy.
Invest America, Trump Accounts and Musk’s SpaceX Gift
Gerstner, the main architect of Trump Accounts, said the program should survive Democratic control of Congress because of broad bipartisan support, and credited earlier proposals by Bill Clinton in 1998 and Barack Obama in 2014. Money flows in from three places: the federal $1,000, employers, and philanthropy through a one-page donor agreement with the Treasury. He said Anthropic’s public benefit plan to put money into citizen accounts spurred Elon Musk to pledge a SpaceX share for every child, which Gerstner valued at $11.5 billion (other estimates put it nearer $10.5 to $10.8 billion).
Notable Quotes
“Jensen has long been a hero of mine. Heroes have flaws. I am grateful for the company he’s built.”
Sam Altman, on whether Jensen Huang should return to Nvidia
“The price of that is you are going to lose all agency and control and voice in the future. You are going to live like a Labrador.”
Sam Altman, on why the AI industry’s sales pitch fails
“You invest $20 billion and you make $19 billion and that’s it.”
Sam Altman, relaying how the memory business has been described to him
“I hope people will look back at 2026 and be like they had it really rough.”
Sam Altman, on AI as an engine of scientific discovery
“Let’s be clear, nobody is above the law, okay? Not Jensen Huang, not any other founders.”
Brad Gerstner, on what Nvidia’s board should weigh
“It’s clear to me we need to take more time, be deliberate, and get this right.”
Brad Gerstner, on the Nvidia board’s decision
“It’s going to be the largest assault on the wealth gap in the history of the country.”
Brad Gerstner, on Trump Accounts
Watch the full conversations here: Sam Altman at Semafor’s Silicon Valley and the World and Brad Gerstner on Nvidia and Jensen Huang.
Related Reading
- Semafor’s report on Gerstner’s remarks covers the letter he co-wrote with Gavin Baker and what the board is weighing.
- Jensen Huang (Wikipedia) covers the Nvidia co-founder’s career and the events of his resignation.
- Recursive self-improvement (Wikipedia) explains the idea Altman says OpenAI is deliberately holding back on.
- METR is the independent evaluator Altman says is now embedded with OpenAI.
- Altimeter Capital is Gerstner’s firm, a large Nvidia shareholder and an investor in both OpenAI and Anthropic.