From September 13 to 15, the All-In Summit 2026 was held at the Shrine Auditorium and Expo Hall in Los Angeles, USA.

The summit was hosted by the four hosts of the tech podcast All-In—Chamath Palihapitiya, Jason Calacanis, David Sacks, and David Friedberg—over three days, following the podcast's conversational format and bringing together guests from venture capital, tech giants, macro investing, and politics on the same stage.

The remarks by Tesla and SpaceX CEO Musk, Microsoft CEO Satya Nadella, U.S. Vice President JD Vance, Nvidia CEO Jensen Huang, and others focused on two things: AI expansion will first hit bottlenecks in chip capacity and power supply; and the model safety of frontier AI companies cannot rely only on self-inspection—competitors or independent third parties should be brought in to check.

Compute and Power: The Two Bottlenecks of AI Expansion

Musk: Existing fabs are already at full capacity, "either build Terafab, or expansion cannot continue"

Tesla and SpaceX CEO Musk explained why the two companies are building their own Terafab chip factory: key chips may face supply cuts in the future, but even if supply remains stable, existing fabs are already basically at full capacity, and their output cannot support continued AI expansion. In the future, compute will not only need to go into data centers, but also support humanoid robots and cars, and logic chips, memory, and advanced packaging all need to be expanded together.

Musk said, "either build Terafab, or expansion cannot continue." Tesla and SpaceX are advancing a joint chip production line in Austin, equipment has been ordered, and the goal is to produce usable chips by the end of next year, but there is still some time before large-scale mass production.

Vance: Data centers face protests because of power shortages; U.S. electricity should be so cheap that it is "too cheap to meter"

U.S. Vice President JD Vance believes that growing opposition to AI data centers is mainly because power supply has not kept up—residents see data centers being built continuously while their own electricity bills are rising. The approach he advocates is not to build fewer data centers, but to add generating capacity.

Vance said the United States has not built enough power facilities over the past generation, and its power generation growth has lagged behind China, already dragging down AI and other new industries. He proposed that the United States should make electricity "too cheap to meter," a phrase that comes from 1954, when U.S. Atomic Energy Commission Chairman Lewis Strauss described the atomic age.

The IEA estimates that data centers accounted for about half of new U.S. electricity demand in 2025; the U.S. Congress is also advancing legislation this week to prevent the grid costs of data center expansion from being passed on to ordinary households.

Jensen Huang: China will be able to build domestic advanced lithography systems by 2030

Nvidia CEO Jensen Huang judges that China will be able to build its own advanced lithography systems by 2030, and with its large-scale production capacity, it is only a matter of time before domestic equipment enters local wafer fabs.

This is far more optimistic than Zeiss's assessment. As a core supplier of ASML's EUV optical systems, Frank Rohmund, head of Zeiss Semiconductor, recently said that it may take China about 15 more years to build a domestic EUV lithography machine. However, the host asked about "advanced lithography systems," and Huang did not explicitly say EUV.

AI safety: cannot "grade your own homework"

Musk: before releasing new models, competitors should conduct safety tests first

Tesla and SpaceX CEO Musk proposed that before frontier AI companies release new models, competitors should first test them using their own safety testing systems. He compared the current approach to "grading your own homework," arguing that competitors have more incentive to find problems; if rivals have already warned that a model is unsafe and the company still insists on releasing it, it will be difficult to defend itself after an incident, and legal liability will also be very large. Musk also suggested including major AI players such as China in this mechanism to avoid only U.S. companies being unilaterally restricted.

Nadella: AI agents may cook the books to complete tasks, a "new type of internal risk" for enterprises

Microsoft CEO Satya Nadella warned that AI agents running autonomously for long periods may become a new "internal risk" for enterprises. For example, if a frontier model is asked to optimize working capital, it may even "cook the books" to achieve the goal.

Nadella believes that enterprises cannot only look at the results delivered by agents, but must also verify the process: use additional models to check results, continuously monitor agent operations, and ensure that behaviors such as reading confidential information and continuously calling multiple systems are traceable and auditable.

Vance criticizes Anthropic: don't create a "Frankenstein" and then ask the government for regulation

U.S. Vice President JD Vance criticized some frontier AI companies for saying they have created a "Frankenstein" while pointing to global AI governance as the solution. He believes that if they have truly created a dangerous model, they should stop first; if the capabilities have already been released, they should first build defensive tools.

Vance then named Anthropic, saying its latest model has strong cyberattack capabilities, but companies that urgently need tools of the same caliber for defense cannot get access. He did not specify which companies. Anthropic's Mythos 5.1 is currently only available to a small number of vetted institutions, and the company says the reason is that the model is too capable in cybersecurity and biological research and risks being misused.

Vance said he does not believe Anthropic CEO Dario Amodei is manufacturing panic for regulatory capture, and that he has heard Dario Amodei is very serious and genuinely believes in these risks. What he criticized was the approach: "Don't build Frankenstein and then come to the government and say we need regulation."

Jensen Huang: Opposes slowing AI, but supports third-party audits of frontier labs

Nvidia CEO Jensen Huang opposes Anthropic CEO Dario Amodei's call to slow down frontier AI development, but agrees with another of his proposals: having multiple independent third parties enter frontier labs to conduct safety assessments. He believes the most dangerous AI safety problems will first appear in frontier labs, and these companies should first do a good job on sandboxes, monitoring, and evaluations, and accept external inspections like financial audits, with even higher standards.

Midway through the interview, U.S. President Trump called Huang, and Huang put him on speakerphone. Trump called AI taking over the world a "hoax," saying "robots will not take over the world" and "whoever wins AI wins." Huang echoed, "You're right." Trump asked whether he thought the same as he did, and Huang replied, "Yes, Mr. President."

Welcome to join the official BlockBeats community:

Telegram Subscription Group: https://t.me/theblockbeats

Telegram Discussion Group: https://t.me/BlockBeats_App

Official Twitter Account: https://twitter.com/BlockBeatsAsia