Elon Musk, Jensen Huang and Anthropic's Tom Brown With Gavin Baker at the America.gov Event: Superintelligence Factories, Orbital Compute, the Watts-to-GDP Rule and a Joint SI Safety Declaration

Fresh from a long lunch at the White House, Elon Musk and Jensen Huang took the stage at the launch event for America.gov, a new AI-powered government website, in a conversation moderated by Atreides Management founder Gavin Baker, and were later joined by Anthropic co-founder Tom Brown. In about 26 minutes they covered how America wins the race to artificial superintelligence, why energy is the real bottleneck, how SpaceX plans to put hundreds of gigawatts of compute into orbit, what Nvidia is doing to contain and monitor AI agents, what Claude Opus 5.5 means for ordinary Americans, and a joint SI safety declaration the companies signed that day. Watch the full conversation on YouTube.

TLDW

Jensen Huang frames superintelligence (now branded “SI”) as a reinvention of the entire computing stack that rests on energy, then chips, algorithms and applications, and credits President Trump’s pro-energy stance for making the data center buildout possible. Elon Musk says America leads on software but China produces about three times the electricity, so the long-run challenge is power generation plus domestic logic and memory fabs. Musk offers a rule of thumb: US average consumption is about 500 gigawatts, so every 5 gigawatts is a 1% rise in power use and, he bets, roughly a 1% rise in GDP; Huang puts each gigawatt at $40 to $50 billion of annual output. Musk describes orbital compute as the path to far more energy, with SpaceX and Tesla targeting 200 gigawatts a year of solar production, which in space delivers nameplate output around the clock instead of a fifth to an eighth of it on the ground. Huang recasts data centers as “SI factories,” estimates 10 to 20 gigawatts a year of US buildout creating about a million construction and trade jobs, and argues that market demand now funds every form of sustainable energy without subsidies. Musk says Colossus created a labor shortage, roughly doubled the local tax budget, and came with half-price Starlink and a $250 million water recycling plant. Huang explains Nvidia’s open agent safety framework: OpenShell contains agents like a browser sandbox, and BlueField chips monitor them out of band. Musk reports Starship’s first orbital flight deployed Starlink V3 satellites, the largest payload since Skylab, with weekly or twice-weekly launches expected next year and a million tons needed for a self-growing city on the Moon or Mars. Tom Brown says Opus 5.5 handles longer delegated tasks and will make software cheaper and better. Huang describes using Claude for research reports, Musk cites AI catching medical errors, Huang questions how much moral judgment tools should have with a steak knife analogy, and the group describes a joint declaration built on mutual monitoring, internal and external audits, and “grading each other’s homework.”

Thoughts

Musk’s watts-to-GDP bet is the most useful idea in the first half, because it turns an abstract race into something you can put on a spreadsheet. If 5 gigawatts of steady-state power is 1% of US consumption and roughly 1% of GDP, then a single gigawatt is worth around $60 billion a year, and Huang’s more conservative $40 to $50 billion lands in the same neighborhood. The claim only holds if intelligence per watt keeps rising at both the chip and the algorithm level, which is exactly the point Musk makes. It also explains why both men keep returning to China’s roughly three-to-one electricity advantage: if power converts to output this directly, the grid is the scoreboard and chips are just the multiplier.

The orbital compute pitch sounds like science fiction until you run Musk’s numbers. Ground solar yields a fifth to an eighth of its nameplate capacity and needs huge batteries to cover the night, while panels in the right orbit run near nameplate all the time. His argument about timing is the sharp part: launch costs are still high, but the returns on compute are so large right now that the launch bill looks cheap by comparison. When he later says Starship could launch “a few hundred gigawatts a year” of compute, which at 200 gigawatts would equal a 40% annual increase in total US energy consumption, you see why Starship’s cadence matters to AI as much as it does to Mars.

Huang’s explanation of agent safety in the middle of the conversation is more concrete than most AI safety talk from a CEO. The key line is that even a contained model should never be trusted to stay contained, so monitoring has to run on separate hardware outside the agent’s sandbox. That is essentially the security model you would use for an untrusted employee with admin rights: limited permissions (OpenShell) plus an independent auditor that sees everything (BlueField). Tying safety to the early internet, when fear of stolen credit cards held back e-commerce until it was solved, makes a fair case that containment is what lets adoption speed up rather than slow down.

The most interesting moment comes late, when Huang asks if he is allowed to say something controversial and then questions the premise that superintelligences should carry values at all. His steak knife analogy (a knife that cuts medium-well steak but refuses rare steak because it looks too much like live meat) captures a real tension: a tool that judges its user becomes less reliable, and in security work a defender’s actions often look just like an attacker’s. The answer came right after a question about imbuing AI with American values, and it sits awkwardly beside the safety framework he had just described. Containment and monitoring constrain what an agent can do; judgment is about whether it wants to help. How the industry balances those two ideas is still unresolved, and it deserves a longer discussion than it got here.

The joint SI safety declaration, mentioned in the last few minutes, may matter more than anything else in the video. Musk’s phrase “grading each other’s homework” describes an unusual arrangement for competing labs: joint monitoring, board special committees, internal audits that escalate, third-party external audits and shared best practices. Huang says it “has a lot of teeth.” Whether it does depends on the published text and on enforcement, neither of which the panel could show on stage. Still, when the heads of Nvidia and SpaceX and an Anthropic co-founder agree that labs should not self-certify, mutual verification is on its way to becoming the industry default.

Key Takeaways

The speakers repeatedly use “superintelligence” or “SI” instead of “AI,” and Baker calls it good branding. Huang corrects himself on stage more than once and jokes that he is “going to get good at this.”
Huang describes superintelligence as a reinvention of every layer of the computing stack: chip design, how software is trained, and what applications can do.
Generative computing needs energy first. Huang’s stack is energy, then compute, then algorithms, then applications, with land, power and shells hosting the infrastructure.
America is world class in chip design and algorithm development, Huang says, so the main job is to stay vigilant about enabling power generation and power delivery.
Huang credits President Trump’s pro-energy vision, saying the current data center buildout would otherwise have been impossible.
Huang admits the industry has to do a better job of partnering with the communities where data centers are going up and sharing the prosperity with them.
Huang’s four-part formula for winning: the best chips, lots of energy, great algorithms, and a country enthusiastic enough to spread the technology across manufacturing, healthcare and space.
Musk says America is winning on anything intellectual or digital, but China produces about three times as much electricity as the United States.
Musk names the central long-term challenge: generating enough power, and securing enough logic and memory fabrication capacity in the US or a safe region.
Baker notes that SpaceX’s publicly discussed 10 gigawatts would be about 25% of all power added in America, with rocket engineers helping build turbines.
Musk’s framing: average US power consumption is about 500 gigawatts, so every 5 gigawatts of steady-state production is a 1% increase in power use.
Musk bets that a 1% increase in power usage produces roughly a 1% increase in GDP, because useful intelligence per watt keeps improving in both hardware and algorithms. On that math, 10 gigawatts would be a 2% GDP increase.
Huang puts each gigawatt at about $40 to $50 billion of economic output a year. By Musk’s math (1% of GDP is about $300 billion) it works out closer to $60 billion.
SpaceX sees space as the way to reach far more energy than is available on the ground, and Musk says orbital compute will be “a very big deal.”
Musk notes Google agrees and is sending TPUs to orbit, and that an H100 was put into orbit last year, showing GPUs work in space.
SpaceX and Tesla are aiming for 200 gigawatts of solar production per year.
In space, solar panels deliver their nameplate output or better. On the ground they deliver a fifth to an eighth of it and need enormous batteries.
Launch costs are still high, but compared with the economic returns of AI compute, Musk says, they are affordable, which makes now the right time.
Huang renames data centers “SI factories” because they produce something of economic value instead of just storing data.
Just as the internet lets you find anything anywhere, Huang says SI infrastructure will let you “know anything and do anything from anywhere.”
Huang estimates the US will build 10 to 20 gigawatts a year, creating roughly a million jobs in construction, power plants, cooling and trades such as pipe fitting.
Huang calls it the first reindustrialization of the United States in about 50 years and says he wants an economy with both white-collar and blue-collar jobs.
Musk says the region around xAI’s Colossus has gone from unemployment to a job shortage, and the local tax budget may have roughly doubled.
xAI gives the community half-price Starlink access and is building a $250 million water recycling plant. Musk: do things for the community and they will reciprocate.
Huang says that for the first time, market forces alone fund almost every form of sustainable energy (fusion, fission, small and large reactors, battery storage) without government subsidies.
Baker credits Musk and Huang with more than $5 trillion of wealth in Americans’ 401(k)s. Musk says Huang did $5 trillion by himself, and Huang corrects it to about three and a half.
Nvidia’s open agent safety framework treats testing and deployment the same way: contain the agent and monitor it in real time, the same way you give an employee limited rights and access and keep an eye on them.
OpenShell is a containment system, “a browser of agents,” that isolates the agent from the rest of the computer and grants only the rights it should have.
BlueField chips monitor the agent out of band, from outside its sandbox, and escalate anything that breaks policy.
Huang argues safety and capability are not in conflict, comparing it to autopilot, which is both more advanced and safer. He calls the tradeoff a “false choice.”
On its first orbital flight, Starship deployed Starlink V3 satellites with the wingspan of a Boeing 737, which Musk says was the largest payload to orbit since Skylab.
Musk expects Starship to reach a weekly or twice-weekly launch cadence next year.
A self-growing city on the Moon or Mars needs roughly a million tons delivered to the surface, Musk estimates.
Starship should also launch at least a few hundred gigawatts of AI compute a year, possibly a terawatt one day. At 200 gigawatts a year, that equals a 40% annual increase in total US energy consumption.
Tom Brown says Opus 5.5 continues steady model-over-model gains, and the main difference users will notice is that it can handle longer tasks than previous models.
The less visible effect, Brown says, is that models will write huge amounts of software, making every product people use better and cheaper over time.
Baker cites cases of superintelligence identifying a known drug as a safe treatment for a disease it was not previously known to treat.
Huang uses Claude to write comparative research reports with charts, turning four to eight hours of weekend research into about three minutes.
Musk says AI is becoming essential in both professional and personal life, and cites many anecdotes of AI reading X-rays or MRIs correctly after a doctor got them wrong.
Asked about imbuing SI with American values like free speech and commitment to truth, Huang pushes back, saying SI is a tool and a tool with judgment may judge its user and work less well.
In cybersecurity, Huang notes, a defender’s actions can look like an attacker’s, so the SI has to act for the defender even when the action looks aggressive.
The companies signed a joint declaration on SI safety at the White House, with joint monitoring, board special committees and mutual review, which Musk calls “grading each other’s homework.”
Huang says the declaration covers rigorous internal controls, clearly documented product intentions, internal audits that escalate, third-party external audits and shared best practices, and calls it a standard every company should adopt.

Detailed Summary

How America wins the superintelligence race

Gavin Baker, who moderates, opens with a list of American technology firsts (the Wright brothers, the Moon landing, the internet) and asks how America wins what is now being called superintelligence while making that win good for ordinary Americans. Jensen Huang describes SI as a complete reinvention of the computing stack. Computing has moved from retrieving information to generating it, so the first thing it needs is energy: energy runs compute, compute runs algorithms, algorithms feed applications. America leads in chip design and algorithms, he says, so the task is enabling power generation and delivery, and he credits President Trump’s pro-energy stance for making the data center buildout possible. He also concedes that data centers are arriving quickly in communities that need to be brought along. His formula is the world’s best chips, lots of energy, great algorithms, and a population enthusiastic enough to spread the technology into every industry.

Elon Musk agrees that the key is scaling energy and chip production. America wins on software and anything digital, but China produces about three times as much electricity, so the long-term challenge is generating enough power and securing enough logic and memory fabs in the United States or a safe region.

The watts-to-GDP rule of thumb

Baker notes that SpaceX’s publicly discussed 10 gigawatts would be about a quarter of all power added in America, with rocket engineers helping build turbines. Musk frames the scale: the US uses about 500 gigawatts on average, so every 5 gigawatts of steady-state production is a 1% increase. He bets it will turn out to be roughly a 1% increase in GDP as well, because intelligence per watt keeps rising as Nvidia’s GPUs get more compute per watt and algorithms extract more intelligence from each watt. By that logic, 10 gigawatts is 2% of GDP. Huang puts each gigawatt at about $40 to $50 billion of annual output, and the two work through the math on stage: 1% of US GDP is about $300 billion, so 5 gigawatts is $300 billion and one gigawatt is about $60 billion.

Orbital compute and 200 gigawatts of solar

For SpaceX, Musk says, the long-term way to reach far more energy than the ground can supply is space. Orbital compute will be a very big deal. Google is sending up TPUs, and an H100 went into orbit last year, so it demonstrably works. SpaceX and Tesla are aiming for 200 gigawatts a year of solar production, and in space those panels deliver nameplate output or better, while on the ground they deliver a fifth to an eighth of it and need enormous batteries. The returns on computing are so high that even today’s launch costs are affordable, which makes now the right time. Huang responds that putting Nvidia computing in space is the best way to get there, and Baker sums it up as American power and American chips on American rockets.

SI factories, jobs and communities

Asked how data centers help ordinary Americans, Huang says the name is outdated. They were called data centers because they held data, but the new buildings are SI factories: they produce something of economic value, and they combine power generation, computing, networking and cooling. Just as the internet lets you find anything anywhere, SI will let you know anything and do anything from anywhere. At 10 to 20 gigawatts a year across the US, he estimates about a million jobs building power plants and data centers, from construction to cooling to pipe fitting. He calls it the first reindustrialization of the country in 50 years and a chance to build an economy with blue-collar as well as white-collar jobs.

Musk describes the effect around xAI’s Colossus: a region that had unemployment now has more jobs than people, the local tax budget may have doubled, and xAI provides half-price Starlink and is building a $250 million water recycling plant. Huang adds that, for perhaps the first time, market forces are funding nearly every kind of sustainable energy, from fusion and fission to small reactors and battery storage, without government subsidies. Baker thanks the pair for creating more than $5 trillion of wealth in Americans’ 401(k)s. Musk says Huang did $5 trillion by himself, and Huang corrects him to around three and a half.

Nvidia’s open agent safety framework: OpenShell and BlueField

Baker mentions reading about Nvidia’s open agent safety framework on the morning of Starship’s first orbital launch. Huang says Nvidia wants to advance SI optimistically but also safely and securely. Training is a fairly safe environment. During evaluation and testing, though, the agentic system must be tightly contained, and because you should never trust that it is contained, it also has to be monitored in real time, ideally by a separate chip outside the agent’s sandbox. The same approach carries over into deployment, much like giving a new employee limited rights, tools and access and then monitoring them continuously. OpenShell is the containment layer, described as a browser for agents that isolates them from the rest of the computer. BlueField is a chip that watches everything the agent does and escalates anything out of policy. Baker compares it to early internet fears about stolen credit cards: making the web safe is what made Amazon, Netflix, Google and X possible. Huang adds that safety and capability are one and the same, like autopilot, and that the tradeoff between them is a false choice.

Starship reaches orbit

Musk says Starship reached orbit for the first time and deployed Starlink V3 satellites with the wingspan of a Boeing 737, the largest payload delivered to orbit since Skylab. He expects a weekly or twice-weekly cadence next year. Starship’s purpose is to carry enough payload to the Moon and Mars to build a self-growing city, which he puts at roughly a million tons to the surface. It will also launch at least a few hundred gigawatts of AI compute a year. At 200 gigawatts, that equals a 40% annual increase in total US energy consumption, and it might one day reach a terawatt.

Tom Brown on Claude Opus 5.5

Tom Brown, the Anthropic co-founder, arrives midway through (his car took a wrong turn) and is asked what Opus 5.5 does for ordinary Americans. He describes it as a continuation of steady progress, each release adding “more IQ points.” The most visible change is that people can hand it longer tasks than earlier models could handle. The less visible one is that these models will write a huge amount of software, so the products everyone uses will get better and cheaper as engineering work is shared between people and superintelligence.

What superintelligence does for ordinary people today

Baker cites cases of SI finding that an existing drug could safely treat a disease it was not known to treat. Calling himself an ordinary American, Huang describes using Claude for comparative research: he lists every question, asks for a report with charts, and enjoys watching it work, getting in about three minutes what used to take four to eight hours of weekend reading. Musk says AI is helping people organize their lives and do their jobs better, is becoming essential at work and at home, and has in many cases read X-rays or MRIs correctly after a doctor got them wrong, saving lives.

Values, judgment and the steak knife

Baker asks how each of them imbues SI with American values such as the First Amendment, free speech and commitment to truth. Huang, after asking whether he is allowed to say something controversial, argues that SIs are tools, and that a tool with values has judgment and may judge its user, which affects how well it works. A steak knife that cuts medium-well steak perfectly but refuses rare steak because it looks like live meat is not a better knife. He leaves open how much judgment these tools should have and points to cybersecurity, where a defender’s actions can resemble an attack and the SI still needs to act on the defender’s behalf.

The joint SI safety declaration from the White House lunch

Asked what came out of their long White House lunch, the group says they signed a joint declaration on SI safety, to be published by the White House. Musk says it includes joint monitoring, board special committees and grading each other’s homework instead of their own. Huang lists rigorous internal controls, established methods and processes, clearly documented product intentions, internal audits that escalate, third-party external audits and industry sharing of best practices. He says it has a lot of teeth, calls it a standard every company should adopt, and credits the President with running the meeting well.

Notable Quotes

“I would bet anyone that 1% increase in power usage corresponds to roughly 1% increase in GDP.”
Elon Musk, on why energy is the scoreboard for the AI economy

“If it’s on the ground you’re going to get somewhere between a fifth and an eighth of the solar power generation and you need enormous batteries. Whereas it’s essentially always sunny in space.”
Elon Musk, making the case for orbital compute

“We used to call them data centers. We call them data centers because that’s where we hold data. But the things that we’re building now, these are really SI factories, super intelligence factories.”
Jensen Huang, on why AI infrastructure is a factory and not a warehouse

“Even though you have it contained, you should never trust that it’s contained. And so you want to monitor it in real time.”
Jensen Huang, explaining Nvidia’s OpenShell and BlueField agent safety design

“Safety and advancing a technology is one and the same thing. It’s no different than when you have autopilot, not only is it advanced technology, it’s also safer technology.”
Jensen Huang, on safety and capability moving together

“In fact, it’s the largest payload delivered to orbit since Skylab.”
Elon Musk, on Starship’s first orbital flight carrying Starlink V3 satellites

“There’s going to be a huge amount of software being written by these models. And so all of the products that people use are going to be getting higher quality, cheaper over time.”
Tom Brown, Anthropic co-founder, on the quiet effect of Opus 5.5

“It’s a little bit like a steak knife where if you were served medium well, it cuts perfectly, but if you were served rare steak and it looks awfully like live meat, somehow it doesn’t cut.”
Jensen Huang, questioning how much judgment AI tools should have

“Just generally grading each other’s homework, which is a lot better than if people just grade their own homework, for obvious reasons.”
Elon Musk, on the joint SI safety declaration signed at the White House

Energy, chips, orbital compute, agent containment and a new safety pact, all in under half an hour. Watch the full conversation with Elon Musk, Jensen Huang and Tom Brown here.

Elon Musk, Jensen Huang and Anthropic’s Tom Brown With Gavin Baker at the America.gov Event: Superintelligence Factories, Orbital Compute, the Watts-to-GDP Rule and a Joint SI Safety Declaration