Elon Told the G20 the AI Boom Hits a 15 Gigawatt Wall Next Year

Elon just joined G20 technology ministers by video and said the quiet part out loud. The AI boom is already a prosperity shock measured in tens of trillions of dollars. Software written by people loses to software written by machines in 12 to 18 months. A billion humanoid robots, in his conservative case, outproduce every human on Earth inside a decade. Then he named the thing that actually decides who gets that future. Power. He said there is already a crisis of electricity, and that the consensus among the analysts who live in this market is at least a 15 gigawatt shortfall in 2027 for AI chips. Not 2035. Next year.

The session was Day 1 of the G20 Innovation Ministerial at the Carolina Inn in Chapel Hill, North Carolina, on Sept 1, 2026. White House science adviser Michael Kratsios ran the conversation. Elon appeared virtually, then former White House AI czar David Sacks, then Bob Mumgaard, co-founder and CEO of Commonwealth Fusion Systems. Farzad posted the session the same day. The transcript is a briefing to the governments that still think they can regulate this like a drug or an airplane.

Default legal, or default illegal

Kratsios opened with the question every minister in that room is actually asking. What separates a country that turns a breakthrough into a deployed product from a country where progress stalls?

Elon's first answer was not a chip, a model, or a factory. It was the legal default. New things must be default legal, not default illegal. He pointed at the European Union and said the regulation level there is extraordinarily high, that things are generally default illegal, and that this does not stop new technology. It slows it down quite considerably. AFP put EU ministers from France, Italy, and Germany, plus EU tech chief Henna Virkkunen, in the Chapel Hill room. That attendance is reporting, not something Elon named in the clip.

His second answer was capital for the young companies, not protection for the old ones. He compared startups to saplings in a forest. Most countries, he said, over-support the large existing trees and do not give enough to the saplings. The large trees do not need the help. Startups do. The system should be biased toward the small trees. That is rarely the case, because the large companies have the access.

Adoption, in his view, is the same fight. Lean forward. Try the new thing. Tesla's self-driving, he said, is already a tremendous boon to users. Humanoid robotics will be an incredible change. If a government sits in the past, it does not get the productivity. It gets the delay.

$20 to $30 trillion, then Stockfish

Then he put a number on digital AI alone. He said AI will probably increase the global economy by 20 to 30 percent. He called that a rough estimate, meaning on the order of $20 to $30 trillion per year. That is his figure, not an IMF forecast and not a US filing. Treat it as Elon's scale, not a measured print.

The timeline under that number is shorter than most ministers budget for. He said AI will be able to do anything digital, anything that does not require shaping atoms by hand, probably by the end of next year. Software is already incredibly good. His prediction is that it goes Stockfish-level good. Stockfish is the chess engine that beats the world's best players easily. You can run it on a phone and beat Magnus Carlsen. In 12 to 18 months, he said, it will be impossible for a human to compete with AI at writing software. AI will crush all humans at software. It will be extremely good, possibly that same level, at all forms of engineering and anything digital.

Physical work is slower. Digital copies. Physical requires a global supply chain and a lot of atoms. That is why factories take longer than models. It is also why the ministers who only regulate software will miss the second half of the shock.

A billion robots, five times a human

Kratsios cited a stat the group had already heard: in the four years since ChatGPT launched, more than a billion people around the world are already using AI. He asked where robotics goes next, and how fast, given that talk of automated factories in the first Trump administration is now about ten years old.

Elon gave them a formula. The usefulness of a general-purpose humanoid is AI software quality, times the AI chip in the robot, times electromechanical dexterity, especially the hands. All three, he said, are improving exponentially. Usefulness is those three multiplied. Then the robots start manufacturing the robots. Recursive. It starts slow. Then it explodes.

Ten years from now, he said there will be well over a billion humanoid robots. Productivity per robot, probably five times that of a human. He called that conservative. He said he would put serious money on at least a billion robots in ten years, at least five times the output of a human, meaning those robots would be more productive than all humans combined.

If he is even directionally right, the labor math every G20 finance ministry still uses is already wrong. A country that blocks the robots does not protect workers. It exports the output to the country that built them.

The 15 gigawatt hole

Data centers have been a political fight in the United States for the last six to eight months. Kratsios asked where we are on the curve, how much has been built versus how much is needed, and how governments should think about power and compute over the next few years.

Elon did not soften it. There is quite a crisis of power. Follow the AI topic on X, he said, which is where almost all of the AI discourse takes place, and the consensus is a significant power shortfall next year. The number he attached to that consensus is at least a 15 gigawatt shortfall of power in 2027 for AI chips. Elon did not name the analysts. Treat 15 gigawatts as Elon's cited consensus, not a published grid study. Two days earlier, on Aug 29, 2026, he posted on X that the consensus estimate is ~15 GW of AI compute produced in 2027 cannot be turned on in 2027 — idle compute, a slightly different claim than a power shortfall. No Wood Mackenzie, EIA, or Goldman print was found that says "15 GW shortfall in 2027." Goldman (May 20, 2026) has US data-center power demand rising from 31 GW in 2025 to 66 GW in 2027.

The curves are why. AI chip production, in his telling, is rising on the order of 40 to 50 percent a year. Power available outside China is rising at like 10 to 20 percent a year. The faster line overwhelms the slower line. Those are Elon's curves, not a matched public series. TrendForce (April 8, 2026) had high-end GPU shipment growth around 26 percent for 2026. Nvidia data-center revenue is running far hotter; that is dollars, not chip units. EIA's August 11, 2026 STEO has US generation up about 1.8 percent in the first half of 2026, not 10 to 20 percent. He said there are already challenges with power before next year, which is why Google and Anthropic and many other companies are leasing compute from SpaceX. SpaceX could turn AI on faster than anyone else so far, he said, by constructing its own power plants. That is the only way they were able to do it.

That lease claim is Elon at the mic. SpaceX filings this year describe large Colossus leases with Anthropic and Google, and a third reported tenant, Reflection AI. Anthropic has been reported at about $1.25 billion a month for Colossus 1, with 90-day termination. Google has been reported at $920 million a month for about 110,000 GPUs from October 2026 through June 2029; as of this video, September is still the reduced-fee ramp. Reflection has been reported at $150 million a month from July 1. Elon did not recite those dollar figures at G20. He recited the reason. He built the power.

China, he said, has a tremendous amount of electricity. GPU export bans stop the latest chips from going into Chinese data centers. The live question for every other country is electricity growth outside China, and that growth is currently a significant shortfall relative to AI chip production. His offer to the room was blunt. If you want AI data centers, construct a lot of power and offer it to AI companies. Tax them. Charge reasonable fees. The shortfall is an import opportunity for any country that can actually turn turbines on.

Sacks: $800 billion this year, $1.4 trillion next

Sacks followed as the policy voice. Kratsios introduced him as former White House AI czar and co-chair of the President's Council of Advisers for Science and Technology. Sacks has left the White House post. Press around the ministerial said he still advises the president. That is reporting, not his line in the clip. He said the AI boom is already here, and that the job is to nurture it rather than hamstring it.

The capex numbers he used are the ones that should be on every screen. Roughly $800 billion is being invested in the US this year on capex related to building out AI infrastructure, the compute capacity. That number is expected to be $1.4 trillion next year. Every time he sees the numbers, they keep getting revised up. He said we have not seen a buildout of this kind of infrastructure since the railroads in the 1800s, and that it is actually even bigger than the fiber buildout of the internet. A new industrial revolution, in his words, if the United States harnesses it to re-industrialize, upgrade the power grid, and bring back manufacturing in cutting-edge areas.

Those are Sacks's figures at the ministerial, not a Treasury print. Hyperscaler consensus for 2026 has been running near $800 billion. Goldman (Aug 7, 2026) cites that consensus at about $794 billion, then publishes its own preferred AI-related print of about $1 trillion global and $581 billion US — a different definition, not the $800 billion minus a rest-of-world slice. Morgan Stanley in May put the five hyperscalers at about $805 billion this year and about $1.1 trillion in 2027, then raised 2027 toward $1.2 to $1.3 trillion; $1.4 trillion shows up as a 2028 figure in that lane and as a 2027 upside case. Keep Sacks's $1.4 trillion as his G20 number, at the high end of the 2027 range.

Then he drew the regulatory line Elon had already implied. He said the United States is not missing laws. Fraud, discrimination, privacy violations, defective products, copyright infringement, non-consensual imagery, and many other harms are already illegal under existing statutes. If government builds, it should build on the agencies it already has: transparency, disclosures, audits. Antitrust exists if the market tips into a monopoly or a duopoly. He said it is a little early to call that the current situation.

The approach he said would be a disaster is an FDA for AI. New drugs and airplanes can take five years plus to approve. AI companies, he said, are rolling out new models every two months. A pre-approval agency becomes a DMV for AI, with new models stuck in a queue. Silicon Valley's secret sauce, boiled down to one principle, is permissionless innovation. A college dropout in a dorm room. Two PhD students in a garage. He said two trillion-dollar companies were founded that way, Meta and Google, and that those founders would not have navigated a bureaucratic maze in Washington to get the product approved. Innovate first. Regulate later, once it is successful.

On data centers he echoed what he said President Trump had said. If done right, data centers bring electricity costs down, not up, because the AI companies generate net new power. If they compete with local residents on the existing grid, prices can go up. If they generate their own power behind the meter and can send excess back, costs can fall and the grid can finally get upgrades. He said we are seeing hundreds of thousands of new construction and skilled-trade jobs, and that counties that embrace data centers are seeing reduced property taxes. His example: residents of Loudoun County, Virginia, paying $6,000 less on average in property taxes because the AI companies pick up the tab. That $6,000 is Sacks's line at G20. County-level analyses of Loudoun's data-center tax base put the avoided residential bill in a similar neighborhood, around $5,800 in one 2026 industry-funded study, depending on the method. It is a political number. Check it.

He also pushed back on what he called fake news. He said the president's executive order of December 11, 2025 — EO 14365, Ensuring a National Policy Framework for Artificial Intelligence — expressly said the federal government should not preempt state decisions on data centers. That tracks a carve-out in Sec. 8(b): the proposed federal framework should not preempt otherwise lawful state laws relating to AI compute and data-center infrastructure, other than generally applicable permitting reforms. The rest of that order is about challenging state AI laws. Kratsios, in the clip, said the president put out a statement on data centers the day before, Aug 31. That is the Truth Social post that communities rejecting data centers want to end up "backwards and poor" and should "let Data Reign." ABC11 later quoted Kratsios citing that post as leaving it a community decision. The post is not read on camera. Do not quote Trump as saying "community decision."

Kratsios, in this video, refers to "Carolina principles" the G20 nations have "agreed to." That overstates a published G20 agreement. Reuters and The Globe and Mail: US-promoted, nonbinding framework — reserve new regulation for novel considerations, do not stand up a new AI regulator. Kratsios told reporters China signed. He did not publish the document. Until the White House posts the paper, treat the name and the posture as real, and the fine print as unverified.

Fusion is the long unlock, power plants are the short one

Mumgaard went last because the room had just been told the constraint is electricity, and fusion is the fantasy version of more electricity. He is building the non-fantasy version. CFS is the largest private fusion company by his own description. He said the fusion industry now has over 50 companies and over $14 billion of private investment, up from essentially nothing ten years ago. CFS is the largest of those, about $4 billion. That $4 billion matches what the company has said it has raised, including a later $1 billion round. The $14 billion industry figure is Mumgaard's at the mic. Confirm it against the Fusion Industry Association survey before treating it as a census.

The facility outside Boston, SPARC in Devens, Massachusetts, is the machine he said would have been a Manhattan Project-scale government site in the last generation. He said they are getting ready to start first commercial power plant construction in Virginia, up the road from the ministers, and that could put fusion on the grid in the early 2030s. That is CFS's target and a site/permit path, not electrons on the meter and not a notice to proceed as a finished plant. CFS has already announced the Chesterfield County, Virginia site, a joint path with Dominion Energy, and offtake with Google and Eni. Early 2030s is the company's target, not electrons on the meter today.

The policy lesson he wanted in the communique is the same one Elon and Sacks used. Do not invent a brand-new regulator for a baby industry. The United States and the UK, he said, licensed fusion against the closest existing analog, particle accelerators, so investors could put billions in without waiting to find out whether a new statute would kill the plant. Public-private partnerships let companies use publicly funded expertise without the government trying to build the first generation of plants itself, a path he called a path to nowhere.

If you plan for fusion in the 2050s, he said, you get fusion in the 2050s. Plan for the most aggressive speed the tools allow. Other countries, he said, are already shortening their fusion timelines year over year as they watch the US private sector.

What the ministers actually have to decide

Put the three talks on one page and the business model is obvious. Elon is selling a world where digital work is done by models by the end of 2027, physical work is done by a billion robots in ten years, and the scarce input is electricity, not talent. Sacks is selling a legal order that keeps that build legal by default, financed at $800 billion this year and $1.4 trillion next, with no FDA sitting on model releases. Mumgaard is selling the only new energy source in the room that could, if it works, refill the 15 gigawatt hole in the 2030s instead of the 2050s.

The countries that treat new technology as default illegal will still get AI. They will get it late, and they will rent it from the countries that treated it as default legal and built the power plants first. Google and Anthropic are already writing checks to SpaceX for that reason. A third reported tenant, Reflection AI, is in the same lane. None of those leases had been cancelled as of Sept 1, 2026. The G20 can issue a principle. It cannot print gigawatts.

This Exclusive is from the long-form at https://www.youtube.com/watch?v=inEO2FaU20E.