Chamath Palihapitiya thinks the hard part of artificial general intelligence is already done. "I do think that we have achieved AGI," the venture capitalist said. "It cannot diffuse. And the reason it cannot diffuse has nothing to do with the technology. It has everything to do with the buildup." Until enough power plants and data centers exist for most people to use it, he said, "it's just something that exists in a lab."
Palihapitiya is the founder and CEO of Social Capital and a co-host of the All-In podcast. He made the argument as a guest on Moonshots with Peter Diamandis. The episode was recorded on October 20, 2026, and published the next day. AGI usually means AI that can match people across a wide range of intellectual work. Palihapitiya's point was less about the models than about everything they need to run: electricity, buildings, permits and the people who wire them. "Now, that has value," he said of AGI in the lab. "But I would not underestimate the difficulty."
A filing he called shocking
Palihapitiya's main evidence came from Anthropic, the maker of the Claude models, and its S-1, the registration statement a company files before selling shares to the public. He said Anthropic's S-1 had "just dropped today" and listed a risk he had never seen before: that "the hyperscalers will not give them compute." Hyperscalers are the giant cloud providers that rent out computing power. He called it "shocking" for "a company growing this fast and so powerful with all this money."
Why he started with electricity
Palihapitiya said his thesis began with a plain look at American electricity demand. The sources of growth are "relatively pedestrian": electric cars, fridges, televisions, laptops and data centers, in what is becoming a more electrified society. On the supply side, he said, there are "huge issues," even if every planned project gets permitted and built on time and on budget. Netting the two, he said, the United States is "still 1.4 terawatt hours short" by 2050, "effectively six California's worth of electricity." He concluded: "There is no world in which we catch up."
He dismissed nuclear power as the quick answer. Small modular reactors, or SMRs, are compact factory-built reactors often proposed for data centers. Their problem, he said, "is the S." Getting a small reactor approved takes the same effort as a large one, because developers still face the same state and local rules, objections and conflicts "whether the form factor is it goes under your daughter's bed or it sits, you know, over there and it's gigantic." The International Energy Agency's 2025 Energy and AI report expects the first SMRs to start serving data-center growth only around 2030.
Once he realized how short the country would be, Palihapitiya said, "you immediately abandon all of the rhetoric around climate change because you're like, just give me anything. Oil, nat gas, coal, wind, solar, nuclear." Solar and wind came out cheapest, with solar slightly more expensive. But solar was much faster: "It's two years. It's do no harm. It's plunk it on the ground and off you go."
From solar farms to data-center sites
Palihapitiya said he spent a lot of time buying solar assets, and that this "accidentally" led him into data centers. Through his dealings with NVIDIA and others, he saw that the build-out of computing capacity was "so upside down" because of the constraints around it. A data center, he said, is where "people's energy, ire, anger" all meet: permitting fights, local pushback, politics, resentment of "the oligarchy" and wealth inequality. About two years ago he started piecing together sites he thought would be valuable in a world where that capacity is the scarcest asset.
Peter Diamandis, the show's host and the founder of XPRIZE and Singularity University, put Palihapitiya's holdings at "nearly six gigawatts of power." A gigawatt is a measure of generating capacity, roughly the output of one large power plant. Dave Blundin, founder of Link Ventures, said the industry is looking for about 100 gigawatts in total by 2030 to run all of AI, so six is "a big chunk," and assembling it in two years was "just mind blowing." "It wasn't cheap," Palihapitiya said.
Blundin argued that power will matter more over time. Today, he said, builders treat electricity as about 10% of a data center's cost, with the other 90% going to NVIDIA. He expects that to invert as new chip designs undercut chip prices, while power costs stay where they are.
Why owning the sites isn't enough
Even with the assets in hand, Palihapitiya said, turning them on is "incredibly complicated." He listed the funding cycle, the supply chain and a shortage of licensed workers. "You have to go and find like a licensed master electrician. They're very, very hard to find. They cost a lot of money." Put together, he said, "we are probably going to see a massive, massive, massive compute shortage in the next three or four years." Blundin agreed.
The IEA identifies similar bottlenecks in its report. It estimates that infrastructure constraints could delay around 20% of planned data-center projects unless addressed. It points to transmission lines that take four to eight years to build, queues for equipment, and data centers crowded into the same areas.
Blundin cited Elon Musk as saying 15 gigawatts' worth of chips are waiting for liquid-cooled data centers to go into, "a massive asset depreciating idly." Palihapitiya said it is worse than that. Older chips from generations before NVIDIA's Vera Rubin are also sitting unused in inventory, he said. And liquid cooling has its own problems. In his account of a deal between the data-center company Nscale and Microsoft, there were two big coolant leaks in a row, and Microsoft triggered the contract's liquidated-damages clause, a penalty agreed in advance, and walked away.
"Everybody thinks this stuff is easy," Palihapitiya said. He blamed part of the difficulty on 20 years of steering young people toward "TikTok and social media and video games" instead of the trades: "The ability to actually build these things is gone." Blundin said the work goes well beyond the trades. It also needs regulatory work, negotiation and planning; "it really involves every human skill." He said it "drives me nuts" to hear people conclude that math and teaching are "cooked" and that they might as well watch late-night TV, when "a bigger opportunity" has opened up.
The co-op problem
Palihapitiya said people are "wildly, wildly wrong in the short term" about how fast AI will spread, though right in the long term, because of these constraints. His example was a regional power cooperative. Aging infrastructure, he said, pushed utilities in several states to organize into co-ops. One of them runs "right through the spine of America," and its northern states have room for a great deal of data-center capacity that would ease the limits on spreading AI.
He and a partner bought sites there and found they were part of the co-op. Inside it, he said, everything comes down to personal relationships, and the large members in the south decide how it works. "It's like an organized mafia," he said. In the north, "you can't build. You can't get the right offtake agreements," meaning the contracts to buy the power. In his telling, Oklahoma Gas and Electric "calls the shots." He did not name the co-op.
"Man, all that complexity just smells like opportunity knocking," Blundin replied.
Space won't rescue most companies
Blundin noted that Musk wants much of this computing to move into orbit. But "even the most wildly aggressive estimates put a couple percent in space within a decade," he said, and over the next 30 years most of the build-out will happen on the ground.
Palihapitiya said it may make sense for Google, Microsoft, Meta and SpaceX to put all of their computing in space. Everyone else faces a different problem, which he linked to Palantir chief executive Alex Karp's argument for AI data sovereignty: keeping a company's data under its own control. He used Humana as a hypothetical ("I'm just making up a company," he said). Such a company can't launch its own data center into orbit and would wait a long time for space capacity. Data sovereignty would also push it out of shared cloud servers, toward at least a virtual private cloud, a walled-off section of a provider's cloud, or "an air gap facility that's our own," physically cut off from outside networks. Companies like that will build on the ground while they wait, he said, and run into all the same obstacles. "That's probably going to dictate the winners and losers more than anything else over the next five years."
How ordinary people can "get long electrons"
Diamandis asked where an average person should look to get in on energy. Palihapitiya said the utilities' monopoly on generating power is breaking up. The simplest route is residential solar, followed by commercial and industrial solar, which he said are easy to permit and fast to build. Pair the panels with a battery "and now you're in the currency game. You're effectively your own little money printer. You use what you need. You sell back into the grid."
The second route is distributed computing: joining pools of scattered machines, such as Venice, that put their processing power to work. "I'm not involved in any of these," he said. He also described SPAN, which he said had done a deal with NVIDIA, as offering storage, a battery and a graphics processor "all in like this like unit that you can just deploy in your house."
SPAN's April 13, 2026 announcement of XFRA describes something closer to a network that homeowners host than a box they buy. The company's electrical panels use the spare capacity of a building's electrical service to run liquid-cooled NVIDIA RTX PRO 6000 Blackwell processors at homes and small commercial buildings. Those processors would handle tasks such as running AI models and cloud gaming. SPAN named NVIDIA and the homebuilder PulteGroup as partners. The proposed benefits for homeowners include the panel, battery backup, optional solar and discounted electricity and internet. First deployments were planned for later in 2026, with a goal of gigawatt-scale capacity in 2027.
Palihapitiya summed up his advice: people "should find the cheapest and most technically simple method that they can manage to get long electrons." In investor slang, being long something means owning it so that you gain when its value rises.
Blundin added two tips. First, these projects are regulated at local, state and federal levels, so anyone with political connections at any of those levels can get involved, even without knowing the Kushners. Second, he said most companies were "frozen out of compute" because quant trading firms, OpenAI and Anthropic bought it all up. Now that companies realize they need huge numbers of AI agents to compete, an executive could be the first to tell colleagues, "we need to unlock a billion dollars to do a prepay on Chamath's next power supply." That way, he said, "we won't be frozen out" next time. "You don't have to know anything about AI," he said. "You just need to know where the capital is and unleash it for the build out."
The counterargument: demand and smaller models
Alexander Wissner-Gross, a computer scientist and the founder of Reified, accepted part of the premise. At the frontier of capability, he said, supplies including energy are a limiting factor. He was less sure about demand. He recalled that OpenAI "infamously lurched toward and then away from the consumer": last year, he said, it tried to get consumers to pay for reasoning models, which work through problems in many steps and use more tokens (the chunks of text a model reads and writes). "Consumers just didn't want it," he said. They had no good use for it at the time. If new capabilities can't be "metabolized by the economy," he argued, no amount of energy "showering down on the frontier labs" will make "a lick of difference."
His second point was distillation, which trains smaller and cheaper models to reproduce what larger ones can do. Models are getting smaller and using less energy over time, he said. He offered a thought experiment, "arguendo": suppose no one builds another data center, power plant or watt of capacity. Wait a few years, he said, and "just through the pure power of math and recursive self-improvement," he would expect models that use "a fraction of the current energy footprint" and deliver "many orders of magnitude better capabilities."
"Welcome to the singularity," Diamandis said.
Blundin's foot race
Blundin closed by calling it "a foot race" between self-improving models that get lighter, faster and smaller and an explosion of ideas for using that intelligence. He said he is "firmly in the camp that we will not run out of ideas on how to use every bit for decades."
His example was Will Thompson, a Harvard junior he had mentioned on an earlier show. Thompson got access to what Blundin called a 1,200-tokens-per-second capability from Cerebras, a chipmaker that builds hardware for very fast AI responses. Blundin said users can reach a "crazy fast" mode for OpenAI's Astra model by going through Codex. "It's insanely expensive. So try it for like 10 minutes, but you will never go back." Diamandis called it "your first dose of heroin."
OpenAI's published figure for its earlier preview is lower. On August 13, 2026, the company previewed a Cerebras-powered Ultrafast mode for GPT-5.6 Sol for selected API customers, promising up to 750 output tokens per second and up to 14 times its standard speed. OpenAI pitched it for coding, research and incident response, where shorter waits mean more attempts in a workday. For Blundin, that appetite was the point: "We have a long way to go before people run out of ideas on how to use, how to use more and more and more AI."