Elon Musk has given his answer to the question people in AI keep asking: what is holding it back? "Elon also said that the real bottleneck for AI is no longer chips," Peter Diamandis, founder of XPRIZE and Singularity University, said on his podcast Moonshots. "It's power and cooling." Diamandis added that both areas are among Musk's strengths. He recalled that on August 29 Musk had written on X that SpaceX and Tesla were building solar production capacity as fast as possible, with natural gas to bridge the gap while solar ramps up.
A bottleneck is the narrowest point in a system, the one stage that caps how much the whole thing can produce. On the episode, recorded October 23, 2026, the other three hosts each rejected Musk's answer, each for a different reason.
Blundin: chips still set the limit
Dave Blundin, founder and general partner of Link Ventures, called Musk's framing "a little misleading." "The bottleneck to all of AI is actually chips," he said, "and right now it's actually packaging," the step that turns finished chips into usable modules. Diamandis said he disagreed with him.
Blundin granted Musk the present moment. About 15 gigawatts of chips are sitting idle, he said, because their owners cannot find data centers with power and liquid cooling, the systems that pipe coolant to the hottest hardware instead of blowing air over it. In Blundin's telling, the industry bought hardware designed for liquid cooling without first securing buildings that could provide it: "we overbuilt for liquid cooling without actually finding the liquid cooling." So for now, he said, Musk is "definitely right" that the constraint is liquid-cooled data centers.
Look past the current wave of construction, though, and Blundin sees chips binding again. Musk's planned Terafab chip plant, he said, will not come online fast enough to keep up with demand. He also pointed to memory, saying its prices are up tenfold. "You can't say like, oh, chips aren't the bottleneck anymore while memory is 10x higher than it should normally cost," he said.
He also suggested a way around the cooling shortage. Inference, the work of running an already trained model to answer requests, is increasingly handled by new hardware from companies such as Positron and Lightwave that is air cooled, and Blundin called those technologies "crazy fast." He expects inference to move largely to air-cooled machines so it can get online sooner, and eventually into space.
Wissner-Gross: bottlenecks are an economic fiction
When Diamandis offered to settle the debate, Alexander Wissner-Gross, a computer scientist and founder of Reified, cut in: "You're both wrong." ("Of course we are," Diamandis replied.) "The correct answer is there are no bottlenecks," Wissner-Gross said. "They're an economic fiction."
He acknowledged the obvious way to argue otherwise. Glance at the supply chain, he said, and you could make a "min-cut, max-flow" argument. The phrase comes from a result in network theory: the most that can flow through a network of pipes equals the capacity of its narrowest cross-section. Treat the AI supply chain as such a network and you end up arguing over which link is narrowest, "maybe it's this, maybe it's that."
Wissner-Gross said he does not buy that reasoning. Even if every physical input were held constant, he expects recursive self-improvement, meaning AI improving the methods used to build AI, "purely at the algorithmic layer" to keep making intelligence "too cheap to meter." Or, he said, the real limit could be something not on anyone's radar, such as the substrate, the physical material computing runs on. Diamandis read this as "capitalism coming in to solve problems." Wissner-Gross agreed: "hold everything else constant and let capitalism do its work. And there are no bottlenecks in the long term."
Ismail: the wrong metaphor
Diamandis then turned to Salim Ismail, founder of Open ExO and a general partner at Exponential Venture Capital. Ismail answered that all three were wrong, "just to be consistent across the board." In his view, "bottleneck" is the wrong way to talk about the problem at all.
He proposed Stewart Brand's pace layers instead: a set of nested layers, each with its own "metabolism." As Ismail described it, government moves slower than culture, and technology moves very fast. Grok, the AI model from Musk's companies, is improving quickly, he said, but chip design and fabrication, software, power infrastructure and data centers each have their own cadence and iteration time. Coordinating them is where he sees the difficulty.
His evidence came from a pilot project he described. "We get a couple of workflows running off to the side, we're running at hyperspeed and the rest of the organization can't cope," Ismail said. When one part runs at 100 times the performance of the rest, he said, "it breaks the whole system." Lining up the cadence across all those parts is, in his words, "a whole new skill set that we have to develop."
The framework appeared in Brand's 1999 book The Clock of the Long Now, and he and Paul Saffo revisited it in a discussion on January 27, 2015. Their account lists six layers from fastest to slowest: fashion, commerce, infrastructure, governance, culture and nature. In that ordering governance changes faster than culture, the reverse of how Ismail put it. Brand and Saffo describe the faster layers trying out innovations while the slower ones preserve continuity, and they say differences in pace can create stabilizing friction as well as conflict. Their account stresses how the layers interact rather than asking every layer to speed up equally: a fast-changing commercial practice still leans on durable infrastructure and institutions.
Diamandis took a blunter lesson from Ismail's answer: "Yeah, humans are the bottleneck for sure. There's no question about that."
The gas question
When the talk moved to how Musk plans to power all this, Wissner-Gross called the move into natural gas pipelines "a heck of a plot twist for the purported sustainable energy magnate." He said Musk is well positioned to become one of the largest owners of liquefied natural gas (LNG) pipelines in the United States, if not the largest.
Blundin said critics often point to this when they ask Ismail why he is such an Elon fanboy: Musk was "Mr. Green," and then needed natural-gas-powered data centers in Tennessee. Blundin defended him anyway. "He's Mr. Practical," he said, arguing that the benefit to humanity from those AI data centers far outweighs the carbon they emit. "And of course, they're all going to move to sustainable energy when it's practical."