Elon Musk, speaking remotely from Shenzhen at the iConnections TomorrowX Summit, November 19, 2026
Elon Musk used a wide-ranging, friendly fireside conversation at the iConnections TomorrowX Summit to put numbers on the AI infrastructure build across his companies. He said SpaceX, which now includes xAI, is bringing on roughly half a gigawatt of AI compute per month and expects to pass a gigawatt per month by the end of next year. He also said Wall Street estimates for the cost of 10 gigawatts of compute are too high. Musk spoke from Shenzhen, where he said he had arrived the day before.
Compute build-out: half a gigawatt a month, at roughly 60% of consensus cost
Musk said SpaceX is "right now, in rough terms, bringing on about half a gigawatt per month of AI compute," which he equated to roughly 220,000 GB300s plus associated CPUs, networking and cooling. He said he measures power at the power plant level, because that is the limiting factor, and that the figure comes to "a little over 500 megawatts" once turbine outages and peak cooling loads are included. "I'm not sure if any other organization is bringing on half a gigawatt of advanced AI compute per month," he said. He added that the pace should exceed a gigawatt a month of AI, "probably at that point superintelligence compute," by the end of next year.
Asked by the hosts about concern on Wall Street over how the 10 gigawatts of compute will be funded, Musk said most published estimates are "off by a factor of two, and sometimes more than a factor of two." Against the industry consensus number, he guessed his cost is "somewhere around maybe 60% of that number." He cited his record on capital efficiency: "I might be the best at capex efficiency, but if not I'm pretty close to being the best."
He attributed the advantage to doing the physical work himself. He compared conventional data center construction to a meme "where there's one guy digging a hole and eight people watching him dig the hole," and said, "I aspire to be the guy digging." He said he personally performs part of every task at the data centers, including pulling power conduit "the size of an anaconda," plugging in fiber and swapping compute nodes. SpaceX also runs an in-house team, which he called Army X, that breaks construction into tasks trainable in days. Recruits include former baristas, shelf stockers and In-N-Out employees. "If you get some aces out of In-N-Out who've been making epic burgers all day long, they're going to do a damn good job of network cables," he said. The Memphis site runs around the clock, and he said he sees "more or less the same productivity" at 2 a.m. as at 10 a.m.
Tesla Megapacks are part of the design. Musk said GPU loads swing widely and training runs are difficult without power smoothing "at scale at a wide range of frequencies." Without it, he said, "you'll break the crankshaft on the turbines." He called Tesla "unequivocally the best electrical engineering talent in the world."
Orbital compute: "Star Mind" described as simpler than Starlink V3
Musk gave an unusually specific description of the orbital AI satellite, which he called Star Mind. He said he does not understand why people assume it would be hard, given that SpaceX has "over 11,000 satellites in orbit, which is I think twice as much as the rest of the world combined." He called the Starlink V3 communications satellite, with multiple radio bands, roughly 30,000 lasers across the constellation and what he described as the most advanced phased array antennas ever made, "far more advanced than anything that any military has at any price."
By comparison, he said, the compute satellite "really just consists of a very large array of solar panels and a radiator." He put solar output at around 250 watts per square meter. He estimated the double-sided radiator, pointed away from the sun and facing space at 3 Kelvin, would shed about 1,200 watts per square meter, implying about five times as much solar area as radiator area. The compute module at the center would be "physically about a quarter the size of a Vera Rubin NVL72 rack" while delivering the full capability of one. It would link to the Starlink network by laser, and Starlink would relay data to the ground. "I'm not saying it's trivial, but it's not as hard as what we're already doing," he said.
On supply, Musk said SpaceX is building out solar cell and silicon fabrication and aims for 200 gigawatts a year of photovoltaics for space manufacturing. In space, he said, the full 200 gigawatts of nameplate power is usable. He noted that figure is roughly 40% of average US power usage, which he called "impossible" to add annually by conventional means on the ground. He said Starship's tonnage to orbit is the other requirement.
Terafab: test chip targeted for June 2027, volume production by late 2028
Musk said Terafab, announced earlier this year, is "still in its infancy" but has recruited a strong team and is hiring. He said the facility is intended to be the largest chip fabrication plant on earth and the largest building of any kind, ultimately about 100 million square feet. The plan is to start with relatively standard equipment, then push suppliers toward special versions of their machines. In some cases SpaceX may build its own tools.
He targeted a research fab in operation and a first test chip around June of next year, with volume production for phase one toward the end of 2028. Musk praised the chipmakers as "incredibly competent" but said their risk aversion is understandable given the capex tied up in each fab. "We're going to take a lot of risks, iterate fast, and ultimately I think have a much more efficient production process for chips than anyone else. That's our hypothesis," he said. He added that he is "cautiously optimistic" about a significantly better process by 2028 and did not want to jinx it.
Grok: training stack rewritten from scratch, Grok 4.8 in reinforcement learning
Musk was candid about xAI's early shortcomings. He said he started it as a small side project to give large-model researchers a home outside OpenAI, Anthropic and Google without distracting Tesla. He said he "didn't grow it with the care and rigor that I grew SpaceX and Tesla," and that it "didn't grow in quite the right way." He also called his year in campaigning and government "a side quest that went on a little longer than it should have," which brought some good "albeit at the cost of animosity from certain quarters." As a result, he said, he has had to rebuild xAI, which is now part of SpaceX.
The rebuild centered on discarding third-party tooling. Musk said the team is rewriting the entire training, reinforcement learning and inference stack from scratch, primarily in C and some C++, instead of using JAX and SGLang, which he described as "free software from Google, more or less," but "not as good as what they've got internally by far." His reasoning: "You just can't build world-class AI using mediocre freeware." Fewer abstraction layers mean higher performance and less of a black box. He said the rewrite is taking six to eight months, and "I think we might have the best training stack software in the world."...