"So here's the number for this week, 830%," Peter Diamandis, founder of XPRIZE and Singularity University, said at the start of an episode of Moonshots with Peter Diamandis recorded on October 23, 2026. He described it as how fast SpaceX's business had grown in one year, and added that AI, not rockets, was now the company's largest business.
Before going further, Diamandis disclosed a conflict of interest. He is an early SpaceX investor, he said, with "a huge position personally," and he asked listeners to take his point of view with that in mind. "I'm a fan." He said he had been on CNBC the day before, arguing that "we watched SpaceX stop being a rocket company. It's now an AI company."
What the 830% measures
The figure comes from SpaceX's third-quarter results, released on October 20 for the quarter ending September 30. It describes one segment, not the whole company. SpaceX calls its AI business "Super Intelligence," or SI. That segment brought in $7.560 billion in the quarter, up from $813 million a year earlier, which works out to roughly 830% growth. Total company revenue grew 161%, to $14.575 billion.
The ranking Diamandis gave matches the filing. SI was just over half of all revenue, ahead of Connectivity, the Starlink internet business, at $5.248 billion, and Space, the launch business, at $1.767 billion. Rockets did come third. SpaceX credited the AI growth to cloud agreements, its acquisition of the coding tool Cursor, and adoption of its Grok models.
The same filing shows what that growth costs. The SI segment made $431 million in operating income for the quarter and spent $16.444 billion on capital investment, the money that goes into things like data centers and chips.
What Musk said he will build
Diamandis played a clip he had cut from Elon Musk's remarks on the earnings call. Much of it was measured in gigawatts. A gigawatt is a billion watts of electrical power, roughly the output of a large power plant, and AI companies now describe data-center size by how much power the buildings draw.
In the clip, Musk said SpaceX would have "maybe two and a half, three gigawatts" of compute online by the end of 2026, and that he was "fairly confident" of adding another 10 gigawatts in 2027. He argued that the limit on deploying superintelligence "is more in power and cooling and infrastructure than it is in chips." To keep that from slowing the company down, he said, SpaceX would aim to build 20 gigawatts of power and cooling infrastructure next year, more than the chips need. That left "a non-zero possibility, maybe it's a small possibility at this point," of having more than 20 gigawatts of compute running in 2027.
He also said SpaceX planned "the most powerful training cluster" by the end of the second quarter: 440,000 Vera Rubin computers, built on NVIDIA's next generation of AI hardware, to train Grok. Musk said that cluster would help make Grok "potentially the best model, but certainly to be extremely competitive."
Much of that compute may end up running Grok Bot. Musk called its growth the fastest he had "ever seen in my life" for a product, and said each user's usage rises over time instead of fading. SpaceX was already in early deployment with "some trusted customers" and expected wide enterprise deployment within a few weeks, so that companies could use it in teams "as though it were a virtual human." He said Grok Bot "may end up being what uses most of our superintelligence compute."
The rocket news was in the clip too. Musk said Flight 15 of Starship, which he expected the following month, would try to catch both the booster and the ship. If it succeeded, he said, it would be the first time an orbital launch vehicle had been fully recovered. He said Flight 14 had put 26 terabits per second of Starlink bandwidth into orbit while only about half loaded, which he equated to 10 launches of the smaller Falcon 9. With more satellites, each carrying much more capacity, he expected Starlink to have about 100 times its current bandwidth "probably within three years."
The hosts' arithmetic
Diamandis called the clip "a drop-the-mic moment" and added figures of his own. He said SpaceX was spending almost $20 billion a quarter, mostly on AI data centers, and bringing about 200,000 NVIDIA chips online each month. He said management claims each new data center pays for itself in under a year, and that the company's guidance is a $100 billion annual revenue run rate by December. The quarterly results put AI capital spending at $16.444 billion. The chip rate, the payback period and the run-rate guidance are not in that release.
Diamandis also estimated that 10 gigawatts in 2027 would be "a quarter, 25% of the entire amount of AI being brought on" that year. That is his own estimate, and he did not say where it came from.
Dave Blundin, founder of Link Ventures, began by imagining the new capacity in orbit: "if you put 10 gigawatts of inference in space." Diamandis stopped him. Next year's capacity would come from Colossus, "the data centers he's building on the ground right now." "He's got three gigawatts now. He's adding a minimum of 10," Diamandis said. "It's all terrestrial right now."
Blundin then argued that gigawatts understate how much intelligence the build-out will produce. He described the cluster as "half a million" Vera Rubins (Musk's figure was 440,000) and said each has about twice the performance of NVIDIA's Blackwell chips, making the cluster equivalent to about a million. He expected algorithms to become about 10 times more efficient and computing to move to photonic hardware, which uses light rather than electrical signals. On that reasoning, he said, the useful inference from 10 to 20 gigawatts would be "100 times bigger than even what that implies." He compared it to parameter counts, the number of adjustable values in a model. People used to rank models by them until, he said, Sam Altman told them to stop, because models were using their parameters more efficiently.
NVIDIA's own material shows how much these comparisons depend on the task. Its Vera Rubin NVL72 is a rack with 72 Rubin GPUs and 36 Vera CPUs. NVIDIA claims up to 10 times the tokens per megawatt of its GB200 NVL72 rack, but on one inference workload: the Kimi-K2 Thinking model with 32,000 input and 8,000 output tokens. Its projected fourfold cut in the number of GPUs needed for training assumes one specific job, a ten-trillion-parameter mixture-of-experts model trained on 100 trillion tokens within a month. NVIDIA labels these projections subject to change.
Blundin liked that Musk went beyond his prepared remarks. "When he goes off script like that, you get so much truth in such a short period of time," he said. "I bet his finance team is looking at him like, oh, here we go."
Starlink and spectrum
Diamandis called Starlink "the undersung hero" of the quarter. He said it had 13.6 million subscribers after adding 1.6 million. The results confirm both figures: 13.6 million consumer subscribers on September 30, up from 12.0 million on June 30. The results also say SpaceX's first revenue-generating Starship mission delivered 26 of its V3 satellites.
Diamandis said Musk had "just bought 800 megahertz of spectrum," meaning radio frequencies licensed for wireless service. The filing describes something less final: an agreement to acquire licenses associated with 800 MHz spectrum, still subject to approvals and expected to close in the first half of 2027. Diamandis also said the Federal Communications Commission had approved a 15,000-satellite constellation for direct-to-cell service, in which satellites connect straight to ordinary phones. Given a choice between T-Mobile, Verizon or AT&T and "a system that gives you coverage everywhere all the time," he predicted, Musk would "basically just wipe the slate and take over that business."
Terafab and power
Diamandis then turned to two other plans from the call: making chips and making power. He described Terafab as the SpaceX and Tesla chip factory in Texas, with Musk's goal of one terawatt of AI computing per year. According to Diamandis, Musk said on the call that the first test chips could work by mid-2027, with volume production around late 2028. Diamandis noted that Musk added, "I put big error bars on this because we don't know if we will actually achieve this or not."
For power, Diamandis pointed to an August 29 post on X. He said Musk wrote that SpaceX and Tesla were each building 100 gigawatts a year of solar production capacity as fast as possible, with natural gas bridging the gap while solar ramped up. Diamandis put the combined total at 200 gigawatts a year and compared it with roughly 600 gigawatts of solar worldwide in 2024. He did not say whether that figure meant the year's installations or total capacity.
Gas has its own problem. "Gas turbines are sold out," Diamandis said. In his account, SpaceX does "what Elon always does" when a supply chain is blocked, which is to build the missing piece itself. SpaceX is building its own foundry for turbine blades in Texas, he said, and says it will bring power plants online "within 18 months or faster."