Musk says superintelligent compute capacity will be in long-term shortage, SpaceX plans to expand compute to 13 gigawatts by end of 2027

In the early hours of October 21 Beijing time, SpaceX (SPCX.O) held its Q3 2026 earnings call. The company's CEO Musk said at the meeting: "I think there will be a shortage of superintelligent compute, and this shortage will last for a long time."

Superintelligence is SpaceX's name for its AI business, which mainly includes renting out compute externally, as well as AI products such as Grok. Musk also said at the meeting that he would be surprised if, by the end of 2027, SpaceX had less than 13 gigawatts of compute in use. As of the end of September 2026, that scale was 1.7 gigawatts (measured by data center power capacity). In other words, to achieve this expectation, SpaceX's compute would need to expand to more than 7 times the end-of-September level in just over a year.

SpaceX's Q3 2026 results data showed quarterly revenue of $14.575 billion, up 161% year on year, above market expectations; of that, superintelligence business revenue was $7.56 billion, surpassing the connectivity business led by Starlink. Affected by this, on October 21 US Eastern Time, SpaceX's stock closed up 7.79%.

In Musk's view, compute demand will exceed supply for a long time, and what currently constrains supply is more power, cooling and infrastructure, with chips secondary.

Supply falls short of demand

SpaceX's compute business is currently oversubscribed, with external customers wanting to rent more and its own products also needing to use more.

Musk said at the earnings call: "Every major cloud service provider, and dozens of small companies, are constantly approaching us, wanting to rent more compute." The company's Chief Financial Officer Bret Johnsen also said that many customers are seeking compute from SpaceX, and the company is actively expanding supply to meet internal and external demand.

Over the past year, SpaceX's compute scale increased from 0.6 gigawatts to 1.7 gigawatts, and superintelligence business revenue jumped accordingly. According to Johnsen, in Q3 2026, revenue from this business increased by about $5 billion quarter on quarter and by about $6.7 billion year on year, mainly from the gradual ramp-up of cloud service agreements.

Compute used by SpaceX's own products is also tight. Musk said that to free up more compute for training models, the company had tried to hand some inference tasks (i.e., computation when models respond to user requests) to other cloud service providers. Because compared with training, inference tasks are easier to split onto other platforms for running. He also believes that what ultimately consumes most of the company's superintelligent compute may be the agent assistant Grok Bot.

Grok Bot was launched in August. Musk said it is the fastest-growing product he has ever seen; it is currently in early use among some customers and will be deployed broadly to enterprise customers within weeks. According to the company's President and Chief Operating Officer Gwynne Shotwell, earlier in October, Grok Bot's paid weekly active users had reached 800,000, more than doubling in a month, with enterprise customers being the fastest-growing category.

Musk estimates that with the expansion of compute plus demand for Grok Bot, by the end of 2027, SpaceX's annualized recurring revenue (full-year revenue converted from the revenue of the month at year-end) will be between $500 billion and $1 trillion. Johnsen mentioned at the earnings call that the target for this figure by the end of this year is $100 billion.

"We believe the value of compute per watt will rise over time, not fall," Musk said at the earnings call. As AI software keeps improving and models get better, each watt of compute can produce more "useful intelligence," and compute becomes more valuable, "which is why you see that even old models like the H100 (an AI chip launched by Nvidia in 2022) now rent for more than when they were first launched."

"So I think the smart thing to do is sign short-term contracts and bet on the rapid growth of useful intelligence per watt," Musk said. According to him, SpaceX rents out compute at prices close to spot market prices and does not sign long-term contracts; based on renewals, the price per watt of compute is higher than before. If compute indeed becomes more valuable, signing long-term contracts would amount to locking the price at the current level.

Documents previously made public by SpaceX show that customers renting its compute include AI companies Anthropic and Google. Although both agreements run until 2029, after the initial stage either party can terminate with 90 days' notice.

Johnsen also said at the earnings call that current compute prices are slightly above the upper end of the $30 to $50 per watt range mentioned at the company's Q2 2026 earnings call; at such prices, newly deployed compute can pay back in less than a year, and operating cash flow can support expansion. Musk believes that in 2027 the payback period for compute investment may shorten to 6 months, although this is not "certain to happen."

Power bottleneck

Musk said at the earnings call that more and more people are beginning to realize that deploying superintelligence is constrained more by power, cooling systems and infrastructure than by chips. The compute scale of AI data centers is measured by power capacity, and every additional gigawatt of compute requires corresponding power generation and cooling facilities to keep up.

According to Musk, SpaceX plans to put into use by the end of Q2 2027 a training cluster composed of 440,000 Nvidia Vera Rubin chips. In addition, he is also advancing a self-built chip factory but admits the company is still "exploring" in this area. The timetable he gave is to use the research fab in Austin to produce the first working test chip in May or June 2027, and to have the planned chip factory Terafab achieve mass production around the second half of 2028.

On the power side, because connecting to the grid often takes years, many AI data centers choose to build their own power sources using gas turbines, which has made gas turbines scarce. Gas turbine manufacturer GE Vernova (GEV.N) data shows that as of the end of the second quarter of 2026, its gas-fired power generation equipment had combined orders in hand and capacity reservation agreements (customers paying to lock in future production slots) totaling 116 gigawatts, while its planned annual gas turbine production capacity will only rise to 20 gigawatts in the third quarter of 2026. The company's CEO, Scott Strazik, said at the July earnings call that negotiations with customers already concern deliveries in 2032 and beyond.

Musk also said that gas turbine suppliers told SpaceX that delivery of blades and guide vanes lags other complete-machine components by 12 to 18 months, and that casting is a "very difficult bottleneck" among them. Blades rotate at high speed with the rotor, while guide vanes are stationary and direct the flow of gas; both directly withstand the scouring of high-temperature gas and must be made by precision casting of superalloys, and few manufacturers can supply them stably.

U.S. aviation parts manufacturer Howmet Aerospace (HWM.N) is one of the main suppliers of such blades. The company's Executive Chairman and CEO, John Plant, said at the August earnings call that Howmet Aerospace holds more than 50% of the global market for gas turbine blades used in power generation, and that capacity at the company's new plant in Japan is already essentially fully booked.

According to Plant, the new capacity will only be released gradually from 2028 to 2030, and the company has also had to reserve capacity from some machine tool suppliers. Howmet Aerospace Chief Financial Officer Patrick Winterlich attributed the growth of the gas turbine business to rising power demand, especially demand for gas-fired power generation from data centers.

Right, Musk's solution is still for SpaceX to cast blades and guide vanes itself. After he publicly announced the plan at the end of August, Howmet Aerospace's stock fell more than 7% on the first trading day afterward. At the aforementioned earnings call, Musk explained that the casting capability SpaceX developed for rocket engines can be used to produce these components, "so we can get power connected 12 to 18 months earlier than anyone else." He also said that in addition to gas turbines, SpaceX hopes to use solar, wind, and all the power it can get.

Johnsen said that by the end of October, SpaceX's computing capacity is expected to approach 2.2 gigawatts. On that basis, about 0.5 gigawatts would be added in October alone, compared with about 1.1 gigawatts added in the previous year. Musk also said that building power, cooling facilities, electrical equipment, batteries, and networking to support the large number of new chips put into operation each month is "simply crazy," but the pace of deployment in 2027 will accelerate further.

Musk's plan is that by the end of 2026, SpaceX's computing capacity will reach 2.5 to 3 gigawatts. In 2027, SpaceX is targeting 20 gigawatts of new power, cooling, and infrastructure. His view on this is: "Hard, but not impossible." At least, he is "very confident" about adding 10 gigawatts of computing capacity in 2027.

Of course, related capital investment is also increasing. In the third quarter of 2026, SpaceX's capital expenditure was $19.593 billion, and Johnsen expects it to reach $35 billion to $40 billion in the fourth quarter of 2026. In addition, specific computing capacity deployment is demand-oriented and advances according to contract schedules; funding comes not only from operating cash flow but also from investment-grade credit and multiple debt funding pools.

Whether these plans can be fulfilled on schedule is something Musk himself has left room on. He admitted at the aforementioned earnings call that he does not expect everything to go smoothly.

This article comes from the WeChat public account "Economic Observer," author: Zheng Chenye, published by 36Kr with authorization.