Why does Elon Musk believe that there will be a long-term shortage of computing power?
In the early morning of October 21 Beijing time, SpaceX (SPCX.O) held its Q3 2026 earnings call. Elon Musk, CEO of the company, stated at the meeting: "I believe there will be a shortage of superintelligent computing power, and this shortage will last for a very long time."
Superintelligence is the name SpaceX uses for its AI business, which mainly includes external computing power leasing and AI products such as Grok. Musk also stated at the meeting that he would be surprised if SpaceX's computing power put into use by the end of 2027 does not reach 13 GW. As of the end of September 2026, this figure stood at 1.7 GW (calculated based on the power consumption capacity of data centers). This means that to meet this expectation, SpaceX's computing power needs to expand to more than 7 times the level at the end of September in a little over a year.
SpaceX's Q3 2026 earnings data shows that the company's revenue in the quarter was 14.575 billion US dollars, a year-on-year increase of 161%, which exceeded market expectations; among which the revenue of the superintelligence business reached 7.56 billion US dollars, surpassing the Starlink-dominated connectivity business. Affected by this, SpaceX's stock price closed up 7.79% on October 21 Eastern Time.
In Musk's view, computing power demand will outstrip supply in the long run, and the current constraints on supply are more from power, cooling and infrastructure, with chips being a secondary factor.
Supply Falling Short of Demand
SpaceX's computing power business is currently in a state of supply shortage, as external customers want to rent more, and the company's own products also need to consume more computing power.
Musk said at the earnings call: "Every large cloud service provider, as well as dozens of small companies, are constantly reaching out to us, wanting to rent more computing power." Bret Johnsen, the company's CFO, also said that many customers are approaching SpaceX for computing power, and the company is actively expanding supply to meet internal and external demands.
Over the past year, SpaceX's computing power scale has increased from 0.6 GW to 1.7 GW, and the revenue of the superintelligence business has jumped accordingly. According to Johnsen, in Q3 2026, the revenue of this business increased by about 5 billion US dollars quarter-on-quarter and about 6.7 billion US dollars year-on-year, mainly driven by the gradual volume release of cloud service agreements.
The computing power used by SpaceX's own products is also tight. Musk said that in order to free up more computing power for model training, the company once tried to offload part of the inference tasks (the calculation when the model responds to user requests) to other cloud service providers. This is because compared with training, inference tasks are easier to split and run on other platforms. He also believes that Grok Bot, the intelligent agent assistant, may eventually consume most of the company's superintelligent computing power.
Grok Bot was launched in August. Musk said it is the fastest-growing product he has ever seen. It is currently being used by some pilot customers, and will be widely deployed to enterprise customers within a few weeks. According to Gwynne Shotwell, President and COO of the company, in early 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 segment.
Musk estimates that with the expansion of computing power and the demand from Grok Bot, SpaceX's annualized recurring revenue (the full-year revenue converted from the revenue of the month at the end of the year) will be between 500 billion US dollars and 1 trillion US dollars by the end of 2027. Johnsen mentioned at the earnings call that the target for this figure by the end of this year is 100 billion US dollars.
"We believe that the value of per-watt computing power will rise over time, not fall," Musk said at the earnings call. As AI software continues to improve and models become better, more "useful intelligence" can be generated per watt of computing power, making computing power more valuable. "That's why you can see that even old models like the H100 (the AI chip launched by Nvidia in 2022) now have higher rental prices than when they were first released."
"So I think the smart move is to sign short-term contracts and bet on the rapid growth of useful intelligence per watt," Musk said. According to him, SpaceX rents out computing power at rates close to spot market prices, and does not sign long-term contracts; from the perspective of renewals, the price per watt of computing power is higher than before. If computing power does become more and more valuable, signing long-term contracts would mean locking the price at the current level.
Documents previously disclosed by SpaceX show that its computing power rental customers include AI company Anthropic and Google. Although both agreements are set to expire in 2029, after the initial phase, both parties can terminate the agreement with a 90-day prior notice.
Johnsen also said at the earnings call that the current computing power price is slightly higher than the upper limit of the $30-50 per watt range mentioned by the company at the Q2 2026 earnings call; at this price, the newly deployed computing power can pay back its cost in less than a year, and operating cash flow can support expansion. Musk believes that the payback period for computing power investment in 2027 may be shortened to 6 months, although this is not "guaranteed to happen".
Power Supply Bottleneck
Musk said at the earnings call that more and more people are beginning to realize that the deployment of superintelligence is more constrained by power, cooling systems and infrastructure than by chips. The computing power scale of AI data centers is measured by power consumption capacity, and every additional 1 GW of computing power requires corresponding power generation and cooling facilities to keep up.
According to Musk, SpaceX plans to put a training cluster composed of 440,000 Nvidia Vera Rubin chips into operation by the end of Q2 2027. In addition, he is also pushing forward the construction of self-owned chip factories, but admitted that the company is still "figuring things out" in this regard. His timeline is to produce the first working test chip at the Austin research wafer fab in May or June 2027, and achieve mass production at the planned chip factory Terafab around the second half of 2028.
In terms of power supply, since accessing the grid often takes several years, many AI data centers choose to build their own power sources with gas turbines, which has led to a shortage of gas turbines. Data from gas turbine manufacturer GE Vernova (GEV.N) shows that as of the end of Q2 2026, its total backlog of orders for gas power generation equipment and capacity reservation agreements (where customers pay to lock in future production schedules) reached 116 GW, while as planned, its annual gas turbine production capacity will not increase to 20 GW until Q3 2026. Scott Strazik, CEO of the company, said at the July earnings call that the company is already negotiating with customers for deliveries in 2032 and beyond.
Musk also said that gas turbine suppliers told SpaceX that the delivery of blades and guide vanes will be 12 to 18 months later than other components of the whole unit, and casting is a "very difficult bottleneck" among them. The blades rotate at high speed with the rotor, while the guide vanes are fixed and responsible for guiding the flow of gas. Both are directly scoured by high-temperature gas, and need to be precisely cast with high-temperature alloys, with only a few manufacturers capable of stable supply.
US aerospace component manufacturer Howmet Aerospace (HWM.N) is one of the main suppliers of such blades. John Plant, Executive Chairman and CEO of the company, said at the August earnings call that Howmet Aerospace holds more than 50% of the global market share of gas turbine blades for power generation, and the capacity of its newly built factory in Japan has been almost fully booked.
According to Plant, new production capacity will not be released gradually until 2028 to 2030. In addition, the company has to reserve capacity with some machine tool suppliers. Patrick Winterhalter, CFO of Howmet Aerospace, attributes the growth of the gas turbine business to the rising demand for power generation, especially the demand for gas power generation from data centers.
Right, Musk's solution is still for SpaceX to cast the blades and guide vanes on its own. After he made this plan public at the end of August, Howmet Aerospace's stock price fell by more than 7% in the first trading day that followed. At the aforementioned earnings call, Musk explained that the casting capabilities SpaceX developed for rocket engines can be used to produce these components, "so we can get the power connected 12 to 18 months earlier than anyone else". He also said that in addition to gas turbines, SpaceX hopes to make use of solar power, wind power, and all available power sources.
Johnsen said that by the end of October, SpaceX's computing power scale is expected to approach 2.2 GW. Calculated based on this, about 0.5 GW of new computing power will be added only in October, while the new added scale in the previous year was about 1.1 GW. Musk also said that building supporting power, cooling facilities, electrical equipment, batteries and networks for a large number of new chips put into operation every month is "absolutely crazy", but the deployment pace will further accelerate in 2027.
Musk's plan is to bring SpaceX's computing power scale to 2.5 to 3 GW by the end of 2026. In 2027, SpaceX is targeting an additional 20 GW of power, cooling and infrastructure. His view on this is: "Difficult, but not impossible." At least, he is "very confident" in adding 10 GW of new computing power in 2027.
Of course, the relevant capital investment is also increasing. In Q3 2026, SpaceX's capital expenditure was 19.593 billion US dollars, and Johnsen expects it to reach 35 to 40 billion US dollars in Q4 2026. In addition, the specific deployment of computing power is demand-oriented and advanced according to the pace of contracts. In addition to operating cash flow, funds can also be obtained by leveraging investment-grade credit and tapping multiple debt funding pools.
Even Musk himself has left some leeway on whether these plans can be fulfilled as scheduled. He admitted at the aforementioned earnings call that he does not expect everything to go smoothly.
This article is from the WeChat official account "Economic Observer", written by Zheng Chenye, and published with authorization from 36Kr.