Different Stories About Compute: What SpaceX Put on the Record About 2027

Crowd protesting AI expansion outside a data center; headline forecasts 51.2 gigawatts by 2027.

For ETF investors, the useful takeaway is not simply that AI buildout may continue, but that the bottleneck appears to be shifting among different parts of the stack. A theme that depends on data-center growth can remain intact even if local permitting slows, because capital can move toward jurisdictions, campuses and partners that can solve power, cooling and interconnection faster. That matters for thematic ETFs because the opportunity is often indirect: the fund may hold not just chip designers, but also firms exposed to electrical equipment, infrastructure, networking and industrial capacity.

That broadening is a diversification story, but not a free one. When a theme becomes crowded, index weights can concentrate quickly in a few large names and in the same small set of beneficiaries. Investors should look closely at how a fund defines its universe, whether it caps single-name exposure, and how often it rebalances. A rules-based ETF can reduce company-specific risk relative to owning one stock, yet it can also increase exposure to the theme at the exact moment valuations become more demanding.

Execution risk also matters. If physical infrastructure, not demand, is the immediate constraint, then funding, build schedules and supply-chain timing can shape which companies actually convert interest into revenue. For long-term holders, that argues for treating AI-related ETFs as exposure to a multi-year industrial rollout rather than a straight-line growth trade. Fees, spreads and tracking difference still matter, especially when the underlying basket is volatile and turnover may rise as the index adapts to fast-changing winners and bottlenecks.

In practice, the main portfolio question is whether the fund gives you broad participation in compute expansion or an unduly concentrated bet on a few perceived leaders.


Key Takeaways

  • Public opposition may redirect AI data center construction geographically, changing where capacity is built without necessarily reducing the overall buildout.
  • SpaceX targets 10 gigawatts of additional compute in 2027, with management identifying power and cooling infrastructure as greater constraints than demand.
  • SemiAnalysis forecasts accelerating data center construction in 2027 but warns that chip shortages could leave some capacity unused by the end of 2028.

The prevailing story about artificial intelligence infrastructure has turned sour, and it deserves to be taken seriously. A Gallup survey fielded in March 2026 found that seven in 10 Americans oppose building an AI data center in their local area, with nearly half strongly opposed.1 A Heatmap Pro poll of registered voters in August found that three-quarters would oppose a data center near them, and the share who would strongly support a nearby project had fallen to 4% from 13% a year earlier.2 The sentiment has moved into politics. Voters have now weighed in on dozens of local ballot measures on data centers this year, including 41 on November 3, and most of the outcomes favored more limits.

New York’s governor signed an executive order in July pausing state environmental permits for new hyperscale data centers for up to one year,3 and in September Texas’ governor ordered the state’s environmental regulator to halt all data center permits until a grid-operator audit is complete, which is expected in mid-December.4

From that vantage point, the natural expectation is that the buildout slows. The question this piece asks is narrower and more testable.

What has a company that is actually building compute at scale said, on the record, about what comes next?

For any of us watching equity markets, the “AI buildout” has been a driving force behind a large chunk of economic growth as well as the business results of many companies. Thinking about the expected pace versus the actual pace in 2027 brings many critical questions into focus.

A Witness with Something to Measure

On October 20, SpaceX5 held its third-quarter earnings webcast.6 The company is a useful witness for a few specific reasons.

  1. It is deploying compute at a pace few others disclose in detail
  2. It rents a large share of that compute to outside customers
  3. Its management attached numbers and dates to its statements

Chief Financial Officer Bret Johnsen said the company ended the third quarter with 1.7 gigawatts of nameplate compute capacity, expected to reach nearly 2.2 gigawatts by the end of October and between 2.5 and 3 gigawatts by year-end. Third-quarter capital expenditures were $19.6 billion, and he guided fourth-quarter capital expenditures to a range of $35 billion to $40 billion because of the additional capacity coming online.

Elon Musk then looked further out. He said he is fairly confident of bringing an incremental 10 gigawatts of compute online in 2027, which would put the company at roughly 13 gigawatts by the end of that year. He described the current deployment pace as on the order of 200,000 advanced GPUs per month, and said 2027 would be dominated by the next generation of Nvidia hardware.

Demand, as Management Described It

The more interesting statements concerned demand, because a slowdown thesis ultimately depends on it. Johnsen said that at current pricing, new compute deployments are generating payback in under a year. He reported a backlog of $63 billion, of which about $42 billion is expected to be recognized within the next 12 months, and guided total company revenue for the first quarter of 2027 to a range of $33 billion to $37 billion.

Musk said the company aims to split its compute roughly evenly between internal use and rental to third parties, with that mix shifting month to month. He said SpaceX rents at roughly spot-market rates rather than signing long-term deals, that short-term contracts are renewing at higher dollars per watt rather than lower and that the company could rent out its capacity several times over if it said yes to every request. His reasoning was that as models improve, the useful intelligence delivered per watt rises, which supports the value of each unit of compute.

Looking beyond SpaceX on the Scale of the Build

SpaceX is one builder, so it is worth asking whether its account fits the wider picture. SemiAnalysis tracks data center construction building by building, and its model puts net additions of global data center information technology (IT) capacity, excluding China, at 15.6 gigawatts in 2025, 25.0 gigawatts in 2026 and 51.2 gigawatts in 2027.7 Of the 2027 figure, 39.4 gigawatts comes from capacity already online or under construction and 11.9 gigawatts from buildings not yet started.The authors themselves describe 50 gigawatts of additions as essentially the global cumulative data center market through 2024 being built in one year.

The evidence behind that forecast is physical and financial. SemiAnalysis counts roughly 58 gigawatts of IT capacity under construction worldwide as of October 2026, 24.3 gigawatts of it at the 40 largest campuses. It describes most of that campus capacity as due for delivery in 2027, with most of the sites more than halfway built and exterior cooling and electrical equipment already installed. Construction starts by the four largest hyperscalers rose from 1.7 gigawatts in the fourth quarter of 2024 to 6.4 gigawatts in the second quarter of 2026. Because developers typically spend only 5% to 10% of project capital before they have a customer, signed leases lead the buildout. SemiAnalysis identifies $1.36 trillion of lease value contracted in the first three quarters of 2026, against $236 billion in all of 2024.

Where the Constraint Moved

Perhaps the most consequential statement was about what limits the build. Musk said the binding constraint on deploying superintelligence is increasingly power, cooling and electrical infrastructure rather than chips. This is, effectively, the same terrain on which public opposition operates. Local resistance acts on siting, permitting, grid interconnection and energy costs, not on the supply of silicon.

SpaceX management described how it intends to work within that constraint. Musk said turbine suppliers told the company that blades and vanes, which are cast components, arrive 12 to 18 months after the rest of a turbine, and that SpaceX is applying its rocket-engine casting capability to make them itself. He said the company is planning for as much as 20 gigawatts of incremental power and cooling infrastructure in 2027, deliberately exceeding what is needed to switch on the GPUs, while describing that figure as difficult but not impossible. He also described manufacturing modular inference compute at the factory, with power and cooling built in, and distributing those clusters widely.

Looking beyond SpaceX, SemiAnalysis reports that Google has replaced a significant amount of outdoor on-site electrical cabling with prefabricated electrical modules, that AWS is prefabricating data halls and that the developer DayOne has cut build times to about nine months from more than 18 months at traditional developers. It also argues that the power problem has broadly been solved through behind-the-meter generation from fuel cells, truck engines and refurbished turbines that bypasses grid interconnection timelines of 5 to 10 years. That sits in some tension with the turbine-component bottleneck SpaceX described, and it is a reminder that different builders are solving the same constraint in different ways.

On funding, Johnsen pointed to SpaceX’s investment-grade status, multiple pools of debt capital in discussion, operating cash flow from short payback periods and partnerships along the vendor and customer chain. He also described the pace of deployment as demand-led and prudent, which is a useful reminder that the plan is conditional on customers continuing to show up.

How the Stories Fit Together

Read together, the polling, the SpaceX earnings call and the broader SemiAnalysis data center review do not necessarily contradict each other. They describe different parts of the system. The polling measures public willingness to host facilities. The SpaceX call describes demand for what the facilities produce and the physical limits on supplying it. SemiAnalysis puts a lens across essentially every data center project it can track. Opposition can change where capacity is built, how fast and at what cost, through moratoria, ballot measures and slower approvals. SpaceX’s statements suggest that, at least for one large builder, demand is not the factor currently setting the pace.

SemiAnalysis adds that the buildout is diversifying geographically, which sidesteps some U.S. labor, power and permitting constraints, with construction outside the United States accelerating sharply in 2027 and expected to gain share in 2028, largely because of U.S. challenges. In our reading, U.S. opposition may therefore redirect some capacity as much as it reduces the total.

One implication is worth flagging as our reading rather than management’s. A strategy built around factory-made, modular compute with integrated power and cooling is structurally different from a strategy that depends on winning many individual rezoning votes. Management did not present it as a response to local opposition, and we should not assume it is one. But if siting friction persists, builders with more control over their own power and cooling supply chains may be better positioned to keep moving.

What to Weigh

There are good reasons for caution. Musk himself placed wide error bars on several targets. He called more than 20 gigawatts of compute in 2027 a small but non-zero possibility, and described the 2027 power plan as one where not everything is expected to go right. Opinion polls measure sentiment rather than permits, and neither the polling nor SpaceX’s account settles how many projects will ultimately be delayed or cancelled.

SemiAnalysis calls its 2027 to 2028 forecast aggressive and recounts one investor putting 2027 additions at about 25 gigawatts, roughly half its figure. The portion of its 2027 forecast that sits in buildings not yet under construction, 11.9 of 51.2 gigawatts, is nearly a quarter of the total. More important for the question of which constraint binds first, it expects the limit to move from facilities to chips. It describes the industry as in a silicon shortage, with high-bandwidth memory and leading-edge logic the two binding constraints, and projects high-bandwidth memory bit supply growing only about 27% in 2027, with no new fabs contributing until 2028. It sees a risk of roughly 10 gigawatts of data center oversupply by the end of 2028 relative to available chips, with new leases and orders possibly slowing toward the end of 2027 or early 2028. That is a different binding constraint from the power and cooling emphasis in SpaceX’s remarks, and a reason to separate a 2027 acceleration from a 2028 outcome. We have to force ourselves to remember all of the moving parts: just as we are forecasting data center construction projects, a different research effort could track chip fabrication construction projects.

The encouraging feature of the SpaceX statements is that they can be tested soon. Investors can compare reported nameplate capacity with the 2.5 to 3 gigawatt year-end figure, fourth-quarter capital expenditures with the $35 billion to $40 billion guide and first-quarter 2027 revenue with the $33 billion to $37 billion range. November 3 offered an early answer on how far local opposition translates into binding restrictions: voters approved 27 of 41 data center measures, most of them limits on development.

The narrative that the buildout is stalling and the evidence that one of its largest participants is accelerating can both be true at once. The more useful question for investors is which constraint binds first, with the choices being:

  1. Public consent
  2. Physical infrastructure
  3. Demand

On the record so far, strongly, SpaceX is telling investors that demand is not the one. It’s important for all of us to remember that, while a data center construction slowdown might feel like the consensus forecast, it is not a completely foregone conclusion.

1 Source: Jones, J. M. (2026, May 13). Americans oppose AI data centers in their area. Gallup.

2 Source: Meyer, R. (2026, August 19). Exclusive: 75% of Americans now oppose local data center development. Heatmap News.

3 Source: Governor Kathy Hochul. (2026, July 14). First statewide moratorium on new hyperscale data centers launched by Governor Kathy Hochul [Press release].

4 Source: Pope, D. J., Berg, J. & Paulson, M. (2026, October 7). Abbott orders statewide pause on Texas data center permits. The Texas Lawbook.

5 As of October 19, 2026, SpaceX represented a 17.11% weight in the WisdomTree Space Economy Fund (WSPC), and a 0% weight in the WisdomTree Artificial Intelligence and Innovation Fund (WTAI). Holdings subject to change.

6 Source for details about SpaceX, unless otherwise stated: SpaceX. (2026, October 20). SpaceX Q3 2026 earnings call transcript [Transcript]. Benzinga.

7 Source for SemiAnalysis-related research details, unless otherwise stated: Eliahou Ontiveros, J., Knuhtsen, R. & Patel, D. (2026, October 19). Datacenters 2027: The year of acceleration, +50GW in a single year. SemiAnalysis.

  Original Post

Enjoyed this article? Sign up for our newsletter to receive regular insights and stay connected.

Leave a Reply