AI build-out forecasts are increasingly reliant on Elon Musk telling the truth

Hyperscalers, we’re told, are getting more serious about balancing AI expenditure with free cash flow generation. A tough October for tech stocks has left them with no alternative.

Results last week from Alphabet and Amazon broke the upgrade cycle by leaving 2026 capital expenditure guidance unchanged. Meta’s results narrowed its 2026 capex range by a fraction, while Microsoft shaved its guidance to reflect capacity constraints. Oracle used an investor day to reassure investors that 2028 would represent peak capex.

Then there’s SpaceX.

In its maiden results as a quoted company, SpaceX approximately doubled its fourth-quarter capex guide to between $35bn and $40bn.

Officially speaking, SpaceX only entered the data centre business in February by buying xAI from CEO Elon Musk. It’s now accelerating past competitors because the skills required to be a top-tier rocket scientist are transferable to fitting out warehouses, Musk said on the results call:

This is a bit like the New York Yankees playing in Little League. It’s not fair. But the net result is being able to deploy power, cooling and infrastructure at a rate that is vastly faster than anyone else, and may at some point be faster than everyone else combined. So, just let that sink in for a minute.

This is a bit like the New York Yankees playing in Little League. It’s not fair. But the net result is being able to deploy power, cooling and infrastructure at a rate that is vastly faster than anyone else, and may at some point be faster than everyone else combined. So, just let that sink in for a minute.

Musk told investors that SpaceX would have “maybe 2.5 to 3GW” of compute online by the end of 2026, and that he was “fairly confident of being able to bring an incremental 10GW of compute online in 2027”.

The second of these comments added more than $500bn to hyperscaler capex forecasts for the next two years. Exclude SpaceX, and total expectations would have barely changed this quarter:

To put it another way, over the past three months, SpaceX alone has been responsible for 100 per cent of the upgrade to current-year hyperscaler capex expectations. It also accounted for 62 per cent of the total capex upgrade for 2027 and 2028.

Cash-financed capex isn’t the whole picture. Data centre lease financing has been growing in importance lately and, because it bypasses the cash flow statement, doesn’t show up in analyst forecasts or management guidance.

Morgan Stanley analysts George Webb, Mark Hyatt and William Richards show how SpaceX’s IPO in June has provided an artificial bump to long-term forecasts just as they were showing signs of levelling out:

Though that doesn’t include all the capex that’s shifting off-balance-sheet:

Musk’s own predictions suggest the above estimates might be much too cautious. For 2027, SpaceX will aim to install twice the required amount of “power, cooling, and infrastructure” needed to turn the GPUs on, he said in the Q3 results call:

There’s a non-zero possibility — maybe it’s a small possibility at this point, but it’s not zero — of SpaceX having over 20 gigawatts of Super Intelligence compute online next year. A small possibility, but it used to be zero. It’s not zero any more.

There’s a non-zero possibility — maybe it’s a small possibility at this point, but it’s not zero — of SpaceX having over 20 gigawatts of Super Intelligence compute online next year. A small possibility, but it used to be zero. It’s not zero any more.

The likelihood of a Musk “non-zero possibility” is hard to estimate, however. You may remember SpaceX from such guidance as:

Further reading:
— Fact-checking some dull bits of Elon Musk’s biography (FTAV)
— Of SpaceX, Tesla, and metaphysical horror (FTAV)