UBS Puts Hyperscaler AI Capital Spending at $4.1 Trillion Through 2028

UBS now estimates that the largest cloud and platform companies will spend roughly $4.1 trillion on capital projects between 2026 and 2028, against $1.292 trillion over the prior six years. The bank also lifted its global artificial-intelligence capital-expenditure forecast to $423 billion for 2025 and $571 billion for 2026, up from $375 billion and $500 billion, and sees roughly $1.3 trillion a year by 2030.

The number that matters more for the equity case is the ratio, not the total. UBS calculates that Amazon, Alphabet and Microsoft together will spend the equivalent of 102 percent of their cloud revenue on capital expenditure in 2026. Every dollar the three companies' cloud businesses earn is being reinvested in land, power, buildings and accelerators, and they are spending beyond that. A summary of the note circulating on Monday put the three-year totals at about $938 billion for Google, $683 billion for Meta and $672 billion for Microsoft.

For engineers, the relevant consequence is that capacity is being built ahead of demand, which is why serving prices keep falling even as frontier models get larger. For anyone holding the equity, the consequence runs through depreciation. Servers and accelerators carry short useful lives relative to buildings, so a spending wave of this size converts into a fixed cost that lands on the income statement over the following three to six years whether or not inference revenue arrives on schedule. Several of these companies have already moved part of the buildout into leases, joint ventures and special-purpose financing vehicles, which shifts the exposure from equity holders toward credit markets and toward the utilities and developers signing long-dated power contracts.

Note the source. This is a sell-side projection, not disclosed guidance, and UBS has raised its own numbers twice in a year. The projection is best read as a description of announced intent by four companies with unusual balance-sheet capacity, not as a forecast that has been independently tested.