Five of the largest technology companies in the United States are set to spend about $1.2 trillion next year building the computing capacity that AI needs, according to Goldman Sachs strategists. The figure is more than 50 percent above the roughly $800 billion the same five are expected to spend in 2026, and above the roughly $1.1 trillion that analysts on Wall Street expect on average. It also carries a question the spending cannot answer on its own. Goldman estimates the companies would need about $300 billion a year in AI revenue to earn that outlay back, and says the AI revenue they have today falls short of it.
What the $1.2 trillion buys
The five are Amazon, Alphabet, Microsoft, Oracle and Meta. The first four sell cloud computing, the rented processing power and storage that other businesses buy instead of running their own machines. The $1.2 trillion goes into data centres, the chips inside them and the electricity supply that keeps them running. Goldman's forecast has the growth rate of this spending falling from nearly 100 percent in 2026 to 54 percent in 2027 and 12 percent in 2028, even as the amounts themselves keep rising. Ryan Hammond, a Goldman strategist, said that measured against the size of the economy, the 2027 cycle would be the largest investment of this kind since railroad construction in the 19th century.
Where the money comes from is changing
Goldman's estimate of the revenue needed is its test of whether the build-out pays for itself: about $300 billion a year in AI sales. Cloud revenue growth across these companies rose from 25 percent in 2024 to 48 percent in the second quarter of 2026, the clearest sign in the numbers that customers are spending more on the services the new machines will run. That is still short of $300 billion a year.
The money itself now has to come from somewhere else. Goldman says the spending has outgrown the cash these companies generate from their operations, so more of it will have to come from borrowing. The borrowing already shows across the wider market: US companies issued about $1.9 trillion in bonds through August, 30 percent more than in the same period a year earlier, according to SIFMA data reported by the Hindustan Times, and bond sales by companies linked to AI have passed $400 billion this year, roughly 90 percent of it by US firms.
Kevin Warsh, the chair of the Federal Reserve, described what that does to the price of money. "The competition for capital is real," he said, and he said it partly explains the increase in yields on 10-year Treasury notes, the interest rate the US government pays to borrow for ten years and the reference point for many other loans.
The bill for worn equipment arrives later
There is a second cost that does not land when the money is spent. Depreciation is the accounting charge a company takes each year as its equipment ages and loses value, spreading the cost of a machine bought today over the years it is used. Goldman analyst Ben Snider estimated in a September 18 note that depreciation at these companies will subtract about 5 percentage points from S&P 500 earnings growth in 2027, offsetting nearly half of the 11-point boost that the spending itself gives to earnings. By 2028, he expects the depreciation to cancel that boost out entirely.
Snider also traced where the spending lands meanwhile. It becomes revenue for semiconductor, hardware, industrial and utility companies, which he said account for roughly half of the earnings growth analysts expect from the S&P 500 this year.
The grid decides how fast this can be built
Goldman expects US data centre power demand to more than double from its 2025 level to 66 gigawatts by 2027, a measure of how much power has to be available at once to run the machines. Electricity does not arrive like chips in a shipment. It needs generators, transmission lines and a connection to the grid, and in Texas those connections have slowed. The state has paused progress on new data centre connections while the Electric Reliability Council of Texas audits a queue of about 474 gigawatts of requests from very large electricity users. Pablo Vegas, the council's president and chief executive, said the forecast behind the queue looked too big: "We believe this forecast to be higher than expected future load growth."
Thomas Gleeson, who chairs the Texas Public Utility Commission, said the projections were "extremely high", adding that "a lot of that load will not actually come here. A lot of it is speculative." Utilities say the stricter connection rules are meant to separate serious projects from speculative ones. Brad Viator, president of Power for Tomorrow, a trade association for utilities that own their own power plants, said that by the time a customer reaches a contract, "the utility has high confidence they're dealing with a real project, not a tire-kicker."
Not all of the resistance comes from the companies themselves. The energy advisory firm Relae, formerly Carbon Direct, found that more than $170 billion in AI data centre capacity has been blocked, withdrawn or stalled by community opposition since January 2024. Karl Rábago, a former Texas utility commissioner, told the same publication: "They are promising big earnings-per-share growth to shareholders, and the transmission and generation they build to serve data centers is the only way to get that."
What would settle it
Goldman's forecast is not that the spending stops. It is that the spending grows more slowly, from 54 percent in 2027 to 12 percent in 2028, while the depreciation from everything installed now keeps arriving. On Snider's estimate, that charge cancels out the earnings boost from continued spending by 2028. The number that decides whether any of it was worth doing is AI revenue, which Goldman says needs to reach roughly $300 billion a year to cover the outlay. Cloud revenue growth is moving toward that line, at 48 percent in the second quarter of 2026, and has not reached it.