The forecast is 1.58 trillion dollars, and it covers one year. Morgan Stanley expects the largest cloud providers and the companies that rent them computing capacity to spend that much on capital projects in 2027, up about 60 percent from roughly 980 billion dollars this year, and to spend 1.8 trillion dollars in 2028.
The report, published this week, is unusual for what follows the number. Instead of treating the spending as given, Ariana Salvatore and her colleagues on the strategy team set out the ways it could fail to translate into revenue, and organized the risks into three categories.
The first is politics, and it operates at the level of the state rather than the federal government. Interconnection requests, tariffs for large industrial loads, siting decisions and property tax abatements are all decided by state utility commissions and local boards. Arizona has suspended a set of relevant exemptions through June 2029. The states the bank flagged account for nearly 60 gigawatts of data center capacity either under construction or planned, and Texas alone accounts for 24 gigawatts of it.
That is a description of a permitting system under strain. A gigawatt of computing is roughly the load of a large city, and the decisions about who pays for the transmission to serve it are made by commissions with elected members and limited staff. The bank’s point is not that projects will be rejected. It is that the timetable for approval has become a variable that cannot be modeled from a capital expenditure forecast, and that a project delayed by two years has a different return than the one in the spreadsheet.
The second risk is the price cycle, and the bank argues it is the one investors most consistently underestimate. The example it uses is NAND flash at the end of 2022, when shipments fell 5 percent and prices fell 20 to 30 percent, and profits collapsed far faster than either number would suggest. Memory is a commodity with high fixed costs, and when demand turns, the price adjustment is larger than the volume adjustment and it arrives sooner.
Applied to the current buildout, the lesson is that the suppliers earning record margins on accelerators, memory and networking equipment are exposed to a correction that would show up in their results before it shows up in the demand data. The customers have committed to spending; the suppliers have committed to capacity. When the two curves cross, the correction happens in price.
The third risk is the most concrete. Even counting nuclear plants, gas turbines and fuel cells that are planned or under construction, the United States will be short about 32 gigawatts of power in 2028 against the 97 gigawatts the buildout requires, roughly a third of the total. In 2029 the gap widens to about 72 gigawatts, because demand keeps growing and the generation that would close the gap takes longer to build than the data centers it serves.
That arithmetic is the one that turns a forecast into a constraint. A gas turbine ordered today arrives in three to four years. A nuclear plant takes a decade.
A interconnection queue in most regional markets takes two to five years to move through. Data centers are built in eighteen months. The mismatch is structural, and it means the spending can be committed and the capacity still cannot be energized on schedule.
The bank kept a constructive view, with conditions. That formulation is doing real work. It means the analysts believe the demand is genuine and the buildout will happen, and that the returns depend on which of the three risks materializes first, in which state, and for which company.
The distributional consequence is the part the headline number hides. Companies with their own generation, or with contracts for power that is already interconnected, will earn the returns in the forecast. Companies dependent on the queue will spend the money and wait. The difference between the two disappears in the aggregate.
Every large infrastructure cycle in the past century has been financed on the assumption that the constraint was capital, and slowed by the realization that it was permits, equipment lead times or fuel. The fiber builders learned it in 2001. The data center developers are learning it now.
What makes this cycle different is the speed of the demand signal. A cloud customer can sign a contract for a hundred megawatts in a week, and the utility that must serve it needs several years to build the line. In previous cycles the demand arrived at the pace of the physical build. Here it arrives at the pace of a software release.
Several banks have raised their capital expenditure forecasts repeatedly this year. The value of this report is that it puts the number and the objections in the same document.
For investors, the practical question is which of the three risks is priced. The credit market has begun to price the financing risk, with spreads on data center borrowers widening sharply this month. The equity market has not priced the power risk, because the companies most exposed are utilities and equipment makers whose shares have risen on the demand. The political risk is not priced anywhere, because it is decided in hearings that no one models.
The report’s implicit conclusion is that the buildout will be smaller than the forecast and later than the schedule, not because the demand fails but because the physical system cannot absorb it that fast. That is a different investment thesis from the one the sector has been trading on, and it favors the companies that own the bottleneck rather than the ones that buy through it.
One trillion five hundred eighty billion dollars is a large number, and the bank that published it is not arguing the money will not be spent. It is arguing about when, where, and who pays for the connection.
