Bank of America put a number on the buildout on November 8. Global spending tied to data centers will reach about $1 trillion this year and double to $2.1 trillion by 2028, the bank estimated. The Philadelphia Semiconductor Index, the closest thing the market has to a pure read on the companies supplying that buildout, fell 12.2% in a single day last week.
Both facts belong in the same paragraph because they describe the same disagreement. The estimate is about demand, and it is large. The selloff is about the price investors pay for that demand, and it was violent.
A spending forecast that doubles in two years implies growth of roughly 45% a year between this year and 2028, a rate that has few precedents in capital-intensive industries and none that lasted without a shortage somewhere in the chain. That is the arithmetic a portfolio has to underwrite. It is also why the forecast and the selloff can be right at the same time.
The useful question is not whether the trillion is real. It is where in the income statement the money lands, and how much of it converts into earnings the market has already paid for. Bank of America’s framing puts chips, power equipment and software on the receiving end.
Those three have very different economics. A chip is a product with a cycle, a gross margin that moves with utilization, and a customer who can delay an order by a quarter. Power equipment is a backlog business with long lead times, which makes revenue visible and capacity decisions expensive. Software is the only one of the three that can scale without a new factory, and it depends on the other two being installed first.
That ordering matters more than the headline. If the estimate is right, the spend has to show up in sequence: orders for equipment, then construction, then the power connections that let a shell become a data center, then the software that runs on it. Each step has a lag measured in quarters, and each step can stall without cancelling the one before it.
A forecast of $2.1 trillion in 2028 is therefore a statement about 2027 orders as much as about 2028 revenue. The market’s reaction last week suggests investors are not yet willing to underwrite the middle of that chain. They are buying the first step and questioning the third.
The comparison that matters is the last capex cycle with the same shape. Telecom operators spent a decade building fiber and mobile networks on the argument that traffic would grow without limit, and traffic did. What did not grow fast enough was the price per unit of traffic, which fell every year, so the spending produced assets and the assets produced less revenue than the models assumed.
The companies that did well were not the ones that built the most. They were the ones that sold the picks, shovels and software to everybody who built, and the ones that owned the scarce input. Power is the scarce input in this cycle, and that is one reason the equipment names have led.
There is a second comparison, closer to home. The index that fell 12.2% is the same index that has spent two years pricing a buildout that had not yet been delivered. A single-day drop of that size is not a verdict on the forecast; it is what happens when a group of stocks has been bought on a story and the story’s arithmetic starts to demand evidence. Order books are bottom-up and they are reported quarterly, which is why they move stocks and estimates do not.
The bull case, stated fairly, is that this cycle differs because the buyer is not a carrier with a saturated market. The buyer is a group of companies selling a product whose unit costs fall as capability rises, and they have balance sheets to fund the build without waiting for revenue. That case is credible. It is also unfalsifiable until the spending appears in the suppliers’ numbers.
The bear case is simpler, and it is the one a sell-side estimate answers least well. A doubling in two years requires power, land, permits and transformers, and those are the parts of the chain that no bank can will into existence. If the constraint binds, the dollars do not disappear; they move later, and a forecast that assumed 2028 becomes one that assumes 2029.
The distinction the forecast blurs is between spending on capacity and spending on replacement. A trillion dollars of data center investment includes servers that will be obsolete in four years, buildings that will last thirty and grid connections that will outlive both. Only the capacity portion responds to demand for AI.
The replacement portion responds to depreciation, and it continues whether or not the models get better. A supplier with a large installed base collects that stream regardless of the cycle, which is one reason the equipment names have held up better than the forecast alone would suggest.
What to watch is specific and dated. The January and April reporting seasons will carry the first real test: order backlogs at the power equipment makers, new interconnection agreements filed with utilities, and any language from the software companies about attach rates on new capacity. If those three accelerate through the first half of 2027, the $2.1 trillion figure has a spine; if they flatten while the forecast stands, the gap is the trade.
