The data-centre boom: big numbers, thin margins

A headline from a market-research firm recently promised that the "data centre solutions" market will grow from $535bn in 2026 to $1.34trn by 2031, a giddy 20.1% annual pace. For the retail investor, the instinct is to treat such a number as a buy signal for everything with a server in its name. Resist the instinct. A forecast about a category is not a forecast about a company, and this particular one measures a market whose largest slice is precisely the least profitable part of the chain.

Start with what the figure actually is. It comes from MarketsandMarkets, a firm that sells research reports, and the press release is advertising. Its scope is broad by design: servers, storage, networking, power gear, cooling, racks and software all count as "solutions". Note too that the number keeps its own base moving. A year ago the same firm forecast $448.95bn in 2025 growing to $1.105trn by 2030; now it reports a $535bn base for 2026 growing to $1.34trn by 2031. Each edition re-anchors the starting point to the larger, actual spend and rolls the horizon forward, which is a fine way to make a 20% compound rate look durable.

The direction is nevertheless real. This is the biggest industrial buildout of the decade. J.P. Morgan estimates the five largest American hyperscalers will spend $697bn on capital projects in 2026 alone; McKinsey reckons global data centres will absorb roughly $6.7trn of capital expenditure by 2030, nearly $7trn once traditional IT is included. Pointing out that the headline is imprecise is not to deny the boom. It is to say that the total tells you almost nothing about which investor wins.

The first place the total misleads is in its composition. The report's own numbers show IT infrastructure—servers, storage and networking—accounting for $415bn of the $535bn market in 2026, about four-fifths of the whole. That is the commodity layer. Building a server is honest, useful, low-profit work: the value sits in the chip at its heart, not the chassis around it. The contrast in margins is stark. Super Micro, one of the three big packagers of Nvidia's GPUs, was treated as having a blowout quarter this year when it guided gross margin to 15–17%, up from about 8%—levels that a chip designer like Nvidia would regard as a rounding error. The assemblers do the volume; the silicon keeps the profit.

The second way the total misleads is older and more consequential: the boom's binding constraint has moved from chips to electricity. Merely buying the hardware no longer decides the pace of the build. It now takes five to seven years to connect a data centre to the grid against twelve to eighteen months to build the building, and nearly half of America's planned 2026 projects have been cancelled or delayed, a shortfall of some 7GW. Transformer lead times have stretched to five years as demand has risen 119% since 2019. McKinsey puts 60% of AI workload capital into semiconductor and IT suppliers, but a quarter—over $1trn—into the "energisers": the utilities, transmission lines, cooling and electrical equipment that actually make the machines switch on. As compute demand roughly doubles electricity consumption by 2030, whoever can deliver power reliably and on time commands the rent. That is increasingly the scarce input, and scarcity is where margin lives.

The lesson for the reader is anticlimactic but durable: a growing market is not a growing business. The dollar count of this boom is dominated by the assemblers who earn the least from it, while the enduring pricing power sits in two scarce places—the chokepoint silicon and the power that feeds it. A $1.3trn forecast is best read as a statement about pace, not about winners. The investor's job is not to buy "the market". It is to ask, at each link in the chain, who is doing the volume and who gets to keep the margin, and to hold the yes-men of the boom to the same arithmetic.