AI Agent UpdateBank of America: data-center spending hits $1 trillion this year — and doubles by 2028
Bank of America says the buildout doubles by 2028 and nearly triples by 2030. The power grid, not demand, is the only thing that can stop it.
The News
The AI buildout just got a number big enough to be hard to say out loud. Bank of America Global Research said in a note dated Friday, October 23, that global data-center spending will reach $1 trillion in 2026 and more than double to $2.1 trillion by 2028, on its way to $2.93 trillion by 2030 — a 37% compound annual growth rate from 2025 to 2030. For scale, the market was $275 billion as recently as 2022. In four years it roughly quadruples; in six, it grows more than tenfold. Bank of America cites AI demand as the driver, with chipmakers and server developers the biggest beneficiaries so far.
The cost breakdown explains where the money goes — and why chip stocks move when data-center forecasts move. Building costs about $49 million per megawatt of capacity, with IT equipment making up 78% of that spend. Servers alone run $31.4 million per megawatt; networking adds $4.8 million per megawatt. The newer 'neocloud' operators — companies built specifically to rent AI compute — cost even more, about $57 million per megawatt. In other words, a data center is mostly a very expensive box for holding Nvidia chips, and the forecast is really a forecast for how many of those chips the world will buy.
The counterweight arrived the same day. S&P Global Ratings reported October 23 that 146 US data-center projects have been delayed or canceled, with about three-quarters of the delays stemming from local opposition, and warned that grid-capacity constraints, environmental controversies and zoning changes 'may increase downward pressure on the credit quality of data center projects.' Bans, slowed permits or policy reversals causing substantial delays could trigger rent abatements and even tenant termination rights, though S&P notes the actual rating impact so far remains limited. The money and the megawatts are both real; whether they meet on schedule is the open question.
A data center is mostly a very expensive box for holding Nvidia chips — and the world just ordered a lot more boxes.
Why It Matters
Start with why a bank forecast moves markets. Data-center spending is the master number behind the AI trade: every dollar of it eventually becomes revenue for someone — chipmakers, server assemblers, networking vendors, landlords, utilities. When Bank of America says the number doubles in two years, it is effectively saying the revenue pools for all of those companies double too. That is why a forecast like this matters more than any single company's earnings: it sizes the entire pond everyone is fishing in.
The breakdown matters as much as the total. If 78% of build cost is IT equipment, then the bottleneck — and the profit — sits with the companies making servers and silicon, not the ones pouring concrete. It also explains this week's market behavior: when investors worried about AI financing, equipment stocks fell first and hardest, because their revenue is most leveraged to whether the buildout keeps getting funded. The forecast says it does, at a pace that would make data centers one of the largest capital programs in economic history.
Finally, the forecast reframes the week's bad news. Stalled projects, grid fights and local opposition are real, but they are constraints on the pace of building, not on the desire to build. A $2.1 trillion forecast slipping by a year is still a $2.1 trillion market arriving slightly late. The distinction between 'delayed' and 'canceled' is the whole investment question for the next two years, and Bank of America's numbers argue the demand side of that question is already settled.
Analyst's Take
The verdict is Bullish, and the size of the numbers is the argument. A market growing 37% a year for five years does not need every project to succeed — it needs the average project to keep getting funded, and the funding case rests on AI demand that has not yet shown a crack. Chipmakers and server builders are named as the biggest beneficiaries so far, which lines up with the rest of this week's tape: equipment and memory names fall hardest when financing wobbles precisely because they gain most when it flows. For readers holding Nvidia, Equinix or Digital Realty, this forecast is the demand-side anchor for the whole thesis.
The bear case deserves its paragraph, because it is not theoretical. One hundred and forty-six stalled US projects, plus the Australian pipeline where only 23% of planned capacity has power secured, show the constraint has moved from 'can we afford the chips' to 'can we plug them in.' S&P's warning about credit quality is the financing world's way of saying delays cost money even when projects eventually get built. But note what S&P did not say: demand is not falling, and actual rating impact so far is limited. The risk to this forecast is timing, not direction. Bullish on the buildout, clear-eyed on the grid. Readers' poll: does spending really double by 2028, or does power push it out?
By The Numbers
FAQ
What does $1 trillion of data-center spending actually buy?
Mostly the computers inside the buildings. Bank of America breaks construction down to $49 million per megawatt of capacity, with IT equipment making up 78% of the spend — servers alone cost $31.4 million per megawatt, networking another $4.8 million. The concrete and cooling are the cheap part; the silicon is the money.
What is a megawatt, in plain terms?
A megawatt is a million watts — roughly enough electricity to power 750 to 1,000 homes at once. Data centers are measured in megawatts the way apartments are measured in square feet: it tells you how big the machine inside can be.
What could break this forecast?
Power and permission. S&P Global Ratings reported October 23 that 146 US data-center projects have stalled, with about three-quarters of delays tied to local opposition, grid limits, environmental disputes or zoning changes. The money is willing; the grid connections and permits are the bottleneck.
Source: Reuters / Bank of America Global Research