The AI buildout, in context.
The numbers behind AI infrastructure get quoted so often that they have stopped meaning much. A figure with "billion" in it shows up in a headline, people nod, and the conversation moves on. It is worth slowing down once and putting the spending next to the other times a country built something this large, because the comparison says more about what to plan for than any single headline does.
How big the spending is.
Start with the four largest cloud companies. According to their filings with the SEC, Microsoft, Alphabet, Amazon and Meta spent a combined $166 billion on capital expenditures in 2022. CNBC's tally of the guidance those same companies gave for this year puts their planned spending at about $641 billion, once both years are expressed in 2025 dollars. That second figure is a plan, and much of it has yet to be spent.
Capital expenditure is the spending that builds physical things: data centers, chips, power connections, cooling. It is committed before the demand arrives, often years before. When a few companies commit at that scale, the rest of us inherit what they build, in the form of capacity that sits there waiting to be used.
The comparison with past buildouts.
Stijn Van Nieuwerburgh of Columbia University, in a paper published by the Brookings Institution this fall, set AI against earlier buildouts as a share of the economy. He projects that AI infrastructure investment will average about 3.6% of U.S. GDP every year through 2032. By the same paper's measures, the highway years of the 1950s to the early 1970s ran at about 1.1%, the telecom and fiber boom around the turn of the century also ran at about 1.1%, and electrification in the early 1900s ran at about 0.5%.
Speed tells a similar story. Take the first four years of each buildout, adjust for inflation, and measure growth from year one. On the four companies' filed numbers, AI capital spending grew 126% from 2022 to 2025. Using the historical series in Van Nieuwerburgh's paper, electrification grew 47% over its first four years, telecom and fiber 34%, and highways 24%. Add this year's planned spending and the AI line reaches 286% above its starting point.
The buildout will keep going whether or not your organization is ready for it, so spend your effort on what you build on top.
What lasted last time.
Every one of those earlier buildouts left something that people still use. The grid still runs homes and factories. Trucks and commuters still drive the interstates. The telecom boom ended badly for many of the companies that paid for it, as the Federal Reserve Bank of Richmond documented in its history of that boom and bust, but the fiber stayed in the ground and later carried the video calls and cloud software most offices now run on.
So the history suggests the capacity will outlast the current round of investors. That matters more to a town office or a mid-sized manufacturer than any forecast about which company wins, because it means the tools will keep getting cheaper to reach and harder to ignore.
The one difference worth naming.
There is an honest gap in the analogy. Rail lines, highways and fiber lasted for decades. A large share of AI money goes into chips, and the same Brookings paper notes those can become obsolete in three to six years, so some of this spending will have to happen again. The buildings, the power connections and the know-how of running them wear out much more slowly.
Organizations face a smaller version of the same split. Some of what you spend on AI is like the chips: a particular tool, a particular model, a subscription you will likely replace within a few years. Some of it is like the power lines: a team that knows how to brief an assistant, a process that has been mapped and rewritten, clean records that any future tool can read. The first kind depreciates fast. The second kind keeps paying out regardless of which vendor you use next year.
We see teams get this backward fairly often. The license gets bought in a week and the process around it stays exactly as it was, so when the tool changes, nothing that was learned carries over. A clinic that wrote down how its intake forms move from the front desk to the chart can point a new tool at that map in an afternoon. A clinic that only bought a tool has to start again.
Sorting your own spending.
The buildout is a good prompt to look at what your organization has put into AI so far and ask which parts will still be standing in five years. A short list is enough. Do it once a year, ideally when budgets are being set.
A durable-or-disposable check
- List every AI cost from the past twelve months: subscriptions, seats, usage fees, consultant hours, staff time spent learning.
- Mark each line as disposable (tied to one product or model) or durable (skills, documented processes, data that is clean and findable, a written policy).
- For each disposable line, write down what would be lost if the vendor disappeared tomorrow. If the answer is "everything we learned," move some of next year's money to writing that learning down.
- For each durable line, name the person who keeps it current. A process map nobody updates turns disposable within a year.
- Compare the two columns. If nearly all the spending is disposable, plan the next round of spending around training and process work before adding another tool.
- Keep the list with the budget, so next year's review starts from what was actually built rather than from memory.
None of these steps needs the newest model or a large budget. A school district, a professional firm and a nonprofit can all run the list in an hour with whoever pays the invoices and whoever does the work in the room together.
