Okay, beginning with the dollar. Aggregate capital expenditure across the four largest cloud providers was roughly US$226 billion in 2024.13 Then US$410 billion in 2025.14 And now US$700 billion is guided in 2026. So take the ratio: the number compounds roughly at 1.7-1.8x per year, across three years and four firms that would prefer to spend less than each other.15 For 2027, sell-side projections already exceed US$1 trillion, which isn’t a bold forecast.16 Why? There isn’t a reason to bend that line. Now play it forward further and the annual figures start looking absurd: Goldman Sachs’s baseline of US$7.6 trillion in cumulative investment across compute, data centers, and power between 2026 and 2031 already factors in growth rate deceleration.17 No one modelling this line believes that it can run at 1.8x forever. If you plot these on a graph, you’ll find yourself looking at something like a straight line.
Let’s talk about the demand that reflects this cumulative investment. A primary driver is that capacity is being monetized as quickly as it’s being built, and that demand exceeds supply. In 2025, AWS added 3.9 GW of power capacity, expects to double that power capacity by the end of 2027, and is monetizing that capacity as fast as it is being installed whilst continuing to face capacity constraints that yield unserved demand (For scale: 7.8 GW of installed data center power capacity is enough to power the peak demand of Singapore, depending on usage assumptions).18 Microsoft has energized nearly 1 GW in a single quarter and as of April 2026, we have Microsoft spending nearly US$200 billion in one year on aggressive expansion of their data centers and they still expect capacity to remain constrained at least through mid-2026, even as they bring GPU, CPU, and storage capacity online quicker.19 And then there’s Oracle, which is late to the game and is compensating with one of the most aggressive balance sheet campaigns in corporate history. Issuing debt, diluting shareholders, just to fund its transition to become a cloud infrastructure company.20 Whether Oracle wins is besides the point. The point is that we have a fifty-year-old software company that is leveraging its balance sheets just to build data centers, giving you a rough idea of the demand on the other side (I will go into what the credit market thinks further down in this essay). Finally, we have neoclouds whose businesses have grown rapidly because hyperscalers cannot meet all GPU demand. CoreWeave, a company that barely existed five years ago, holds 3.5 GW under contract.21
Now, back to that straight line I mentioned earlier, the dollar line isn’t going to run forever. If you extrapolate at 1.7-1.8x annual growth, you will eat the combined capex of the entire S&P 500 within a few years, and it isn’t difficult for anyone with a calculator to piece that together.22 Lines like this don’t break on financing. Money is relatively flexible in the sense that banks can always expand lending, bond markets can absorb an enormous amount of issuance, private equity and sovereign wealth funds can always supply capital, and the reasons can go on. Lines like this break on physics: land, transformers, and grid interconnection. Hence, using the dollar going forward will be the wrong unit of account. Let’s swap over to the watt now.
The infrastructure buildout has three tiers that are occasionally conflated and should not be. There is aspirational capacity, contracted capacity, and active energized capacity. Aspirational capacity is announced infrastructure plans and multi-year ambitions, which is where the press releases live. Contracted capacity reflects gigawatts attached to signatures, binding commitments such as power purchase agreements, take-or-pay compute contracts, and project financing already drawn. Finally, we have active energized capacity, essentially what is already physically drawing power into compute and generating revenue today.23 The market knows how to judge the first tier, and completely ignore the second. The evidence in this essay revolves around contracted and active energized capacity.
Starting with energized capacity, as I mentioned above, Amazon added 3.9 GW of power capacity in 2025, roughly the output of three to four nuclear reactors, brought online in just one singular year by one company, and expects to double its total capacity by 2027.24 Microsoft stood up more than 2 GW of capacity in 2025, added a further 1 GW in only the first quarter of 2026, and is on track to double its overall footprint over the next two years.25 Contracted capacity is being accumulated in chunks that would have been national news a decade ago. CoreWeave and Nebius both have roughly 3.5 GW under contract.26 Playing this forward, you get a buildout that stops looking like a corporate spending program and starts looking more like the electrification of a mid-sized industrial country.
If the extrapolation sounds unserious or if there is any doubt regarding the contracted and energized capacity, just take a look at what the infrastructure builders are doing with their own money: buying the watts in dollars with irreversible, decade-scale commitments. Microsoft recently signed a twenty-year agreement with Constellation Energy to restart Three Mile Island.27 Amazon secured up to 1.2 GW of carbon-free energy directly from the Comanche Peak Nuclear Power Plant and is building “Project Spectrum,” which is a US$5 billion data center on adjacent land.28 Meta agreed to a 20-year Power Purchase Agreement with Vistra (PPA) for 2.609 GW of carbon-free energy from its Perry, Davis-Besse, and Beaver Valley nuclear plants.29 Alphabet couldn’t seem to find a sufficient power contract, so it simply bought the power developer: Intersect Power for US$4.75 billion.30 I hope this clears some doubt with regards to how seriously the infrastructure players are treating power capacity. You don’t simply recommission an entire nuclear plant to serve a trend you expect to break.
One more measurement, because it helps to quantify the transformation that this essay is named for. Capital intensity, which is capex as a share of revenue, ran near 10% for these firms across the software era.31 The entire asset-light business model rested on that number staying small, and dominating through intangible assets like software, network, and brand power. Now, this number runs between 45-57%.32 For calibration, that is not the profile of a software buildout; that is the profile of an intense utility buildout: railroads in the 1880s, a telecom in the fiber buildout, or an electric utility company during a major generation and grid expansion. Income statements are starting to register it. For example, Alphabet’s depreciation expense rose 38% from US$15.3 billion to US$21.1 billion in a single year.3334 Its depreciation is simply past capex arriving on the income statement with a delay. The metamorphosis in the title is not a metaphor; it’s real, and it’s visible in the filings to anyone reading the right numbers in the right units. These software and technology companies with their traditionally asset-light roots are pivoting away from that and into asset-heavy utility and infrastructure companies.
All this information that I’ve provided is disclosed figures, public contracts, and the markets have seen every number in this section. They see the capex line and see that spending must revert. It’s not a cost curve; it’s a supply curve that is chasing a demand curve that it has not once even caught.