AI data centers: the hidden costs of the compute buildout

01The scale of the bet

The fastest way to understand the AI boom is that it stopped being a software story and became an infrastructure story. Training and running frontier models requires warehouse-scale buildings, hundreds of thousands of accelerators, high-voltage substations, cooling plants measured in river-water volumes and, above all, firm electricity around the clock. The Economics Explained channel, whose analysis of the buildout is embedded above and has drawn an audience of roughly 387,000 viewers, makes the central point bluntly: the world is executing the largest capital expenditure program in corporate history on the expectation of revenue that has not arrived yet.

McKinsey has estimated that meeting projected AI demand through 2030 would require roughly 5.2 trillion dollars of cumulative capital spending on data centers worldwide, including the compute, the power systems and the buildings. For comparison, that is on the order of the annual GDP of a G7 economy being poured into sheds full of silicon, most of it spent not on the glamorous model layer but on concrete, copper, transformers and grid interconnects.

The bet is unusual in another way: it is being made almost entirely by a handful of private companies. Microsoft, Amazon, Alphabet and Meta alone have guided to combined capital budgets in the hundreds of billions of dollars per year, with the majority flowing toward AI infrastructure. When four firms make an investment decision of that size, the failure mode is no longer a bad product launch; it is a macroeconomic event.

02The energy bill nobody budgeted for

The International Energy Agency, in its Energy and AI report published in April 2025, put numbers on the physical footprint. Global data center electricity consumption was around 415 terawatt-hours in 2024, roughly 1.5 percent of world electricity demand. In its base case, the IEA projects that figure roughly doubling to about 945 terawatt-hours by 2030, which would push data centers past industries such as aluminum, cement and chemicals in share of global electricity.

The American numbers are even starker because the United States hosts the largest share of the world's AI compute. The Lawrence Berkeley National Laboratory's 2024 United States Data Center Energy Usage Report found data centers consumed about 4.4 percent of US electricity in 2023, with a plausible range of roughly 6.7 to 12 percent by 2028. EPRI, the Electric Power Research Institute, framed the 2030 range similarly: from around 4 percent of US electricity today to between 4.6 and 9.1 percent by 2030 depending on adoption and efficiency assumptions.

What makes these numbers hard is not their size but their shape. An individual hyperscale campus draws tens to hundreds of megawatts, and gigawatt-scale campuses have been announced, all of it as firm, 24/7 load. That is the opposite of the demand profile the grid spent two decades optimizing for behind weather-dependent wind and solar generation.

03Capital costs and the depreciation trap

Inside a modern AI data center, the building is not the expensive part. Industry rule-of-thumb puts shell and power infrastructure at roughly 10 million dollars per megawatt of IT capacity, but the accelerators that fill it can cost several multiples of that per megawatt. An NVIDIA GB200 NVL72 rack lists in the millions of dollars, and a single campus now routinely represents billions in GPUs.

Here is the hidden cost the video centers on: the depreciation mismatch. The power infrastructure, the transformers, switchgear, substations and interconnects, is engineered and financed against 20- to 40-year asset lives, the way utilities have always worked. The GPUs sitting inside degrade economically in three to five years, because each accelerator generation delivers dramatically more useful compute per watt and per dollar. Owners are therefore signing multi-decade energy contracts to house assets that will be scrapped on a smartphone refresh cycle.

The financing structures reveal the strain. Hyperscalers fund the bulk of capex from operating cash flow, but the marginal projects increasingly use special-purpose vehicles, sale-leasebacks, and partnerships in which utilities or infrastructure funds take the long-lived assets while the tech company leases the short-lived ones. The structure is rational; it is also an admission that these are two different businesses wearing one building.

04The grid is the bottleneck

Electricity, not land and not capital, is now the binding constraint on buildout schedules. Interconnection queues at major US grid operators stretch for years, with thousands of gigawatts of generation and storage requests waiting across the country. Large transformers and gas turbines carry multi-year lead times, and high-voltage equipment supply chains are booked out. A developer can raise ten billion dollars in a quarter and still wait four years for the substation.

The responses define the current era of energy deals. Microsoft signed a power purchase agreement tied to restarting a unit at the Three Mile Island site in Pennsylvania, renamed the Crane Clean Energy Center, with the plant's full output contracted to a single data center customer. Other hyperscalers have signed nuclear, geothermal and behind-the-meter gas arrangements, effectively becoming utilities' most aggressive counterparties. The common thread is that tech companies are no longer buying power from the market; they are commissioning generation.

The IEA notes a further subtlety: where a data center is built matters as much as whether it is built. Regions with cheap, low-carbon firm power, such as those with hydro and nuclear fleets, are attracting disproportionate shares of new capacity, while load growth in constrained regions feeds directly into the next section's problem.

05Who actually pays

The bills for the buildout arrive at four doors. The first is the hyperscalers' own income statements, which so far are absorbing the cost because cloud and advertising cash flow can carry it. The second is the electric ratepayer. Adding gigawatts of firm demand in a constrained grid raises wholesale capacity and energy prices for everyone in that market; analysts have attributed part of recent capacity auction increases in the PJM interconnection, the grid serving the Mid-Atlantic and Chicago, to data center demand growth. Regulators in several states are now designing large-load tariffs so that data centers pay the true cost of the dedicated infrastructure they require, rather than socializing it across households.

The third door is the taxpayer, through accelerated depreciation, investment tax credits and state incentive packages competing for campuses. The fourth is the most speculative: the end customer of AI services, whose willingness to pay is the revenue assumption underneath every one of the financing structures above.

The distributional question is the one Economics Explained returns to repeatedly: the gains from the AI buildout are privately captured, while several of the costs, higher capacity prices, grid strain and stranded-asset risk if demand disappoints, fall on the public. Whether that trade is acceptable is not an engineering question, and it is being litigated right now in state public utility commissions.

06Limits, risks and the outlook

Two forces could bend the curve. The first is efficiency. Compute per watt has improved relentlessly, and model-level innovations, such as the mixture-of-experts techniques and cheaper training runs demonstrated by labs including DeepSeek in early 2025, showed that AI capability per dollar can move in discontinuous jumps downward. Every such jump delays some marginal data center that no longer pencils out. The IEA's low-demand scenarios, where efficiency gains outpace demand growth, still show substantial growth, but well below the base case.

The second is the revenue test. The buildout is justified by projected AI income that remains, as of 2026, far smaller than the capital being deployed to chase it. History's cautionary example is the fiber-optic buildout of the late 1990s, which bankrupted its builders and then, because the dark fiber was real and cheap, became the substrate for the next two decades of the internet. The optimistic read of the AI buildout is identical in structure: even under disappointment, cheap compute persists. The pessimistic read notes that fiber does not depreciate in five years and does not need a gigawatt.

The signals to watch are concrete: hyperscaler capex guidance each quarter, vacancy rates and secondhand GPU prices as a proxy for compute oversupply, utility rate cases with large-load tariff design, and the PJM capacity auction results. Those four lines will tell you, earlier than any earnings call, whether the largest infrastructure bet in corporate history is converging with its revenue, or diverging from it.