The AI build-out is creating real jobs and real returns. It is also creating liabilities that investors, acquirers and local governments have barely begun to price.

Amazon, Alphabet, Microsoft and Meta spent about $410 billion on capital expenditure in 2025. Their 2026 guidance sums to roughly $720–760 billion, depending on how Microsoft's lease-accounting change is treated. The revenue to justify that is much smaller. OpenAI and Anthropic together passed $105 billion in annualised revenue by August, and Microsoft's AI business passed a $37 billion run rate.

Those are run rates, not recognised revenue, and they overlap: Microsoft's AI revenue includes what OpenAI pays it. They also leave out AI income earned elsewhere. Run rates also flatter durability. Where AI is billed by usage, revenue can shrink as quickly as customers ration consumption, while the data centres and leases behind it are committed for years. So read the gap as an order of magnitude, in the hundreds of billions of dollars a year, not a measurement. Analysts disagree on whether demand closes it in time. I won't settle that here.

My point is that calling data centres an economic boon hides a liability structure that deal teams, boards and governments should map before they sign anything.

A smelter, not a software company

A hyperscale data centre is a 24/7 industrial facility. The right comparison is a smelter or a desalination plant, because of what it consumes and what it passes on to others.

The IEA reports that data-centre electricity demand grew 17% in 2025, against 3% for global demand overall, and expects it to roughly double by 2030. Globally that is still a modest share of demand. The strain is local. One large facility can draw up to 5 million gallons of water a day, about what a small town uses.

None of this makes data centres bad investments.

It makes them resource-intensive assets whose costs fall outside the operator's ledger.

Where the costs land

Electricity. Utilities build substations and transmission lines for large loads, and traditional cost-sharing spreads that bill across all ratepayers. PJM's independent market monitor attributed 63% of the price increase in its 2025/26 capacity auction to data centres, about $9.3 billion recovered from customers across the region. In September the US House voted 417–3 for a bill to push the largest data centres toward covering their own grid costs. But it only requires states to consider the standard, and a Senate objection stalled it the next day. The direction is clear. The law is not.

Water. Ceres modelled water stress in Phoenix-area basins rising by as much as 17% a year if all planned data centres come online. Microsoft reported that 42% of its 2023 water use came from water-stressed areas, and Google reported 15%. Lower-water cooling designs exist. The diligence question is which design is actually in the contract, and in which basin.

Noise. Some costs never reach a ledger at all. Almost a third of Virginia's data centres sit within 200 feet of residentially zoned land, because zoning treats them like offices. Neighbours complain less about volume than about a hum that never stops, especially at night. Low-frequency sound barely registers on standard decibel meters, so ordinances are hard to enforce. The health evidence is unsettled: a 2026 review found no study that measured the low-frequency spectrum at an operating data centre or tracked outcomes in nearby communities. Nor does the cost show up in house prices, since the one formal study, in Northern Virginia, found homes near data centres selling for more, probably because the facilities cluster in well-served areas. So the cost surfaces in permitting fights, not yet in prices. It is also largely avoidable by design. The diligence question is whether the site plan includes shrouds, barriers and setbacks, or whether the neighbours will pay for the omission.

Tax. Thirty-eight US states offer dedicated data-centre incentives. Good Jobs First reports that Virginia's sales-tax exemption cost state and local governments over $1.9 billion in fiscal 2025. The incentives are now contested, with lawmakers in at least 28 states introducing bills to curb them in 2026. A model that assumes today's regime for twenty years is assuming a political equilibrium.

What happens when the load doesn't arrive

Deals already show the strain. One 2026 industry guide reports at least three mid-market transactions that collapsed after letters of intent. The causes were a 48-month transformer lead time, a change-of-control clause in a ground lease, and a portfolio where 61% of contracted revenue rolled off within 36 months. It is a single trade source and anecdotal, but the pattern is plausible. Power, lease quality and approvals now matter as much as headline megawatts.

The pipeline shows it too:

• 26% of the 110 projects expected online in 2025 were delayed.
• Sightline Climate expects 30–50% of large 2026 projects to slip.
• Data Center Watch counted at least 75 US projects, worth about $130 billion, blocked or delayed by local opposition in Q1 2026 alone.

The visible capex is only part of the commitment. Moody's has estimated about $1.2 trillion in data-centre lease commitments across six major firms, with more than $820 billion on facilities not yet delivered, and so not yet on balance sheets. Morgan Stanley puts the hyperscalers' undiscounted commitments, including purchase contracts, above $2.7 trillion, roughly three years of operating cash flow. This is not concealment: the facilities don't exist yet, and Moody's treats the rents as debt-like. It does mean today's balance sheets understate what boards have signed up for. If the AI revenue gap doesn't close, those commitments become real liabilities, and for non-cancellable leases and take-or-pay contracts, the operator can't walk away.

There is a counter-view, and it depends on whose risk you mean. Janus Henderson argues that the bigger risk through 2030 is under-delivery, not overbuild: too little power reaching the projects that need it. That is an investor's risk, and for investors in utilities and power infrastructure, scarcity can mean pricing leverage. Ratepayers face a different risk. Scarce power raises their bills, and if projects stall, the grid upgrades built for them stay in the rate base. The Janus Henderson analysis doesn't address them.

A board should ask who bears each risk.

Contracts that make operators carry the cost, such as take-or-pay terms and direct infrastructure contributions, are how the answer gets decided.

Latin America: the advantage and the bill

Arizton estimates Latin American data-centre construction at about $3.0 billion in 2025, rising to about $6.9 billion by 2031. ByteDance's reported $39 billion campus in Ceará, Brazil, shows the ambition.

Chile shows the tension. Counts of its facilities vary with the definition used. The national data-centre plan counted 22 medium and large facilities in late 2024, with about 30 more announced. By capacity, JLL puts Chile at 166 MW, about 15% of the region and third behind Brazil and Mexico. The grid operator projects data-centre demand rising from about 325 MW in 2025 to about 1,207 MW in 2030, roughly 3.7 times, mostly in the Santiago region. That region has endured more than 15 years of mega-drought.

Chile's renewable power costs about $50/MWh, against more than $100 in the US.

That advantage is real. For cross-border M&A teams, the task is to map the grid and water assumptions behind it and stress-test them

against drought and regulatory change.

Five questions for every term sheet

1. Who pays for the grid upgrade, and is that allocation locked in by contract or open to regulatory revision?
2. What does the site draw on, and what does it sit beside? Water withdrawal under 1.5°C and 2°C basin scenarios, plus the distance to the nearest homes and the noise controls in the site plan.
3. How much contracted revenue rolls off within three years? Hardware refreshes far faster than the building ages.
4. What happens to the project if local tax incentives are revised?
5. What is the exit if the AI revenue gap does not close on schedule?

Data centres are not bad investments. They are complex industrial assets, and

too many deals and incentive packages price them as something simpler.