Skip to content
ReviewRamble

The weekly brief

AI, reviewed, in your inbox.

  • New reviews of tools, models, MCP servers and repos, with the verdict up front
  • Re-checked prices and ratings when something changes
  • New analysis pieces as they are published

Models, agents, MCP servers and the tools actually worth your time.

One email a week. Unsubscribe with the link in any brief. How we use your email

Analysis AI infrastructure

What an AI data centre really costs, and who carries the risk

By ReviewRamble editors ·

The electricity is the cheap part of an AI data centre. The expensive part is the chips inside, which wear out in about five years, while the building and the debt that paid for it last much longer. That gap, between hardware that ages fast and money borrowed for decades, is where most of the risk in the AI boom sits, and a growing share of it is held by people who never signed up for it.

This explainer is for anyone who pays for an AI tool and wonders what stands behind the price. Every figure comes from the source we link to: the International Energy Agency (IEA), Lawrence Berkeley National Laboratory (LBNL), property firms JLL, CBRE and Cushman & Wakefield, Ireland's Central Statistics Office, the US Energy Information Administration (EIA), PJM's market monitor, the Federal Reserve and company filings. Forecasts are forecasts; we say whose they are.

How much money is going in

According to the IEA's Key Questions on Energy and AI, capital spending by the largest technology companies passed $400 billion in 2025 and is expected to rise by another 75% in 2026, which would put it near $700 billion. Meta alone told investors in its September quarter filing that it expects to spend $135 billion to $145 billion this year. Our news desk tracked the latest round of spending plans as they came out.

JLL's 2026 Global Data Center Outlook puts the bill for the whole industry at up to $3 trillion by 2030. It expects nearly 100 GW of new capacity between 2026 and 2030, doubling what exists, and roughly $870 billion of new debt to pay for the property alone.

What one megawatt costs

Data centres are priced per megawatt (MW) of power they can deliver to computers. One MW running all year is about what 800 average American homes use, on the EIA's figures below, and a large AI campus is measured in hundreds of MW or several gigawatts (GW, a thousand MW).

There are two very different bills. The first is the building: land, shell, power and cooling plant. JLL says the global average construction cost rose from $7.7 million per MW in 2020 to $10.7 million in 2025, and forecasts $11.3 million in 2026. That figure covers only the shell and core.

The second bill is what goes inside. JLL notes that tenants usually pay for the technology fit-out, and for AI that can reach $25 million per MW. Across the industry it expects tenants to spend $1 trillion to $2 trillion on GPUs and networking by 2030. So for an AI site, the computers can cost more than twice as much as the building that holds them.

A worked example: 100 MW for a year

Take a site with 100 MW of computing load, running flat out all year. There are 8,760 hours in a year, so the computers use 876,000 megawatt-hours (MWh). Add cooling and power losses, measured by a ratio called PUE (explained below). At Google's reported fleet average of 1.09, the site draws about 954,840 MWh. At the industry average of 1.54 that Google cites from the Uptime Institute, it draws about 1,349,040 MWh.

The EIA's average US industrial electricity price for 2025 was 8.62 cents per kilowatt-hour, or $86.20 per MWh. That puts the annual power bill at about $82 million for the efficient site and about $116 million for the average one. The efficient site's use matches roughly 88,000 American homes, at the EIA's average of 10,791 kWh a year.

Now compare that with the hardware. At JLL's figures, the building costs about $1.13 billion and an AI fit-out up to $2.5 billion more. Meta's 2025 annual report depreciates servers and network gear over five to 5.5 years and buildings over 25 to 30 years. Spread $2.5 billion of chips over five and a half years and you get about $455 million a year, more than five times the efficient site's power bill.

Two caveats. Real sites rarely run at full load all year; LBNL assumed 50% average use when it turned energy into power figures. And big buyers negotiate their own power prices, which can be well above or below the national average. The shape of the result holds either way: the chips are the dominant cost, and they age fastest.

If you want to see the same arithmetic at the scale of one machine, our GPU calculator compares the hourly cost of renting the cards a model needs, and our NVIDIA hardware guide covers what a GB300 rack costs.

How much electricity data centres use

The IEA estimates that data centres worldwide used about 485 terawatt-hours (TWh) in 2025 and will use about 950 TWh by 2030, roughly 3% of the world's electricity. Use by AI-focused sites triples over the same period. A TWh is a million MWh, so 950 TWh is about a thousand of our 100 MW example sites.

In the US, LBNL's 2024 United States Data Center Energy Usage Report found data centres used 176 TWh in 2023, 4.4% of US electricity, and projects 325 to 580 TWh by 2028, or 6.7% to 12.0%. The width of that range is itself the story: nobody knows how many GPUs will ship or how hard they will be run.

Where the power comes from

The IEA's Energy and AI report breaks down where data centres get their electricity today: coal about 30%, renewables about 27%, natural gas 26% and nuclear 15%. The mix depends heavily on the country. In the US, gas supplies over 40%, wind and solar 24%, nuclear around 20% and coal around 15%. In China coal supplies nearly 70%.

To 2030, the IEA expects renewables to meet nearly half of the extra demand, with gas and coal together meeting over 40%. Where grid connections are slow, some operators are building their own gas plants on site: the IEA reckons 15 to 27 GW of on-site gas may power data centres by 2030, mostly in the US, and that such plants must be oversized by 30% to 70% to be reliable.

Nuclear is the long bet. Technology companies have plans to finance more than 20 GW of small modular reactors, but the IEA expects the first ones to be commissioned only after 2030. For the next five years, the new electricity mostly comes from what can be built quickly, and that is wind, solar and gas. The IEA projects data centre emissions roughly doubling to about 350 million tonnes of CO2 by 2035, still about 2% of power sector emissions. Our news notes covered an MIT study on the air-quality cost of that electricity.

Water: why the numbers disagree so much

Ask how much water a data centre uses and you will get answers that differ enormously. Most of the gap comes from three choices: how the site is cooled, where it is, and whether you count the water used to make its electricity.

Evaporative cooling towers chill water by letting some of it evaporate, which saves electricity and uses water. Dry coolers and closed liquid loops use little water but usually more electricity. LBNL's simulations found water use varies most with the cooling system, with the highest figures at sites using water-cooled chillers without economisers, and climate matters too: the same design uses more in a hot, humid place.

The totals reflect that. LBNL estimates US data centres consumed about 66 billion litres on site in 2023, and that large cloud sites alone could consume 60 to 124 billion litres by 2028. But the water used at power stations to make their electricity was nearly 800 billion litres, about twelve times the on-site figure, at a national average of 4.52 litres per kWh.

For one company's view, Google's 2026 Environmental Report says its data centres consumed 10,523 million gallons, about 10.5 billion gallons, in 2025. Its campuses in The Dalles, Oregon, used 469 million gallons. Google says it chose water cooling there because local sources were not at high risk of scarcity, which is the honest version of the trade-off: water cooling is the efficient choice where water is plentiful and the wrong one where it is not.

PUE, explained

Power usage effectiveness (PUE) is the total electricity a site draws divided by the electricity that reaches its computers. A PUE of 1.0 would mean nothing is spent on cooling or power conversion; 1.5 means half a watt of overhead for every watt of computing. Our glossary entry has the detail.

LBNL estimates the average US data centre fell from 1.6 in 2014 to 1.4 in 2023. Google reports 1.09 across its fleet in 2025. The catch with PUE is that it rewards water cooling: a site can lower its PUE by evaporating more water. It also says nothing about whether the computers are doing useful work. Read it alongside water figures, never on its own.

Where capacity is concentrated

AI capacity clusters where power, fibre and land come together, and the biggest clusters are running out of room. In CBRE's first-half 2026 survey of eight primary North American markets, supply rose 33.7% in a year to 10,903 MW, yet vacancy fell to a record low of 1.4%. A record 7,481 MW was under construction and 80.4% of it was already leased.

Where the big clusters are

  • Northern Virginia, US: about 4,497 MW, the largest market in North America (CBRE)
  • Atlanta, US: about 1,803 MW, with 2,882 MW under construction, the most in North America (CBRE)
  • Dallas–Fort Worth, US: about 1,433 MW (CBRE)
  • Phoenix, US: about 1,069 MW (CBRE)
  • Chicago, US: about 911 MW (CBRE)
  • Oregon, US: about 491 MW around Hillsboro (CBRE)
  • Montréal, Canada
  • Alberta, Canada: more than 16 GW of requests in 2025 against a 1,200 MW interim grid limit to 2028 (AESO)
  • Dublin, Ireland: data centres used 23% of Ireland's metered electricity in 2025 (CSO)
  • London, UK: capacity up 21% in a year (CBRE)
  • Frankfurt, Germany: about 1,223 MW (CBRE)
  • Amsterdam, Netherlands: about 64 MW added in a year (CBRE)
  • Paris, France: about 667 MW (CBRE)
  • Stockholm, Sweden
  • Singapore: the highest rents among Asia-Pacific's major markets, about $403 per kW a month (CBRE)
  • Johor, Malaysia: about 1,110 MW in operation (Cushman & Wakefield)
  • Tokyo, Japan: more than 1 GW (CBRE)
  • Seoul, South Korea
  • Mumbai, India: about 890 MW in operation (Cushman & Wakefield)
  • Sydney, Australia: about 917 MW in operation (Cushman & Wakefield)
  • São Paulo, Brazil: about 537 MW, the largest in Latin America (CBRE)

The figures come from different firms and dates, so treat them as a rough ranking rather than a league table. The US markets are CBRE's first-half 2026 inventory figures; Frankfurt, Paris, London, Amsterdam, Tokyo, Singapore and São Paulo are from CBRE's Global Data Center Trends 2026, with first-quarter data; Johor, Mumbai and Sydney are Cushman & Wakefield's first-half 2026 operating figures. We found no comparable published figure for Montréal, Stockholm or Seoul.

Two places show what concentration does. In Ireland, the Central Statistics Office found data centres used 23% of all metered electricity in 2025, up from 5% in 2015 and more than every urban household combined at 18%. In Alberta, the grid operator set an interim limit of 1,200 MW of new large loads until 2028 after 29 projects asked for more than 16 GW. For scale, Edmonton's whole load is about 1,400 MW.

Who holds the debt

For most of the past decade, big technology companies paid for data centres out of their own cash. That is changing. The Federal Reserve's Financial Stability Report of 13 November 2026 says debt issued for advanced computing passed $680 billion by October 2026, about five times the $130 billion issued in all of 2023. Such borrowing went from under 6% of investment-grade non-financial corporate debt issuance in 2023 to about 31% in 2026.

The bigger number sits off the balance sheet. The Fed counted leases that have not yet started at over $1 trillion, and guarantees and other contract commitments at over $2 trillion, across Amazon, Alphabet, Broadcom, Meta, Microsoft, Nvidia and Oracle. These figures are not discounted, so they are not directly comparable with debt, but the Fed's point is plain: these firms' exposure to a downturn "may be substantially larger than suggested by their on-balance-sheet liabilities". Our news desk noted Nikkei's similar count earlier this month.

How the off-balance-sheet deals work: Meta's two campuses

Meta's two joint ventures show the structure clearly. In October 2025 it formed a venture with funds managed by Blue Owl Capital for its Hyperion campus in Richland Parish, Louisiana. Blue Owl's funds own 80% and Meta 20%, against about $27 billion of development costs. Blue Owl put in about $7 billion in cash, Meta took a one-time payout of about $3 billion, and part of the money came from bonds sold to PIMCO and other investors.

In July 2026 it repeated the pattern with BlackRock in El Paso, Texas: a 1 GW campus costing about $14 billion, 80% owned by BlackRock's funds, funded partly by $12.5 billion of debt, with capacity coming online from 2028. That works out at about $14 million of development cost per MW.

In both cases Meta leases the whole campus back. The leases run four years at first, with options that can stretch them to 20. To make lenders comfortable with such a short commitment, Meta guarantees the campus's value for the first 16 years: if it walks away and the site is worth less than a set threshold, Meta pays the gap, up to a cap. Because Meta does not control the ventures, it does not put their debt on its own balance sheet.

The September 2026 10-Q lists Meta's maximum exposure to the two ventures at $68.4 billion, including $41.5 billion of maximum residual value guarantees and $18.5 billion of lease commitments. Meta says those guarantee payments are not probable, so it records no liability for them.

The short lease is the point. It gives Meta the option to leave if AI demand cools, and the guarantee caps what that exit costs. The bondholders and fund investors on the other side are betting that a building designed for AI chips will still be worth its threshold in a decade. Some of that money is ordinary savings: as one economist argued, retirement accounts are part of the funding, and the Fed notes that the riskier slices of these deals sit with high-yield funds, life insurers and private credit vehicles, while large US banks have close to $1 trillion of exposure to the sector.

The risks, in plain terms

  • The chips age before the building. Meta writes servers off over five to 5.5 years and buildings over 25 to 30. Every new GPU generation makes the old one worth less, and the Fed notes that GPU makers sometimes guarantee the resale value of chips used as loan collateral. If that value falls faster than expected, the losses land on whoever lent against it.
  • Demand may not match the plans. The range in LBNL's 2028 forecast, 325 to 580 TWh, shows how uncertain usage is. PJM's market monitor says the data centre forecasts behind its latest capacity auction were unverified and overstated. A four-year lease lets a tenant leave long before the building is paid off.
  • Households may pay for grid upgrades. PJM runs the grid across a large part of the eastern US, Northern Virginia included. Its independent market monitor found that $6.3 billion of the $16.4 billion in its latest capacity auction, 38.2%, was a result of data centre load, and $29.4 billion across the last four auctions. In the first half of 2026 that added $11.11 per MWh, or 9.7%, to the total cost of wholesale power, a cost spread across all of the grid's customers. The monitor wants data centres moved into a separate auction so that they pay for their own capacity.
  • Local water and air. A site's water and emissions land on its neighbours, whatever its owner's global offsets say.

Some operators are offering to give something back. OpenAI has pledged to cut its data centre power use when the grid is strained, and the IEA suggests grid operators could offer non-firm connections in exchange for that kind of flexibility.

What this means for the price you pay

The price of an AI tool has to cover the chips' short life, not just their electricity. That is why providers push so hard to keep GPUs busy, and why fast, specialised hardware like that compared in our Cerebras, Groq and SambaNova guide competes on throughput per chip. If you rent compute yourself, GPU clouds such as Runpod (4.3) and Modal (4.5) pass that hardware cost through by the second, and token sellers such as Together AI (4.5) build it into the price per million tokens.

Screenshot of the Together AI website
Together AI, official site

The catch: most of the numbers here are forecasts or company-reported figures, and the most important ones, how long the chips stay useful and how much AI demand there will really be, are the least certain. The Fed itself says the full size of private credit in this sector is hard to measure.

Our take: think of an AI data centre as a fleet of delivery vans parked inside a warehouse. The warehouse lasts thirty years and the vans five, and the lender has been told the warehouse will always be full. The power bill is the fuel, real but small. When you judge whether the boom is sound, watch the vans and the lease, not the light switch, and when your local utility asks to raise rates for new grid capacity, ask who it is being built for.

The ReviewRamble brief

AI, reviewed. Once a week, in your inbox.

The verdicts that matter across AI, robotics and tech, with what to use and what to skip. No hype, no digest of everything.

Models, agents, MCP servers and the tools actually worth your time.

One email a week. Unsubscribe with the link in any brief. How we use your email