Is There an AI Data Center Bubble? – Stranded Asset Risk for Power
Every boom gets the bubble question eventually. Data center spending has climbed so fast that people who build power plants, string transmission lines and sign gas contracts are asking it out loud. They have good reason. A server can be replaced in a few years. A power plant is built to run for decades, and somebody has to pay for it either way. Is there an AI data center bubble, and if it pops, who gets stuck with the power assets?
Nobody can say for certain whether there is an AI data center bubble, but the current evidence points to real demand with pockets of speculation layered on top. If it pops, the people most exposed are whoever paid for long-lived power plants and lines built for a single customer. Vacancy in major North American markets is extremely low, but grid requests far exceed what will be built and many GPU contracts are much shorter than the plants behind them. Utilities and regulators are answering with minimum bills, collateral and phased builds.
What Would an AI Data Center Bubble Actually Look Like?
A bubble would look like capacity being built far faster than paying customers can fill it, financed on the assumption that growth never slows. The tell is not big numbers. It is big numbers without signed tenants and revenue behind them.
In practice, I would watch for a few signs:
- Vacancy rising in finished buildings, not just in planned ones.
- Tenants cancelling or deferring leases they already signed.
- Developers building "spec" capacity with no tenant and no power contract.
- Equipment orders and grid requests piling up with no customer named.
- Lenders loosening terms to keep deals moving.
None of those is proof on its own. Together, they are the pattern every overbuilt market has shown.
What Does the Demand Data Say Right Now?
The demand data still looks tight, not loose. In its September 2026 market report, CBRE said North American data center demand continues to outpace supply, with vacancy in primary markets around 1.4% and more than 80% of capacity under construction already preleased. CBRE counted about 7,481 MW under construction in those markets, and said power availability is the deciding factor for where deals land.
Those are not bubble numbers on their face. A bubble usually shows up as empty space. Here, operators are fighting over the space and the power to run it.
But tight vacancy can hide future risk. Preleasing means a tenant signed. It does not tell you how long the tenant is committed, whether the tenant can pay if its own business slows, or how much of the pipeline will ever break ground.
What Are the Warning Signs?
The warning signs are mostly about mismatched commitments and inflated pipelines. Three stand out.
Leases That Lapse
In early 2025, analysts at TD Cowen reported that Microsoft had let some data center agreements go. Data Center Knowledge covered the report that Microsoft dropped some AI data center leases, roughly a couple hundred MW voided and more than 1 GW of agreements allowed to lapse. Microsoft said its spending plan was unchanged. One company adjusting its portfolio is not a crash, but it was a reminder that announced capacity can shrink.
Short Contracts Inside Long Buildings
Many GPU cloud contracts run two to five years, while data center leases often run 10 to 15 years, according to JLL figures reported by Bisnow in 2025. Power plants and gas pipelines are financed over even longer periods. When the shortest contract in the chain is the one that carries the revenue, the longer assets depend on renewals nobody can promise.
Phantom Load in the Queue
Utilities are flooded with large load requests, and many are duplicates. A developer might ask three utilities for the same 500 MW and build in one place. Utility forecasts built on raw requests overstate demand, and power built for those forecasts could sit idle.
What Could Change the Demand Picture?
Efficiency is the wild card. If each unit of computing needs less power, the same buildings and power plants can serve more work, or the industry may need fewer of them. Google reported in August 2025 that the energy used for a median text prompt in its Gemini apps fell 33 times over 12 months, thanks to better software and hardware.
Historically, cheaper computing has led people to use more of it, not less, so efficiency gains do not automatically shrink demand. But they make long-range load forecasts shakier. Interest rates, chip supply and whether AI products earn enough revenue to justify the spending all matter too. A power plant financed over 20 years has to survive several of those cycles.
Which Power Assets Could Be Stranded?
The assets most at risk are the ones with long lives and a single customer. A stranded asset is one that loses its value before it has paid for itself. Here is how the main pieces compare:
| Asset | Typical life | Stranding risk | Why |
|---|---|---|---|
| GPU servers | A few years | High, but short-lived | Newer chips cut their value fast |
| Data center shell and cooling | Decades | Moderate | Can be refit for other tenants if the power is there |
| On-site gas generation | Decades | Moderate to high | Sized for one campus; engines and mobile turbines can be moved, large plants cannot |
| Utility plants and transmission for one customer | 30 years or more | High for ratepayers if unprotected | Costs can land on other customers if the load never shows |
The building itself is the safest piece if it sits on good power. A powered shell with cleared data center zoning, a substation and a utility agreement has value to many tenants, not just one. A half-built campus on a site with no firm power is the asset nobody wants.
Who Pays if Demand Falls Short?
That depends on the contract and the regulator. Without protections, a utility that builds a plant or a substation for a data center can recover the cost from all its customers if the data center leaves. That is the core worry consumer advocates raise.
Regulators have started closing that gap. In 2025, Ohio regulators approved an AEP Ohio tariff requiring new data centers over 25 MW to pay for at least 85% of the capacity they subscribe for up to 12 years, with exit fees for leaving early. Utilities and regulators in several other states have proposed or adopted similar terms. Our guide to large load tariffs explains how minimum bills, collateral and contract terms work.
For behind-the-meter projects, the risk falls on the developer and its lenders. If the tenant leaves, the generators either find another customer or sit idle. That is why many on-site power deals are structured with long take-or-pay commitments from the tenant.
Has This Happened Before?
Yes, and the closest parallel is the fiber optic overbuild around 2000. Telecom companies laid huge amounts of fiber on the belief that internet traffic would grow without limit. Many of those companies went bankrupt when the money ran out. The fiber stayed in the ground, and much of it was lit years later at a fraction of its original cost, carrying the traffic that eventually did arrive.
The lesson cuts both ways. The demand was real, just slower than the spending. The infrastructure kept its usefulness, but the first owners lost their investment. Power plants and substations could follow the same path: useful eventually, painful for whoever financed them at the peak.
How Are Developers and Utilities Limiting the Risk?
The smart money is building in stages and making customers carry their own risk. Common tools include:
- Phased construction, where each block of capacity waits for a signed tenant.
- Minimum demand charges, collateral and exit fees in utility contracts.
- Modular or mobile generation that can be redeployed if a campus stalls.
- Sites near large gas supply and transmission, so the power has other buyers.
- Designs that can serve conventional cloud and enterprise loads, not only the densest racks.
None of this is investment advice. If you are putting money into a project, talk to your own advisers and read the contracts. But from an operator's chair, the safest power asset is one with a creditworthy customer, a long contract and a site that someone else would want if that customer left.
Bubble or Boom, the Power Contract Decides Who Gets Hurt
Whether you call it a bubble or a boom, the damage in any downturn lands in one place: long-lived power infrastructure built for a single customer on a short commitment. Space is scarce today. The announced pipeline is still far bigger than what will be built.
Whatever the market does next, the same habit protects you on any project you build, finance or approve: look past the megawatt headline. Ask who signed, for how long, with what collateral, and what happens to the plant if they leave. Those answers will tell you more than any forecast.