AI bubble burst and the power grid: stranded data centres
The essentials in 30 seconds
- A stranded asset is a power plant, grid line, or data centre built for demand that does not arrive.
- An AI bubble burst would strand the assets built ahead of signed customers and spare those with long contracts.
- Six checks sort the two groups: contracts, location, reuse, financing, timing, and flexibility.
- The EQOS Asia AI study records about $200B a year invested in AI and a φ(S) score of 0.337 (April 2026).
- This page is general information, not investment advice.
AI needs power, and power needs plants, lines, and permits that take years to build. If AI demand grows more slowly than forecast, some of that capacity waits for customers. This guide belongs to the page on what happens if the AI bubble bursts, and it explains the stranded asset test, as of October 2026. It is general information, not investment advice.
On this page
Compare the EQOS Asia offers and request a 30-minute pre-diagnostic.
What is a stranded data centre or power asset?
A stranded asset is an asset that loses much of its expected value before the end of its life. A data centre without tenants is one. A power line built for a campus that never fills is another. The asset still exists and still costs money to maintain, but the income it was built to earn arrives late, arrives low, or does not arrive.
Stranding needs two ingredients: a long build time and a forecast. Both are present in AI. Power and data centre projects take years, and they start from forecasts of demand. The EQOS Asia AI study records about $200B a year invested in AI and a φ(S) score of 0.337 in the fragile zone (EQOS Asia AI study, April 2026).
Why does the power grid plan so far ahead for data centres?
A grid operator plans lines and plants years before customers connect. It decides based on requests from developers, and developers file requests that exceed what they will finally build, because holding a place in the queue costs less than missing it. The queue therefore overstates real demand, and the overstatement grows when enthusiasm is high.
When the mood cools, requests drop out of the queue. Capacity that was planned for them may already be funded or under construction. The cost of that capacity moves to the other users of the grid or to the budgets of the utility, depending on local rules. Reading the rules tells you who carries the cost, and it is a document you can find before you commit to a site.
Which power contracts protect a data centre owner or a utility?
A take-or-pay contract obliges the customer to pay for reserved capacity whether or not it uses it. It protects the supplier. A flexible contract lets the customer scale down. It protects the customer. Read the minimum volume, the length of the contract, and the credit strength of the customer. Together they show who holds the risk if demand falls short.
Look for exit clauses and review dates. A 15-year contract with a review at year 5 is shorter than it sounds. A contract with a strong customer and a long fixed term is worth more than one with a young customer and an early exit right. The terms matter more than the headline length. Ask also whether the price per unit of power is fixed or follows the market, since a fixed price shifts risk to the supplier.
How do site location and reuse options change data centre risk?
A hall in a market with many potential tenants is easier to re-let than a hall in a remote area. Power access, network links, and distance from customers all make a location valuable. A secured power connection is the scarce part, and it can retain value even if the first tenant leaves, because a new tenant can use it.
Reuse matters too. A building that can host other users, from cloud to enterprise to network operators, holds value better than one built for a single customer. Ask how much work and money it would take to adapt the hall. A low adaptation cost lowers the risk of stranding.
Why does debt financing matter for a data centre after an AI bubble burst?
Stranding is partly a financial event. An asset with little debt can wait for customers. An asset with heavy debt cannot, because interest accrues each month. Check loan size against value, the date of each repayment, and the sponsor’s willingness to inject money. A patient sponsor turns a slow start into a delay, and an impatient lender turns it into a loss.
Equity-funded projects carry no repayment calendar, though they carry the cost of waiting. Debt-funded projects carry a calendar. The difference shapes who is forced to sell and when, and forced sales set prices for everyone else. Watch the debt maturity of owners as an early warning for a wave of sales.
Which grid investments hold their worth after the AI bubble bursts?
Assets with signed customers, secured power, low debt, and flexible use tend to stay valuable. Efficient assets also hold up, because customers choose lower running cost when budgets tighten. Power capacity that serves the wider economy, not only data centres, keeps its customers. A line that feeds a data centre and a district has a second source of demand.
Assets that rely on a single customer type, heavy debt, and rare approvals tend to suffer. The line between the two groups is clear in advance, and it can be traced with the six checks of this page: contracts, location, reuse, financing, timing, and flexibility.
How can you test your own energy exposure to an AI bubble burst?
Score each of the 6 checks from 0 to 2 and add them. A total of 9 or more shows an asset likely to keep its value. A total of 5 or less shows an asset that depends on demand arriving as forecast. Date the score and update it each quarter. This page gives no price forecast and no call to buy or sell.
To score your own case, use the Λ simulator for a first reading in 10 minutes. For a decision that involves AI spending, power contracts, or data centre projects in Asia, a 30-minute pre-diagnostic applies the EQOS Asia scoring and returns a ranked set of options.
See also in this guide
Frequently asked questions
What does stranded data centre mean?
It means a data centre built for demand that arrives late, low, or not at all. The building and its power link still exist and still cost money, while the income they were built to earn falls short. This page is general information, not investment advice.
Would an AI bubble burst strand power plants?
Some capacity planned for data centres could wait for customers. The risk is highest for capacity built ahead of signed contracts and lowest for capacity that also serves the wider economy. Local rules decide who carries the cost.
Why do grid queues overstate demand?
Developers file requests larger than what they will build because holding a place costs less than missing it. When enthusiasm cools, requests drop out, and the planned capacity may already be funded.
Which assets keep their value after a burst?
Assets with signed customers, secured power, low debt, and flexible use tend to keep value. Efficient assets also hold up because customers pick lower running costs when budgets tighten.
Test your knowledge
1. What is a stranded asset?
- An asset that loses much of its expected value early
- A fully leased hall
- A new power line with a signed customer
Show the answer
An asset that loses much of its expected value early.
2. Which contract protects the supplier?
- Flexible
- Take-or-pay
- Month to month
Show the answer
Take-or-pay.
3. How many checks does the page give?
- 2
- 12
- 6
Show the answer
6.
4. What scarce part keeps value if the first tenant leaves?
- A logo
- A secured power connection
- A lobby
Show the answer
A secured power connection.
5. What does heavy debt do to a slow start?
- Turns it into a loss
- Removes it
- Changes nothing
Show the answer
Turns it into a loss.
About the author
Stéphane Jambu
President, EQOS Forecast Asia
Stéphane Jambu leads EQOS Forecast Asia and works as a consultant in e-commerce, search visibility, and AI. LinkedIn profile
Sources: EQOS Asia AI impact study (eqos-asia.com/case-studies/ia-2026.html), EQOS Asia simulator page (eqos-asia.com/simulator.html) and contact page (eqos-asia.com/contact.html), consulted on 22 October 2026. Page updated on 22 October 2026.