Data center power demand and the AI bubble

What is data center power demand in the AI bubble debate?

Data center power demand is the electricity that AI facilities draw from the grid, measured in megawatts for a site and in terawatt-hours for a year. It ties directly to chips: more accelerators in service mean more load. Because electricity is billed every month, it records what a facility truly uses. Treat it as a meter reading, not as a forecast.

The AI bubble debate leans on power for one reason. Demand forecasts claim that AI usage will keep climbing for years. If the forecast is right, consumption rises quarter after quarter and new sites fill quickly. If it is too high, announced capacity waits for users. Power data lets you test the forecast against the meter instead of trusting the forecast.

How do you read megawatts in AI data center announcements?

Announcements mix 3 different numbers. Planned capacity is what a developer hopes to build. Contracted capacity is what a utility has agreed to deliver. Energized capacity is what customers can use today. Only the third number earns revenue. When a press release gives one total without naming which of the 3 it is, ask before you compare it with demand.

Convert every figure to the same unit and the same date. A site quoted in gigawatts, another in megawatts of critical IT load, and a third in grid supply cannot be added together. Keep a simple table with site, status, megawatts, and date. After 2 quarters the table shows how fast plans become power that you can measure.

What do grid connection queues say about AI data center demand?

Utilities keep waiting lists for new large connections. A long list shows many developers asking for power, which supports the boom reading. Yet a request is cheap to file, so a list also attracts duplicates. The same project can sit in 2 queues under 2 names, and the total then overstates real demand. Count projects, not requests, whenever the data allows it.

Read the queue by stage. Requests, studies, signed agreements, and energized sites form a funnel. A healthy funnel passes a steady share of projects from each stage to the next. A funnel that fills at the top and stalls in the middle signals plans without funding. Track the conversion share from request to energized site each quarter and write it down.

How do power purchase agreements test AI data center demand?

A power purchase agreement is a long contract to buy electricity at an agreed price. Signed by a customer with a real workload, it proves demand years ahead. Signed by a developer who has no tenant yet, it proves only intent. Check who signs, for how many years, and what happens to the price if usage falls short of the agreed volume.

Take-or-pay clauses deserve special attention. They oblige the buyer to pay for the volume whether or not the power gets used. That moves risk from the utility to the data center owner, and from there to its lenders. A growing stack of take-or-pay commitments beside flat consumption is a clear sign that plans have outrun demand.

How does chip efficiency change AI data center power demand?

Each chip generation tends to deliver more computing per watt. When software also improves, the same AI service needs less power than last year. Demand for AI can grow while demand for electricity grows more slowly, and forecasts that tie the 2 together one to one can overshoot. This is the strongest argument against reading every announced megawatt as future usage.

The opposite case also exists. Cheaper AI can raise usage so much that total power keeps climbing. You do not need to settle the argument to use the data. Track power per unit of AI work where a company publishes it, and track total power beside it. Falling intensity with rising total shows real growth, while rising plans with flat total show slack.

What does low power use say about the AI bubble?

Utilization compares the power a site actually draws with the power it is built to draw. For example, a hall designed for 100 megawatts that draws 40 is 40% used. High utilization means customers are filling the space. Low utilization after the first year means the build ran ahead of demand, and the owner pays for idle capacity every month.

Pair utilization with the ramp schedule. New sites fill in stages, so a low figure in month 3 is normal. The same figure in month 18 deserves questions. Ask for the ramp plan, compare it with actual draw, and watch whether the gap closes. A gap that stays open is the overbuilding test in its simplest form.

How can Singapore and ASEAN buyers score AI power demand?

Build a grid with 4 rows: announced versus energized megawatts, queue conversion, contract quality, and utilization. Mark each green, amber, or red, with the date. Hubs with tight grids or limited land can show high utilization and low supply at the same time, so compare each hub with its own history. 3 or more red rows in one quarter signal overbuilding.

The EQOS Asia AI study places the enterprise AI market at a φ(S) score of 0.337, a fragile zone, with Energy the strongest of 9 dimensions at 0.588 and Communication the weakest at 0.305 (April 2026). The method page explains the scoring, and a 30-minute session on the contact page applies it to your own AI infrastructure decision.

Frequently asked questions

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