Data center vacancy: the overbuilding test for the AI bubble

What does data center vacancy measure in the AI bubble debate?

Data center vacancy is the share of finished capacity that has no customer contract. A market with 5% vacancy has nearly everything leased, while a market with 25% vacancy has a quarter of its capacity waiting for a tenant. Vacancy turns the abstract idea of overbuilding into one percentage that you can follow for years and compare between markets.

Vacancy answers a different question than spending, and it is harder to flatter with accounting choices. Spending tells you how much money entered the sector. Vacancy tells you whether that money produced space that customers want. In the AI bubble debate the 2 can diverge for a long time, and the point where they separate is the point where a boom starts to look like a bubble.

Does leased space mean used space in AI data centers?

A signed lease shows that a customer reserved capacity, and it says little about whether the customer uses it. Hyperscalers and AI labs often reserve space ahead of need. Vacancy can therefore read low while many halls sit partly idle. Ask each operator for 2 numbers: the share leased and the share with equipment powered on.

The gap between the 2 numbers is a useful signal. A small gap that closes within a year means customers are ramping as planned. A wide gap that persists means they reserved more than they needed, and leases may later be cut, delayed, or left unrenewed. Treat leased-but-idle space as a risk that vacancy figures alone cannot show, and list it as its own line in your notes.

How does absorption show whether AI data center demand is real?

Absorption is the net amount of space that customers take up in a period, measured in megawatts. If absorption is higher than new deliveries, vacancy falls and the market tightens. If deliveries are higher than absorption, vacancy rises. This single comparison tells you each quarter whether demand is catching supply or falling behind it.

Look at the trend over at least 4 quarters, because single quarters swing with the timing of large deals. A few big leases can make one quarter look strong. A steady run of positive absorption across many customers is more convincing. Check also whether absorption comes from many tenants or from the same few names repeating. Broad demand across industries and countries supports the boom reading far more than repeated deals between the same parties.

What do rents tell you about data center vacancy and the AI bubble?

Rents reveal how scarce space is. Rising rents with low vacancy mean that tenants compete for capacity, which is the clearest boom signal. Falling rents with rising vacancy mean that owners compete for tenants, which is the clearest overbuilding signal. Rents move after vacancy, so they confirm a trend that vacancy has already started to show.

Look beyond the headline rent. Free months, fit-out contributions, and discounts for large commitments lower the real price without changing the quoted one. Ask brokers and operators for effective rent, which spreads those concessions over the term. A rising number of concessions while headline rents hold is a quiet sign of softening demand, and brokers often mention it before the numbers move.

How does the construction pipeline change data center vacancy?

Vacancy today reflects past decisions, while the pipeline shows the vacancy of tomorrow. Space under construction will arrive within the next few years, depending on the project. If the pipeline is large compared with recent absorption, vacancy will rise unless demand accelerates. Divide pipeline megawatts by yearly absorption to see how many years of demand are already being built.

Check how much of the pipeline is pre-leased. Space with a signed tenant before completion carries less risk than space built on speculation. A pipeline that is mostly pre-leased supports the boom reading. A pipeline built without tenants relies on future demand, which is exactly what the AI bubble debate questions. Record the pre-leased share for every large project and update it as construction advances.

How do you compare data center vacancy across Singapore, Hong Kong, and ASEAN?

Each hub has its own supply limits, customer mix, and regulatory setting. A tight hub can show low vacancy because supply is limited, not because demand is extraordinary. A looser hub can show high vacancy because it is new and still filling. Compare every hub with its own history over several years rather than with a regional average.

For decisions in Singapore, Hong Kong, or other ASEAN markets, collect vacancy, absorption, and pipeline from brokers and operators for each hub separately. Note the date of each report and the definition it uses, since some count only operating sites while others include planned ones. Consistent definitions make the quarter-to-quarter comparison meaningful, and they stop a change of method from looking like a change of market.

How can you turn data center vacancy into an AI bubble test?

Build a 4-line test: vacancy trend, leased versus used gap, absorption against deliveries, and effective rent trend. Mark each green, amber, or red with its date. 3 or more reds in one quarter point to overbuilding. Repeat the exercise every quarter, keeping the thresholds fixed, so the sequence of results shows the direction of travel.

The EQOS Asia AI study scores the enterprise AI market at a φ(S) of 0.337, a fragile zone, with Stability at 0.412 (April 2026). A fragile system works until a shock arrives, such as a regulation change, an incident, or a recession. The method page explains the scoring, and a 30-minute session on the contact page applies it to your own data center 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