Investing in the Walls: The Rise of AI Data Center Real Estate

Is the AI data center construction boom creating a dangerous overcapacity problem? The risk is real, but current market data does not yet show broad near-term overcapacity. North American data center vacancy remains at a record-low 1.4%, more than 80% of capacity currently under construction is already preleased, and electricity demand from data centers is projected to more than double globally by 2030.

The greater risk may emerge after 2030 if developers continue building at today’s pace while AI hardware becomes more efficient, compute prices decline, AI adoption slows, or hyperscalers reduce capital spending.

McKinsey’s July 2026 analysis reaches a similar near-term conclusion, stating:

“the greater near-term risk might be underbuilding rather than overbuilding.”

Its analysis estimates U.S. data center IT power demand could grow approximately 27% annually through 2030, reaching 121 GW.

The key question for investors is therefore not simply how many data centers are being announced, but whether future supply will exceed economically sustainable demand.

AI Data Center Overcapacity Risk at a Glance

Sources: CBRE, IEA and McKinsey.

1. Data Center Construction Is Growing at Extraordinary Speed

There is a legitimate reason investors are asking about an AI data center bubble. CBRE reports that supply across major North American markets increased 33.7% year over year in the first half of 2026, reaching 10,903 MW.

Construction activity increased another 24.8% to 7,481 MW. Meanwhile, the IEA says global investment in data centers had already reached approximately $500 billion in 2024, nearly doubling since 2022.

This creates a classic infrastructure-cycle concern:

Shortage → Rising prices → Investment boom → Massive construction → Supply catches demand → Oversupply

The critical question is where the AI industry currently sits in that cycle.

2. Today’s Vacancy Numbers Don’t Look Like Overcapacity

If the market were already significantly overbuilt, one would expect rising vacancy and falling prices. Current North American data suggests almost the opposite.

CBRE reports that despite the 33.7% increase in primary-market supply, vacancy fell to a record 1.4% in H1 2026, compared with 1.6% one year earlier.

Net absorption increased 11.7% to 1,456.2 MW.

In simple terms:

New capacity is being built rapidly, but customers are absorbing it rapidly too. More importantly, 80.4% of capacity currently under construction is already committed to customers, up from 74.3% one year earlier.

CBRE estimates that less than 1,500 MW of future capacity remains available across its primary markets—roughly six months of supply at the current absorption rate.

That is not normally what an immediate capacity glut looks like.

3. Announced Projects Can Exaggerate Future Supply

One of the biggest mistakes in evaluating AI overcapacity is treating every announced project as if it will be built.

Consider:

Project announced → Land secured → Power application → Financing → Customer commitment → Construction → Energization → Operation

A project can fail at almost every stage.

Recent research highlighted by Barron’s illustrates how extreme this distinction can become. Around 509 GW of U.S. data center projects were reportedly planned as of July 2026, but Columbia Business School professor Stijn Van Nieuwerburgh estimates roughly 45% of the projects in his analysis could ultimately be abandoned.

The precise outcome remains uncertain, but the principle is crucial:

Announced capacity ≠ Completed capacity

and:

Interconnection request ≠ Operating data center

McKinsey similarly warns that swollen grid-interconnection backlogs can create a distorted picture of actual future demand.

4. Electricity Is Restricting How Quickly Supply Can Grow

AI data centers require enormous amounts of power. That creates a natural brake on construction. The IEA estimates global data center electricity consumption was around 415 TWh in 2024, approximately 1.5% of global electricity consumption.

Its base case projects consumption reaching approximately:

945 TWh by 2030

or just under 3% of global electricity consumption.

Accelerated servers, primarily driven by AI adoption, are projected to increase electricity consumption by approximately 30% annually under the IEA’s base case.

The problem is timing.

The IEA notes that a data center might become operational within two to three years, while major electricity infrastructure can take considerably longer to plan and construct.

Therefore:

Capital + GPUs + Land + Customer ≠ Operating data center

unless power is available.

5. AI Efficiency Could Eventually Reverse the Shortage

Perhaps the strongest argument for future overcapacity isn’t slowing AI adoption. It is improving efficiency.

Suppose today’s AI workload requires:

1,000 GPUs + 10 MW

Future chips and optimized models might perform substantially more useful computation using the same electricity. This means AI usage could continue growing while physical infrastructure demand grows more slowly.

The IEA explicitly models this uncertainty through scenarios including:

• Lift-Off.
• Base Case.
• High Efficiency.
• Headwinds.

These scenarios consider differences in hardware efficiency, software improvements, AI adoption and energy bottlenecks. The implication is important:

AI demand growth ≠ Data center demand growth at the same rate.

Read Also: Evaluating the Risk of Overcapacity in AI Data Center Construction

6. The Real Metric Is Utilization

Buildings themselves don’t generate attractive returns. Customers using computing capacity do. Consider two hypothetical 100 MW AI data centers.

Facility A

Capacity: 100 MW
Utilization: 90%
Effective utilized capacity: 90 MW

Facility B

Capacity: 100 MW
Utilization: 40%
Effective utilized capacity: 40 MW

Both represent identical headline construction. Their economics are completely different.

Investors should therefore monitor:

1. Occupancy.
2. GPU utilization.
3. Compute rental prices.
4. Revenue per megawatt.
5. Lease renewals.
6. Contract cancellations.
7. Customer concentration.
8. Power costs.

Falling utilization accompanied by falling compute prices would provide much stronger evidence of overcapacity than project announcements alone.

7. Preleasing Protects Developers—but Creates Another Risk

With more than 80% of current North American construction already preleased, developers have considerably more revenue visibility than they would under purely speculative construction.

But that introduces another problem:

Customer concentration.

Imagine a $5 billion facility where one hyperscaler contracts 70% of capacity. Vacancy risk may be low.  Counterparty concentration is high. If that customer reduces capacity requirements, renegotiates its contract or encounters financial difficulties, project economics can change rapidly.

Therefore, a heavily preleased market isn’t automatically a low-risk market. The risk has simply moved from:

“Will anyone rent this capacity?”

to:

“Will today’s major customers continue needing this much capacity?”

8. Geography Could Produce Overcapacity Before the Global Market Does

There may never be one moment when the entire global data center industry becomes oversupplied. Instead, overcapacity could appear region by region.

One market might have:

Cheap power + Strong fiber + Hyperscaler demand + Low vacancy

while another has:

Expensive electricity + Weak connectivity + Excess construction + Low utilization

The IEA notes that almost half of U.S. data center capacity is concentrated in just five regional clusters. That geographic concentration makes local electricity and property-market analysis essential.

An investor shouldn’t ask only:

“Is AI infrastructure overbuilt?”

Ask:

“Is this particular market overbuilt?”

9. Watch What Happens After 2030

Near-term demand is comparatively visible. Hyperscalers have announced capital expenditure, developers have secured contracts, and many projects have entered construction or grid-interconnection processes.

After 2030, forecasting becomes much harder. McKinsey expects growth to slow across its scenarios between 2030 and 2035 because of factors including:

Potential chip-demand saturation + Slower AI adoption + More energy-efficient next-generation hardware.

This creates the greatest long-term overcapacity danger. Projects financed today may still be operating in:

2035, 2040 or 2045.

Today’s shortage does not guarantee tomorrow’s shortage.

10. Warning Signs of AI Data Center Overcapacity

Investors should watch for several indicators simultaneously:

Falling vacancy + high preleasing currently suggest tight markets.

But the picture would change if we saw:

1. Vacancy rising sharply.
2. GPU rental prices falling persistently.
3. Hyperscalers cancelling leases.
4. Capital expenditure being cut.
5. Data center rents declining.
6. Utilization dropping.
7. Construction continuing despite weaker demand.
8. Distressed data center debt increasing.
9. Large projects being abandoned.
10. AI efficiency improving faster than workload growth.

Several of these happening together would be much more meaningful than any single metric.

Final Word

Yes—but today’s shortage and tomorrow’s potential surplus can both be true.

Current North American fundamentals remain extraordinarily tight. Supply increased 33.7% year over year, yet vacancy remained at a record-low 1.4%, while more than 80% of construction capacity is already preleased.

Globally, the IEA expects data center electricity consumption to more than double to approximately 945 TWh by 2030, with AI as the most important growth driver.

That supports continued infrastructure investment. But today’s investment decisions are creating assets expected to operate for decades. The longer-term equation is therefore:

Future AI demand + Compute efficiency + Power availability + Construction pipeline + Utilization + Pricing = Overcapacity risk

The warning sign won’t simply be cranes constructing another data center.

It will be:

New capacity continuing to arrive after utilization, pricing and customer commitments begin weakening. That is when an infrastructure boom can become an infrastructure glut.

For investors evaluating AI data centers, the better question is therefore not:

“How much capacity is being built?”

It is:

“How much economically productive capacity will customers actually need when these projects become operational?”

That distinction may ultimately separate valuable AI infrastructure from stranded capital.

FAQ

Is there currently an AI data center bubble?

Current operating-market evidence doesn’t establish widespread near-term overcapacity. North American primary-market vacancy was only 1.4% in H1 2026, while 80.4% of under-construction capacity was already preleased.

How fast is data center demand growing?

The IEA projects global electricity consumption from data centers rising from approximately 415 TWh in 2024 to 945 TWh by 2030.

Could AI efficiency create excess capacity?

Yes. Faster chips, more efficient models and improved infrastructure can increase computing output without requiring equivalent increases in electricity and physical capacity.

What is the strongest indicator of overcapacity?

No single metric is definitive. A combination of rising vacancy, falling utilization, declining compute prices, cancelled contracts and reduced hyperscaler spending would provide stronger evidence.

Source: mckinsey.com