Are we already building too many AI data centers?

Why are people suddenly worried about building too many AI data centers?

AI data center spending has become so enormous that even a small mistake in demand forecasts could leave companies with billions of dollars in unused or unprofitable infrastructure.

The scale of the investment boom explains the concern. Following their second-quarter 2026 results, Amazon, Alphabet, Microsoft, and Meta were collectively planning roughly $720 billion to $745 billion in annual capital expenditure. That includes AI processors, servers, networking equipment, buildings, and other infrastructure. Capex plans of $720 billion or more are huge bets, and we follow which investors are making the biggest ones in our data centers market report.

We're talking about annual spending comparable to the entire economic output of a medium-sized country. And that is before counting specialized cloud companies such as CoreWeave and Nebius, independent data center developers, and electricity providers.

The concern has become more immediate because financial pressure is appearing alongside the construction boom. CoreWeave's borrowing costs have drawn particular attention lately. Some of its bonds have traded at yields above 15%, even as the company reports rapidly growing revenue and a substantial backlog of customer contracts.

There is also growing resistance to new facilities. Communities across the United States are questioning how data centers will affect electricity prices, water supplies, and local infrastructure.

What would prove that we're building too many AI data centers?

We would be building too many AI data centers if companies consistently struggled to fill completed facilities or earn enough from them to cover construction, equipment, electricity, and financing costs.

The distinction between physical oversupply and financial overbuilding is important here.

Physical oversupply means too much usable capacity is available relative to customer demand. We would expect higher vacancy, falling rental prices, and difficulty securing tenants.

Financial overbuilding can happen much earlier. A developer might secure a customer and build a fully occupied facility, yet still lose money because construction costs, borrowing expenses, or equipment replacement costs turn out to be too high.

AI data centers are especially exposed to this second problem because their assets age at different speeds. Buildings and electrical infrastructure can remain useful for decades, while advanced GPUs may become much less competitive within a few years.

There is also a difference between a data center lease and actual computing demand. A company can rent space to another cloud provider without knowing whether the GPUs eventually installed there will generate attractive returns.

For that reason, we need to look at operational vacancy, construction commitments, customer revenue, computing prices, and profitability together.

Are AI data centers already sitting empty?

No. Existing data centers in North America's biggest markets are currently almost fully occupied, despite an extraordinary increase in construction.

CBRE's North America Data Center Trends report, published in August 2026, recorded vacancy of just 1.4% across eight primary markets in the first half of the year. That was down from 1.6% a year earlier and represented a new record low.

What's remarkable is how much supply grew during the same period. Operational capacity reached 10,903 MW, an increase of 33.7% year over year. Net absorption, which measures newly occupied capacity after accounting for space vacated, reached 1,456 MW during the first half, up 11.7%.

Developers added capacity at record speed, and customers continued taking it almost as soon as it became available.

Northern Virginia, the world's largest concentration of data centers, makes the shortage particularly clear. Its operational inventory reached approximately 4,497 MW, yet only 10.8 MW remained available. That corresponds to a vacancy rate of about 0.2%.

Hillsboro, Oregon, had just 1.1 MW of available capacity, while Phoenix had approximately 12.1 MW.

CBRE also reported that average asking rents for deployments requiring 3–10 MW rose 8.3% over the year. Higher prices alongside falling vacancy are difficult to reconcile with an industry already suffering from excess physical capacity.

One limitation is worth keeping in mind: CBRE measures wholesale data center markets, including traditional cloud and enterprise computing. These figures do not isolate AI workloads or tell us whether individual GPUs are fully utilized.

Still, today's operational data center market shows remarkably little evidence of widespread oversupply.

Are companies building AI data centers faster than customers can fill them?

AI data center construction is expanding extremely fast, but most capacity being built in major North American markets already has customers lined up.

CBRE reported 7,481 MW under construction across its eight primary markets in the first half of 2026, an increase of 24.8% year over year. That is equivalent to roughly 69% of the capacity already operating.

At first glance, adding almost 70% to the existing inventory looks excessive.

But 80.4% of that construction was preleased, up from 74.3% a year earlier. We calculate that approximately 6,015 MW had tenant commitments, leaving around 1,466 MW available.

That remaining amount is close to the 1,456 MW absorbed during the first six months of 2026. At the same pace, the entire unleased construction pipeline would represent roughly six months of demand.

It's an unusually strong position for a market expanding this quickly.

There are limitations. Some commitments may be conditional, tenants might not occupy facilities immediately, and construction schedules can change. A prelease also says little about the profitability of the computing hardware that will eventually operate inside.

Even so, the share of construction already committed to customers has increased rather than decreased. Developers appear to be having more difficulty delivering new facilities than finding tenants for them.

North American data center indicator	H1 2026	What it tells us
Operational capacity	10,903 MW	Existing supply
Operational capacity growth	33.7% YoY	Very rapid expansion
Vacancy rate	1.4%	Record-low availability
Net absorption	1,456 MW	Strong first-half leasing demand
Capacity under construction	7,481 MW	Large development pipeline
Construction growth	24.8% YoY	More capacity being developed
Share of construction preleased	80.4%	Most projects have customers
Estimated preleased construction	6,015 MW	Calculated from CBRE data
Estimated unleased construction	1,466 MW	Roughly six months of absorption
Rental growth for 3–10 MW deployments	8.3% YoY	Prices are still rising

Are Amazon, Microsoft, Google, and Meta spending too much on AI infrastructure?

Amazon, Microsoft, Google, and Meta are making unusually aggressive AI infrastructure investments, but their existing businesses still give them considerably more financial flexibility than smaller data center operators.

Meta's second-quarter 2026 results show how extreme the spending has become.

Revenue reached $60.8 billion, up 28% year over year, while operating income declined from $20.4 billion to $18.8 billion. Its operating margin fell from 43% to 31%.

Meanwhile, capital expenditure, including finance lease principal payments, reached $31.1 billion during the quarter. That was equivalent to approximately 51% of Meta's quarterly revenue.

Not every dollar went toward AI data centers. The figure covers broader infrastructure and equipment spending. Nevertheless, it shows how much cash Meta is committing to its future computing requirements.

Alphabet's results provide a different perspective. In July 2026, the company reported quarterly revenue of $119.8 billion, up 24%, while Google Cloud revenue jumped 82% to $24.8 billion.

That growth gives Alphabet a visible commercial reason to keep expanding cloud infrastructure. Google's established advertising business also provides substantial financial support.

Microsoft and Amazon benefit from similar advantages. Their cloud divisions serve existing business customers, and their infrastructure supports established software and internet services alongside newer AI applications.

We would be more concerned about the same level of spending at a company whose entire business depended on renting out GPUs.

Big Tech can afford to build capacity ahead of demand. Whether shareholders will eventually receive attractive returns on all that additional investment remains a separate question. Whether shareholders ever get paid back for all that spending is still unanswered, and we go through the biggest unsolved problems in our report covering data centers.

Is AI computing demand actually growing fast enough to justify the data center boom?

Yes. AI computing demand is growing fast enough to support major new data center construction, although the long-term profitability of that demand is much less certain.

Google Cloud offers one of the clearest examples. Its 82% revenue growth in the second quarter of 2026 came from a combination of enterprise AI infrastructure, AI applications, and traditional cloud services.

CoreWeave provides another useful measure. The specialized AI cloud company reported approximately $104 billion in revenue backlog at the end of June 2026, excluding more than $25 billion of additional net customer commitments secured early in the following quarter.

Its quarterly revenue also rose from $1.212 billion to $2.575 billion year over year, an increase of approximately 112%.

Those figures demonstrate that companies are signing substantial computing contracts and paying for services already delivered.

However, the same customer demand can appear in several companies' figures.

An AI developer may rent capacity from CoreWeave, which leases data center space from another operator and purchases Nvidia GPUs. A hyperscaler might also buy services from CoreWeave to serve its own customers.

We cannot add every cloud contract, hardware order, and data center lease together and treat the result as independent demand from AI users.

CoreWeave's backlog deserves particular care. It includes future contractual revenue subject to service availability and delivery obligations. It represents business the company expects to recognize over time, rather than cash already earned.

Even with that distinction, we can see demand in cloud revenue, signed computing agreements, and physical data center leases. What we cannot yet establish is whether the companies ultimately paying for AI applications will generate enough economic value to support every layer of infrastructure investment.

Could cheaper AI models make all these data centers unnecessary?

Cheaper and more efficient AI models could reduce computing requirements per task, but there is currently little evidence that they will reduce total demand for AI data centers.

Model efficiency is improving through smaller specialized models, better algorithms, lower-precision calculations, and more powerful chips.

These advances mean that tasks requiring substantial computing power today may become much cheaper to perform.

But cheaper computing also makes AI practical for more applications. Businesses can afford to process additional documents, automate more workflows, and serve larger numbers of users.

The distinction between training and inference is especially important.

Training involves the massive computing runs used to develop models. Inference refers to running those models when people or software use them.

As AI becomes part of programming, customer service, research, search, and business applications, inference could account for an increasingly large share of demand.

AI agents may amplify that trend. An agent writing software or reviewing complex documents can make dozens of model calls, check its work, use external tools, and repeat unsuccessful steps.

That creates more computing activity per completed task. But businesses will eventually care about how much each successful task costs, so inefficient agents cannot justify unlimited spending.

The International Energy Agency's updated 2026 outlook projects electricity consumption from AI-focused data centers to triple between 2025 and 2030, even after accounting for improvements in technology.

We should take the possibility of much cheaper AI computing seriously, particularly for owners of older GPUs. For now, however, lower costs appear more likely to change which infrastructure is profitable than to eliminate the need for additional infrastructure altogether.

Is electricity stopping AI data centers from being built?

Electricity shortages are currently blocking or delaying AI data center projects, often even when developers have customers ready to use the capacity.

The problem starts with timing.

A developer can purchase land and announce a new campus relatively quickly. Supplying hundreds of megawatts of reliable electricity may require new substations, transformers, transmission lines, and generation capacity.

Those projects often take considerably longer than constructing the data center buildings themselves.

The International Energy Agency's 2025 Energy and AI report estimated that grid constraints could delay roughly 20% of planned global data center projects through 2030.

Its updated 2026 analysis continued to identify electricity infrastructure and equipment bottlenecks as major reasons why the most aggressive data center growth forecasts may be difficult to achieve.

We can already see the consequences in Northern Virginia, where CBRE reported only 10.8 MW of available data center capacity despite a construction pipeline exceeding 2,400 MW.

Local opposition is making the situation harder. Communities across the United States have challenged major projects over electricity prices, water consumption, noise, and the cost of supporting infrastructure.

For example, the proposed Green Data campus involving New Mexico Tech in Socorro faced problems involving land, energy arrangements, and community concerns. The university ultimately declined to continue the arrangement.

The project had been presented as a potentially enormous development, with approximately 2 GW of planned capacity. Its failure illustrates the distance between announcing a campus and actually constructing one.

Similar obstacles have appeared in markets where grid connections, permitting, and infrastructure upgrades cannot keep up with developers' plans.

We should be careful about treating every reported delay as a cancellation. Large campuses are often developed in phases, and an announced megawatt figure may include capacity scheduled many years into the future.

For developers, the financial consequences can be severe. Existing facilities with secure power connections become more valuable, while unfinished projects can accumulate financing costs without generating rental income.

Is CoreWeave showing what could go wrong with the AI data center boom?

CoreWeave is one of the clearest examples of how an AI infrastructure company can grow extraordinarily fast while taking increasingly large financial risks.

Its second-quarter 2026 revenue reached $2.575 billion, up approximately 112% year over year.

Yet CoreWeave reported a $626 million net loss for the quarter, compared with $290 million a year earlier. With revenue doubling and losses growing, we compare which business models are actually winning in our data centers report.

Net interest expense increased from $267 million to $640 million. That means interest consumed an amount equivalent to almost 25% of quarterly revenue.

The company's adjusted EBITDA doubled from $753 million to $1.51 billion, but adjusted operating margin dropped from 16% to 5%.

Depreciation and amortization also increased sharply, from $559 million to approximately $1.39 billion. This reflects the growing cost of equipment and infrastructure being recognized through the income statement.

The financing concerns have become more visible lately. CoreWeave's bonds have traded at yields above 15%, with some securities reaching approximately 16% in secondary-market trading.

Those yields indicate that investors are demanding substantial compensation for lending to the company, although secondary-market yields do not mean all its existing debt suddenly carries those rates.

A company can have more than $100 billion of contracted future revenue and still require new financing to build the infrastructure needed to deliver it.

CoreWeave's experience gives us a preview of what financial overbuilding could look like across the specialized AI cloud industry.

Can AI data centers lose money even when they're fully occupied?

Yes. An AI data center can be fully leased and still deliver disappointing returns if borrowing, equipment, and operating costs absorb too much of the revenue.

The economics become easier to understand when we separate the physical building from the computers inside.

A data center owner renting powered space under a long-term lease may receive relatively predictable income. The tenant often bears much of the risk associated with the computing hardware.

An AI cloud provider renting out GPUs faces a different challenge. Its equipment must generate enough revenue during its useful economic life to cover purchases, maintenance, financing, and eventual replacement.

Nvidia's newer accelerator generations can improve computing performance and efficiency, which benefits customers but can make older GPUs less competitive.

An operator may therefore need to lower rental prices while its equipment continues depreciating.

CoreWeave's financial results show how quickly these costs accumulate. During the second quarter of 2026, depreciation and amortization exceeded $1.39 billion, more than half of quarterly revenue.

That accounting expense does not represent an immediate cash payment, but it reflects the cost of assets being consumed over time.

The risk becomes more serious when computing prices fall faster than operating expenses or debt obligations.

Which US data center markets are most exposed to overbuilding?

Atlanta currently stands out among major American data center markets because its construction pipeline is exceptionally large compared with its existing operational capacity.

CBRE reported approximately 2,882 MW under construction in Atlanta during the first half of 2026, an increase of 52.3% year over year.

Its operating inventory was approximately 1,803 MW. That means construction represented around 160% of the market's existing capacity.

Northern Virginia, by comparison, had approximately 2,420 MW under construction against 4,497 MW operating, or roughly 54%.

Atlanta's pipeline was therefore almost three times as large relative to its existing market.

That does not automatically mean Atlanta is building too much. Its vacancy remains low, and large hyperscale facilities are often secured by tenants before completion.

Still, the difference tells us where construction schedules and individual customer commitments could have the greatest impact.

A delayed major campus in Northern Virginia would affect an established market with an enormous installed base. A similar delay in a rapidly expanding market could represent a much larger share of its planned growth.

Other major markets face their own constraints. Phoenix combines strong demand with electricity and water concerns. Silicon Valley faces expensive land and power limitations. Hillsboro has extremely limited available capacity.

Market	Operational capacity	Available capacity	Key observation
Northern Virginia	4,496.5 MW	10.8 MW	Largest established market, extremely tight
Atlanta	1,803.2 MW	29.5 MW	Construction equals roughly 160% of inventory
Dallas-Fort Worth	1,432.8 MW	39.6 MW	Fast-growing market with limited availability
Phoenix	1,069.2 MW	12.1 MW	Strong demand, significant resource constraints
Chicago	910.6 MW	20.0 MW	Established cloud and enterprise market
Silicon Valley	509.2 MW	21.2 MW	Land and electricity are expensive
Hillsboro, Oregon	491.4 MW	1.1 MW	Almost no available capacity
New York / Tri-State	190.0 MW	13.7 MW	Smaller wholesale data center market

How much electricity will AI data centers actually need by 2030?

AI data centers are expected to consume dramatically more electricity by 2030, although the exact amount will depend on model efficiency, adoption, and how quickly new infrastructure can be completed.

In its updated 2026 outlook, the International Energy Agency projected global data center electricity consumption rising from approximately 485 TWh in 2025 to 950 TWh in 2030. Electricity demand nearly doubling by 2030 opens real gaps, and we point out where the biggest opportunities are in our report on data centers.

That's an increase of 465 TWh, or roughly 96%, over five years.

For comparison, the IEA expects data centers to account for approximately 3% of global electricity demand by 2030.

AI-focused facilities are expected to grow considerably faster than the industry overall, with electricity consumption tripling over the period.

The distinction helps explain why data center construction is continuing despite improvements in AI efficiency.

Even if each individual AI task consumes less electricity, the number and complexity of tasks could increase enough to drive total consumption much higher.

The IEA also makes clear that the outlook is uncertain. Its scenarios consider different rates of AI adoption, changes in hardware and software efficiency, electricity bottlenecks, and economic conditions.

We should therefore treat 950 TWh as a central projection rather than a fixed amount of electricity that data centers are guaranteed to consume.

Nevertheless, the increase is large enough to support significant construction. The real question is which proposed facilities can obtain power and deliver computing capacity at an acceptable cost.

Could AI data centers push electricity prices higher for ordinary households?

Yes. Large AI data centers can increase electricity costs for other consumers when utilities build expensive infrastructure to serve them and those costs are shared too broadly.

A facility requiring hundreds of megawatts may need new transmission lines, substations, transformers, and sometimes additional generating capacity.

These investments can benefit other electricity users, especially when upgrades improve reliability or expand available supply.

But the financial arrangements matter enormously.

If a utility builds infrastructure for a large data center that later cancels its plans, the utility may still need to recover the cost.

Without adequate contractual protections, part of that burden could fall on households and smaller businesses.

The risk is particularly relevant because data center development happens much faster than major electrical infrastructure projects.

A developer may change its investment plans before a power plant or transmission line built to support it becomes operational.

Electricity pricing concerns have therefore become an important part of local resistance to data centers.

The proper economic test should include more than whether an AI company is willing to pay rent or purchase electricity.

We also need to know who finances the additional infrastructure, what happens if the facility never opens, and whether other electricity customers are protected from the consequences.

A data center might be commercially attractive to its owner while still imposing excessive costs on the surrounding community.

Could AI data centers repeat the dot-com telecom bubble?

The AI data center boom could repeat one of the most damaging features of the telecom bubble: building useful infrastructure so aggressively that many investors lose money even as demand for the underlying technology keeps growing.

During the late 1990s, telecommunications companies invested heavily in fiber-optic networks because internet traffic was expected to explode. Fiber laid ahead of expected traffic is the comparison; we weigh whether data centers are overbuilt or still full in our data center market report.

That prediction proved broadly correct over the following decades.

The problem was that too much capacity was financed at once, often with large amounts of debt. Competition increased, bandwidth prices fell, and many companies could not generate sufficient revenue to repay investors.

AI infrastructure faces a similar possibility.

Companies are committing enormous amounts of capital to data centers, chips, and electricity infrastructure on the assumption that AI usage will keep rising rapidly.

Some of those predictions may prove accurate, while the financial returns still disappoint.

There are important differences, though.

The largest AI infrastructure investors today include Microsoft, Amazon, Alphabet, and Meta, which already operate highly profitable businesses.

Their financial position is considerably stronger than that of many speculative telecommunications companies during the dot-com era.

We also have unusually clear evidence of current data center demand, including record-low vacancy and extensive preleasing.

The closer comparison may be with specialized cloud companies that depend on debt and external financing to expand.

What would be the first sign that we've actually built too many AI data centers?

The clearest warning would be falling AI computing prices combined with weaker leasing commitments and rising unused capacity across several major markets.

We would look for these changes together because each can have an innocent explanation on its own.

GPU rental prices can decline when new processors become more efficient. That is not necessarily evidence of excess capacity if operators also lower their costs.

Vacancy can increase temporarily when several large projects open at once. A short-term change does not establish persistent oversupply.

And weaker construction activity might reflect electricity delays rather than disappearing customer demand.

What would concern us more is a sustained period in which developers struggle to secure leases, cloud providers report weaker utilization, and computing revenue fails to keep up with infrastructure investment.

The financing market might react before operational vacancy increases.

CoreWeave's recent bond-market pressure illustrates that possibility. Some of its debt has traded at yields above 15%, reflecting growing concern about leverage and future financing requirements.

Those conditions can force companies to reduce investment even when they still have customers.

On the physical side, CBRE's latest figures show that wholesale data center rents continue rising and new capacity is being absorbed rapidly.

Right now, financial stress is appearing before any broad evidence of unused operational infrastructure.

Are we already building too many AI data centers?

No, we are not yet building too many operational AI data centers overall. But the amount of money being committed to future capacity is starting to look more dangerous than the demand figures alone suggest.

The evidence for continued shortages is compelling.

North America's major data center markets have almost no available capacity. Most construction already has tenant commitments, and cloud companies are reporting substantial revenue growth.

The International Energy Agency also expects AI-focused data center electricity consumption to triple between 2025 and 2030.

There is a strong commercial case for building additional infrastructure.

The financial case for every project is much weaker.

CoreWeave's results show that revenue can more than double while losses and interest expenses also increase sharply. Its recent bond-market difficulties demonstrate how quickly financing conditions can become a problem for heavily leveraged AI cloud providers.

Electricity availability adds another complication. Developers may have customers ready to sign contracts but still face years of delays before facilities can operate.

We also found important differences between local markets. Atlanta's construction pipeline is around 160% of its existing operational inventory, while Northern Virginia's is closer to 54%. Those markets cannot be expected to face identical risks.

Our judgment is that the industry is currently more exposed to financial overbuilding than physical oversupply.

The next few years could therefore produce an outcome that initially seems contradictory: AI computing demand continues growing, data centers remain essential, and a meaningful number of infrastructure investors still lose money.

Some facilities will be completed too late. Others will carry excessive debt or require expensive equipment replacements before their owners have earned satisfactory returns.

We still need more AI data centers, but the assumption that every announced project will become a profitable one is already looking increasingly difficult to defend.