Are we already building too many AI data centers?
In our data center market deck, you will find everything you need to understand the market
SUMMARY
No, we are not yet building too many usable AI data centers overall. We are already planning too many speculative projects, double-counting some future demand and financing part of the buildout on assumptions that will fail.
The physical market is still tight where customers actually want capacity. Vacancy in the largest North American hubs remains near record lows, rents are rising and completed megawatts are being absorbed quickly.
The headline pipeline exaggerates what will open. Operating sites, leased buildings, construction projects, grid requests and land options are often discussed as though they carried the same probability, when they clearly do not.
Even so, the construction pipeline is less reckless than the announcement pipeline. In North America, 92% of capacity under construction is reportedly precommitted, which limits the risk of large numbers of empty buildings opening at once.
The more serious risk is economic rather than physical. A data center can stay full and still be a poor investment if AI prices fall faster than computing costs or if chips depreciate before the capital invested in them earns an adequate return.
Utility forecasts are especially vulnerable to inflation because developers can shop one campus across several regions. The same intended project may appear in multiple load queues long before a final site is chosen.
Power shortages are acting as an accidental discipline mechanism. They delay strong projects too, but they also force weak proposals to survive years of permitting, equipment procurement, financing and local scrutiny before construction can begin.
AI demand is real and broadening beyond frontier-model training. Inference, coding tools, search, advertising, media generation and enterprise workloads give the next wave of facilities a credible demand base, especially near major cloud regions and population centers.
Efficiency will not necessarily reduce total computing demand, but it will wreck forecasts built on today’s energy cost per task. Some facilities may open into a market where the same workload needs far fewer chips, even as heavier reasoning, video and agentic applications push demand back up.
The weakest part of the boom is likely to be leveraged neoclouds and rigid single-customer projects. Hyperscalers can redirect capacity across several businesses; specialists have fewer escape routes when a site opens late, a customer changes plans or financing costs keep compounding.
Overbuilding is already visible locally. Double-digit vacancy in markets such as Querétaro, Hong Kong and Bogotá shows that global demand cannot rescue every location, every building design or every financing model.
The likely correction is selective rather than global: speculative campuses, duplicate power requests, poorly located sites and debt-heavy providers are in danger first. Well-located, adaptable and contracted capacity in the strongest markets remains scarce.
This market map, featured in our data center market deck, highlights top companies and startups in the data center market
Why does the AI data center boom look excessive now?
The AI data center boom looks excessive today because spending plans and power requests have raced far ahead of capacity that can actually open.
The International Energy Agency estimates that capital expenditure by five large technology companies exceeded $400 billion in 2025 and could rise another 75% in 2026. At roughly $700 billion, the implied annual total would be larger than worldwide investment in oil and gas production. JLL also counts more than 35 gigawatts of data center capacity under construction across North America, an amount comparable with the annual electricity consumption of a large European country.
Those numbers deserve scrutiny, but they mix commitments of very different quality. An operating hall full of GPUs, a signed lease, a building under construction, a utility application and a land option can all appear in the same discussion. Only the first three are close to becoming usable capacity. Developers frequently hold several possible sites while they wait to learn which one can secure power, permits and financing.
That gap explains why the boom can look both enormous and constrained. The industry has committed extraordinary sums, yet customers still struggle to find powered space in the main markets. The real test is whether completed facilities can earn enough from AI workloads to cover chips, electricity, depreciation and financing.
If you want more recent data on this point, please see our latest data center market report.
What would “too many AI data centers” actually look like?
A real AI data center glut would show up through empty powered halls, falling rents, weaker preleasing and painful write-downs. Those conditions remain uncommon today.
We need to watch three different problems. A physical glut appears when completed facilities cannot find users. An economic glut appears when the buildings stay busy but the revenue generated inside them fails to justify the cost. A grid-planning glut appears when utilities build substations, transmission or generation for projects that never arrive.
The economic risk is easy to miss. A hyperscaler can use every accelerator it buys and still destroy value if AI prices fall faster than computing costs. The building may look full while the return on capital disappoints.
Hardware and real estate also age at different speeds. GPUs can lose much of their economic value within a few years. A well-located grid connection, fiber route and adaptable building can remain useful for much longer. That is why we judge overbuilding by usable capacity and financial returns, rather than by the total number of announcements.
As this chart shows, and as featured in our data center market deck, search interest in data centers has increased significantly
Are existing AI data centers already sitting empty?
No. The latest market data still shows a shortage of usable AI data center capacity in the largest North American hubs.
CBRE found that inventory across the four largest North American markets grew 33% year over year in the first quarter of 2026. Vacancy still fell to record lows. Northern Virginia reached 0.3%, Atlanta 1%, Dallas-Fort Worth 1.8% and Chicago 2.2%.
Demand did more than keep pace with new construction. Net absorption across those four markets rose 34% to 2,236 megawatts, with Northern Virginia alone absorbing 1,148 megawatts. These are completed or occupied megawatts, so they tell us more than a list of future campuses.
The rental market points in the same direction. In CBRE’s latest review, asking rents still rose across the four main North American markets, led by a 14.7% increase in Chicago. Customers are paying more while accepting long delivery times, which is hard to square with a broad surplus.
Europe and Asia-Pacific are less tight, with aggregate vacancy near 7% in their leading markets. That leaves room for local mistakes, but it still falls well short of a global glut.
| Market | Inventory growth | Current vacancy | What we learn |
|---|---|---|---|
| Northern Virginia | 37.3% | 0.3% | Huge additions were absorbed |
| Atlanta | 14.5% | 1.0% | Available capacity kept shrinking |
| Dallas-Fort Worth | 43.7% | 1.8% | Rapid building has not produced a broad surplus |
| Major European hubs | 18.9% | 7.3% | More balanced, with greater local variation |
Is the AI data center pipeline bigger than proven demand?
Yes. The announced AI data center pipeline is much bigger than demand proven today, although the projects already under construction have unusually strong customer commitments.
JLL estimates that more than 35 gigawatts is under construction in North America and that 92% of the pipeline is precommitted. Nearly 60% has been leased, while most of the rest is being built directly by hyperscalers for their own use.
That 92% figure sharply reduces the chance of dozens of speculative buildings opening without tenants. It does not guarantee attractive returns. Microsoft, Amazon, Meta or Google can reserve too much capacity for themselves, and a signed customer can later delay a deployment or renegotiate terms.
The shakier part sits earlier in the process. Global plans contain land options, proposed campuses and grid applications that lack final financing or equipment orders. JLL expects close to 100 gigawatts of new global capacity between 2026 and 2030, almost doubling the installed base. Some of that will disappear long before construction begins.
The press-release pipeline is clearly bloated, while the construction pipeline looks aggressive but currently has enough contracted demand to avoid an immediate crash.
| Pipeline stage | Evidence of commitment | Current overbuilding risk |
|---|---|---|
| Operating facility | Paying workloads and measured electricity use | Low in major hubs |
| Under construction | 92% precommitted in North America | Moderate, mainly through tenant overestimation |
| Contracted power or signed land | Partial commercial commitment | High if financing or permits remain open |
| Utility request or public announcement | Often little beyond developer interest | Very high |
This chart, featured in our data center market deck, illustrates yearly venture capital funding for data center startups
Are utilities counting the same future data centers more than once?
Yes, utility forecasts are inflating the apparent AI data center boom because developers often shop the same project across several grids.
A company looking for 500 megawatts may approach utilities in Texas, Ohio, Virginia and Louisiana at the same time. Each utility sees a serious potential customer, but only one location may eventually win. Regional forecasts can then count four projects where the developer intends to build one.
Grid Strategies estimates that the data center portion of United States utility load forecasts could be overstated by roughly 25 gigawatts. That is a large correction. It equals dozens of hyperscale campuses and explains part of the gap between utility forecasts and industry construction estimates.
PJM has started discounting requests that lack stronger proof. Its process now looks at signed agreements, financial commitments, timing and whether proposed loads can fit within broader market forecasts. An Ohio data center queue that reached 38 gigawatts also pushed utilities toward non-refundable study fees and longer minimum-payment obligations.
Utilities can spend real money on transmission and generation while developers are still comparing locations. Deposits and take-or-pay tariffs are sensible because they force the company creating the forecast risk to carry more of the cost.
If you want more recent data on this point, please see our latest data center market report.
Are hyperscalers spending faster than AI can repay them?
Possibly. Hyperscaler returns now provide the strongest serious case that the AI data center boom has gone too far.
The latest IEA assessment puts combined capital expenditure by five large technology companies above $400 billion in 2025, with another 75% increase expected in 2026. Even businesses as large as Microsoft, Alphabet, Amazon and Meta need an enormous new profit pool to earn a good return on spending of that scale.
Demand is growing at the same time. Google Cloud revenue rose 63% in its latest reported quarter, passed $20 billion and carried a backlog above $460 billion. Microsoft reported Azure growth around 40%, while Meta still expects operating income to exceed its previous year despite guiding to $125 billion to $145 billion of 2026 capital expenditure.
The awkward part appears in margins and cash flow. Microsoft says AI infrastructure and heavier AI usage are already weighing on cloud gross margin. Meta’s core advertising machine can fund its buildout, but the company cannot cleanly separate the return from each additional data center. Backlog also records future contracted revenue, sometimes over many years, rather than cash available today.
The largest companies can afford to be wrong, but their shareholders may still pay for years of mediocre returns. A bubble can consist of excellent technology built at a price that future profits cannot support.
This chart, featured in our data center market deck, shows how Equinix is capturing share in data centers
Is real AI usage growing fast enough to absorb more data centers?
Yes, real AI usage is currently growing fast enough to keep new capacity busy, though it has not validated every campus planned for the end of the decade.
The IEA found that global data center electricity use rose 17% in 2025, far faster than worldwide power demand, while AI-focused facilities grew faster again. Electricity consumption is useful evidence because it measures machines doing work rather than executives describing future demand.
Cloud results confirm the direction. Google Cloud passed $20 billion in quarterly revenue, Azure continued growing near 40%, and AWS remains a business measured in tens of billions of dollars per quarter. AI products are only part of those totals, but the acceleration has arrived alongside tighter capacity and heavier infrastructure spending.
The mix of work is broadening as well. Companies are paying for model training, coding assistants, search, advertising, image generation, enterprise copilots and private inference. This reduces dependence on one blockbuster consumer chatbot.
Still, usage and profit are different questions. Free AI interactions can multiply without creating much revenue, and providers keep cutting the price of comparable output. Compute demand is real. Whether customers will pay enough to justify every dollar now being committed is much less certain.
Will AI inference fill the next wave of data centers?
Probably, and inference now gives the industry its clearest demand case for building beyond a small number of giant training campuses.
Training a frontier model creates a huge but occasional workload. Serving that model can generate requests every second for years. Search answers, coding suggestions, customer-service agents, advertising decisions, generated media and automated workflows all add recurring inference demand.
JLL expects inference to become the main driver of AI capacity by 2030. The location needs are also different. Training can sit in a remote area with cheap power and strong fiber. Interactive inference works better when capacity is spread across regions close to users and enterprise customers.
That change helps absorb more facilities, yet it can also expose bad siting decisions. A remote gigawatt campus designed around one training customer may struggle to replace that workload with latency-sensitive business applications. Capacity remains local, even when the AI market is global.
Inference should support continued construction, especially near major cloud regions and population centers. It offers much less protection to isolated projects built around a single model developer.
This chart, featured in our data center market deck, illustrates yearly funding for data center startups
Could AI efficiency strand a lot of new capacity?
Efficiency will break some AI data center forecasts, but it is unlikely to shrink total computing demand.
The IEA estimates that electricity used per comparable AI task has fallen by at least one order of magnitude annually in recent years. Better chips, smaller models, quantization, caching and improved software allow the same facility to produce far more output.
Simple text is already cheap in energy terms. The IEA calculates that replacing every conventional internet search with a basic AI text query would consume less than 4 terawatt-hours a year, below 1% of current global data center electricity use. Forecasts built by multiplying an old, expensive query cost by billions of future requests will badly overshoot.
New applications keep raising the ceiling. Reasoning, video generation and agentic workflows can use hundreds or thousands of times more energy than a simple text response. Lower costs also encourage developers to run longer prompts, more model calls and additional agent steps.
The likely path is rising total demand with much better output per megawatt. That still creates casualties. Facilities planned around today’s power consumption per task may open into a world where customers need fewer chips for the same workload.
If you want more recent data on this point, please see our latest data center market report.
Could new AI data centers become obsolete before they open?
Yes, a poorly designed AI data center can age before launch even while its grid connection becomes more valuable.
AI hardware changes much faster than buildings. Microsoft reported that roughly half of one recent quarter’s capital spending went to short-lived assets, mainly GPUs and CPUs. The rest included sites and infrastructure expected to earn revenue for 15 years or longer.
Power density is moving just as quickly. New AI racks need far more electricity, cooling and network capacity than conventional cloud servers. A facility designed several years earlier may require expensive liquid-cooling, electrical and structural work before it can host the latest systems.
The best sites can survive that transition. A powered campus with fiber, flexible cooling and room for denser equipment can cycle through several hardware generations. Older accelerators may move to cheaper inference or scientific workloads while new chips take the premium jobs.
Rigid projects face a harsher outcome. A remote site built around one customer, one rack density and one cooling design has fewer alternatives. Such buildings may remain technically usable while losing much of the value assumed in their financing model.
This chart, featured in our data center market deck, compares the main business model options for hyperscale data center operators
Are power shortages preventing an AI data center glut?
For now, power shortages are the main reason the AI data center market has avoided a physical glut.
Securing land is relatively easy compared with securing several hundred megawatts of reliable electricity. Grid studies, transmission upgrades, substations, transformers and permits can add years. Some projects in established markets are being quoted connection dates in the early 2030s.
The bottleneck has already changed the map. JLL says 64% of North American capacity under construction now sits in frontier markets such as West Texas, Tennessee, Wisconsin and Ohio. Developers are moving toward power rather than waiting for traditional hubs to expand.
The filter is severe. EPRI’s latest scenarios put United States data centers at 9% to 17% of national electricity use by 2030, compared with roughly 4% to 5% in 2024. Grids cannot connect the high end of that range without major generation and transmission investment.
These constraints make the boom more expensive and slower, but they also remove weak projects. A developer can announce a campus in an afternoon. Bringing a gigawatt of dependable power online forces the project to survive years of technical, political and financial scrutiny.
Do AI data center cancellations prove that demand is weakening?
Not yet. Recent cancellations reveal more about siting problems, public opposition and portfolio cleanup than collapsing AI demand.
Microsoft’s reported decision to cancel or defer up to 2 gigawatts of leased capacity remains the clearest demand-related warning. Analysts argued that the company had reserved more capacity than its revised OpenAI workload forecast required in some locations. Microsoft kept spending heavily elsewhere, which makes the episode look like a real correction inside one portfolio rather than an industry retreat.
The newer cancellation wave has a different cause. Data Center Watch counted at least 75 United States projects worth roughly $130 billion that were blocked or delayed by local opposition in the first quarter of 2026. Concerns about electricity prices, water, noise, land use and pollution are now stopping projects before demand can be tested.
Those delays tighten available supply even when they damage developer economics. They also show why announced megawatts exaggerate what will open. Community approval has become another hard filter, alongside power and financing.
We would take cancellations as evidence of a broad glut once several hyperscalers blame weak usage, preleasing falls and rents decline at the same time. That combination has not appeared so far.
This chart, featured in our data center market deck, shows the revenue mix across customer segments in the data center market
Are neoclouds the weakest part of the AI data center boom?
Yes, leveraged AI cloud providers are the part of the current buildout most likely to break first.
CoreWeave shows why. Its first-quarter 2026 revenue more than doubled to $2.08 billion, backlog reached $99.4 billion and active power passed 1 gigawatt. It also reported more than 3.5 gigawatts of contracted power and aims to exceed 8 gigawatts by 2030.
The same quarter produced a $740 million net loss and $536 million of net interest expense. Interest alone equaled about one-quarter of revenue. CoreWeave has secured large customer commitments and substantial financing, but it must keep building before much of that backlog can become revenue.
This model works brilliantly while demand, credit and equipment supply remain available. A delay can hit several parts of the business at once: the facility opens late, revenue recognition moves back, interest keeps accruing and the customer may have found another option.
Hyperscalers can redirect a campus across cloud, advertising, search and internal workloads. A specialist has fewer escape routes. That makes neocloud lenders, equipment financiers and landlords more exposed than the biggest technology companies.
| CoreWeave measure | Latest reported figure | What it tells us |
|---|---|---|
| Quarterly revenue | $2.08 billion | Demand is already substantial |
| Revenue backlog | $99.4 billion | Future commitments are enormous |
| Active power | More than 1 GW | The business has reached real scale |
| Contracted power | More than 3.5 GW | Expansion still requires major execution |
| Net interest expense | $536 million | Financing risk is already visible |
| Net loss | $740 million | Fast growth has not removed financial strain |
If you want more recent data on this point, please see our latest data center market report.
Are some regions already building too many data centers?
Yes, local data center gluts are already appearing even while the biggest North American hubs remain almost full.
As seen above, Northern Virginia, Atlanta, Dallas and Chicago still have very low vacancy. Several smaller or faster-growing markets look different. Querétaro’s inventory jumped 450% in one year and vacancy rose to 10.6%. Hong Kong remained near 18%, while Bogotá approached 19%.
The contrast shows why global megawatts can mislead. A vacant facility in Bogotá cannot immediately serve a customer that needs 100 megawatts beside an established cloud region in Virginia. An older air-cooled building also cannot always host dense liquid-cooled AI systems.
Tax incentives and available land can attract several developers before a local customer base develops. Once construction begins, each project may assume that the same limited group of hyperscalers will become tenants.
Regional overbuilding will probably remain the most common form of excess. It can hurt individual developers badly without producing a worldwide collapse in data center demand.
This chart, featured in our data center market deck, shows how hyperscale AI-ready campus technology has evolved over time
Is this becoming the dot-com fiber bubble again?
The fiber bubble is a useful warning, but today’s operating data does not yet show the same outcome.
Both booms involve long construction cycles, aggressive demand forecasts and competitors building at the same time because arriving late looks dangerous. Much of the fiber installed around 2000 was eventually used, yet investors who paid the wrong price or carried too much debt still lost heavily.
AI infrastructure could follow that pattern. Powered campuses may remain useful for years, while the original owners write down GPUs, renegotiate leases or sell projects to stronger buyers. Useful infrastructure can still be a terrible investment at the wrong valuation.
The current difference is utilization. Major-market vacancy is near historic lows, rents remain firm and most construction has an intended user. The fiber comparison becomes much stronger if vacancy rises, lease pricing falls and heavily financed providers begin restructuring together.
Losses would spread unevenly. Hyperscaler shareholders would absorb weaker returns and depreciation. Neocloud lenders could face defaults. Landlords might accept lower rents, while utility customers could be asked to pay for grid upgrades built around projects that never arrived.
The historical lesson is straightforward: the industry can need far more computing infrastructure and still waste a great deal of capital while building it.
Are we already building too many AI data centers?
No, we are not yet building too many usable AI data centers overall, but we are already planning too many speculative projects and financing some capacity on assumptions that will fail.
Today, the case against a broad physical glut is strong. Major North American vacancy remains around 1%, completed capacity is being absorbed quickly, rents have risen and 92% of the construction pipeline is precommitted. Electricity use and cloud revenue also show that AI demand has moved well beyond demonstrations.
The overbuilding is showing up earlier in the chain. Utility forecasts may overstate data center demand by roughly 25 gigawatts. Local markets such as Querétaro, Hong Kong and Bogotá already carry double-digit vacancy. Microsoft has trimmed capacity in some locations. CoreWeave’s growth comes with enormous interest expense, and community opposition recently blocked or delayed dozens of United States projects.
Power shortages are currently saving the industry from its own enthusiasm. They slow the strongest projects and kill many of the weakest before concrete is poured. The danger rises when grid equipment, generation and permitting catch up, allowing several years of delayed capacity to open together.
The practical answer depends on the project. The world still needs more well-located, adaptable and contracted AI data centers, while speculative campuses built around duplicate power requests, one customer or optimistic computing forecasts already look excessive.
We would change the answer once major-market vacancy rises into the high single digits, rents fall, preleasing weakens and several hyperscalers cut capacity because customers are using less AI. Until then, the likely correction remains selective. Leveraged neoclouds, poorly located projects and careless utility forecasts are in danger first, while genuinely powered capacity in the strongest markets remains scarce.
If you want more recent data on this point, please see our latest data center market report.
In our data center market deck, we identify pain points entrepreneurs should prioritize
OUR METHODOLOGY
We broke the question into the parts that show different kinds of excess: completed capacity, vacancy and rents, construction commitments, utility load forecasts, hyperscaler spending, AI usage, financing risk, regional conditions and technological change.
We separated operating and occupied capacity from projects under construction, then separated both from land options, power applications and public announcements. That distinction lets us compare capacity that can actually serve customers with capacity that may never be built.
We used vacancy, absorption, rents and preleasing as the clearest tests of a physical glut. We looked at margins, cash flow, backlog, interest expense and asset life to judge whether busy facilities can still produce weak financial returns. Utility queues, deposits and minimum-payment commitments were used to identify planning demand that may be counted more than once.
We prioritized recent first-hand and institutional evidence over broad market commentary. The main sources were the International Energy Agency, JLL’s Global Data Center Outlook, CBRE’s North America Data Center Trends, Grid Strategies, PJM Interconnection and EPRI.
Company economics came from the investor-relations and filing materials of Microsoft, Alphabet, Amazon, Meta and CoreWeave. We also used Data Center Watch for project delays and opposition, and SEC filings where reported company figures needed a primary-source check.
This chart, featured in our data center market deck, shows the revenue mix by region across Europe, Asia, North America, Africa, and South America in the data center market
Related blog posts
- Are data centers the new bubble?
- Is the AI infrastructure boom a bubble?
- Which startup is setting the pace in AI data centers?
