Are data centers the new bubble?
In our data center market deck, you will find everything you need to understand the market
SUMMARY
Data centers are partly the new bubble: demand and scarcity are real, but speculative development, leveraged GPU fleets and increasingly complex financing have already moved beyond what current returns comfortably support.
The broad market is not overbuilt today. Operational supply across major markets grew by 25% in one year while global vacancy still fell, which is the opposite of a conventional demand collapse.
“Data centers” now describes several very different investments. A powered facility with a strong tenant and a long lease is a different asset from an unpowered site, and both are different again from a debt-financed fleet of fast-depreciating GPUs.
The most bubble-like risk sits inside the capital structure, not necessarily inside the buildings. A useful facility can survive for decades even after its original equity investors and lenders lose money.
Big Tech can fund the buildout, but affordability should not be confused with attractive returns. Capital spending is consuming cash rapidly, while depreciation, energy costs and lower cloud margins are starting to show up in reported results.
Huge backlogs prove that customers want the capacity. They do not prove that every layer of the chain—from the AI developer to the cloud provider, neocloud, landlord and power supplier—will earn a healthy margin from the same underlying demand.
Power constraints cut both ways. They slow construction enough to protect existing rents, but they also encourage speculation in land, interconnection positions and frontier markets where the promised capacity may arrive late or not at all.
Cheaper AI is unlikely to eliminate computing demand. It can increase total usage while destroying the economics of older hardware, fixed cooling designs and operators that financed GPUs on the assumption that today’s rental prices would last.
Local gluts can appear long before a global glut. Bogotá, Hong Kong and Querétaro already show that vacant capacity in one market cannot solve scarcity in another because location, latency, power, regulation and technical design are not interchangeable.
The likely outcome is a financial shakeout rather than rows of permanently empty buildings. Projects will be cancelled, refinanced, delayed or sold at lower valuations, while much of the underlying infrastructure keeps operating.
The telecom boom is the useful warning. Infrastructure can become essential, traffic can keep rising and technology can work brilliantly—while early owners still lose because they paid too much and borrowed too heavily.
This market map, featured in our data center market deck, highlights top companies and startups in the data center market
Are data centers the new bubble?
Data centers are in a real boom, and part of that boom has already crossed into bubble territory.
The shortage itself is genuine. In CBRE’s latest survey of 16 major markets, operational supply grew by 25% in one year, yet global vacancy still fell from 8.3% to 6.7%. Northern Virginia reached 0.3%, Atlanta 1% and Singapore 2%. Customers currently want more usable capacity than operators can deliver.
The financial picture looks less comfortable. Big technology companies are spending hundreds of billions of dollars, developers are chasing power in unfamiliar locations, and a growing share of the expansion is financed through debt, private credit and special-purpose vehicles. Meanwhile, the GPUs inside these facilities can lose economic value far faster than the buildings around them.
Several different markets hide under the label “data centers.” A powered building with a strong tenant and a long lease has little in common with an unpowered development site. A hyperscaler using servers across advertising, cloud and consumer products faces a different risk from a leveraged GPU provider serving a handful of AI laboratories.
The useful question is where real demand ends and capital mispricing begins.
What would a data center bubble actually mean?
A data center bubble would mean that today’s prices and financing assume more durable demand than the projects can eventually earn.
We do not need empty buildings across the world to make that judgment. A bubble can develop in land bought before power is secured, campuses valued before permits arrive, GPUs financed at temporary rental prices, or company valuations that assume every announced megawatt will generate attractive returns.
The distinction between a useful asset and a good investment is crucial here. A data center may operate for decades after its original owner loses money. Lenders can restructure the debt, new investors can buy the facility cheaply, and customers can continue using it. The infrastructure succeeds while the first capital invested in it fails.
We are judging four things: whether the capacity gets built, whether customers use it, whether they pay enough, and whether that revenue arrives before financing costs and depreciation consume the return.
If you want more recent data on this point, please see our latest data center market report.
As this chart shows, and as featured in our data center market deck, search interest in data centers has increased significantly
Why did data center spending explode so quickly?
Data center spending exploded because AI demand arrived while chips, electricity and construction capacity were already tight.
Amazon currently expects about $200 billion of capital expenditure for the year. Alphabet raised its guidance to between $180 billion and $190 billion, while Meta increased its range to between $125 billion and $145 billion. Those three companies alone now plan between $505 billion and $535 billion of annual investment. Their budgets cover more than data centers, but AI servers, networks, energy systems and related infrastructure account for a large share.
Microsoft has avoided giving the same full-year figure, but its cash-flow statement shows $80.1 billion of additions to property and equipment during the first nine months of its financial year, up from $47.5 billion a year earlier. That is a 69% increase.
The price of each megawatt is also rising. JLL estimates that average construction costs increased from $7.7 million per megawatt in 2020 to $10.7 million in 2025 and could reach $11.3 million this year. Installing advanced computing equipment can add far more than the basic building cost.
Developers are ordering more capacity at a higher cost per unit, while paying premiums for land, transformers and rapid grid access. An already large expansion became historically large.
| Company | Latest investment disclosure | What changed |
|---|---|---|
| Amazon | About $200B for the year | AI investment is consuming most recent free cash flow |
| Alphabet | $180B–$190B | Guidance was raised again after its latest quarter |
| Meta | $125B–$145B | Higher component prices and future data center capacity pushed the range upward |
| Microsoft | $80.1B during nine months | Property and equipment additions rose 69% year over year |
Is there really enough data center demand today?
Yes, current data center demand is strong enough that the leading markets remain almost full after a huge increase in supply.
North American inventory across Northern Virginia, Atlanta, Dallas and Chicago grew by 33% in CBRE’s latest annual comparison. Vacancy still fell in all four markets. Northern Virginia added more than 1.1 gigawatts of capacity and absorbed slightly more than it added, pushing vacancy down to 0.3%.
JLL provides another useful check. More than 35 gigawatts of data center capacity is under construction in North America, and 92% has already been committed through leases or owner-occupied hyperscaler projects. Customers may later reduce, renegotiate or delay those commitments, but developers are clearly building for identified users rather than hoping somebody eventually appears.
Prices confirm the scarcity. Monthly asking rents currently reach $190 to $235 per kilowatt in Northern Virginia, $200 to $230 in Chicago and an average of $403 in Singapore. A broad glut would normally push those prices downward.
| Market | Latest vacancy | What is happening |
|---|---|---|
| Northern Virginia | 0.3% | Record absorption exceeded the large amount of new supply |
| Atlanta | 1.0% | Vacancy fell sharply despite continued construction |
| Dallas–Fort Worth | 1.8% | 88% of capacity under construction is already leased |
| Chicago | 2.2% | Rents rose while available capacity remained scarce |
| Singapore | 2.0% | Restricted supply keeps prices exceptionally high |
This chart, featured in our data center market deck, illustrates yearly venture capital funding for data center startups
Is AI usage growing quickly enough to fill future capacity?
AI usage is currently growing fast enough to absorb new capacity, although forecasts beyond the next few years remain far less certain.
AWS revenue reached $37.6 billion in its latest quarter and grew by 28%, its fastest rate in 15 quarters. Microsoft says its AI business has passed a $37 billion annual revenue run rate and is growing by 123%. Oracle’s cloud infrastructure revenue rose by 93% in its latest quarter and by 77% over the full financial year. These are large businesses growing at startup-like rates.
Operational evidence tells the same story. Microsoft added another gigawatt of capacity during its latest quarter and still reported that customer demand exceeded what it could supply. Alphabet said its internal and external need for AI computing was unprecedented, while Google Cloud revenue grew by 63% and its backlog nearly doubled sequentially.
The uncertainty begins when we extend today’s growth rates over five or ten years. Early enterprise adoption can produce spectacular percentage gains from a small base. Future models may use more computing, fewer chips per task, or a different mix of training and inference.
Today’s capacity shortage is clear. Whether every campus proposed for the end of the decade will be needed remains unproven.
Can Big Tech actually afford this buildout?
Big Tech can afford the data center buildout today, but the cash cushion is shrinking much faster than many investors assume.
Amazon generated $148.5 billion of operating cash flow over its latest 12 months. Free cash flow fell to only $1.2 billion after property and equipment purchases increased by $59.3 billion, mainly because of AI investment. AWS is highly profitable, yet infrastructure spending has absorbed almost everything generated after other cash needs.
Alphabet spent $35.7 billion in its latest quarter, with roughly 60% going to servers and 40% to data centers and networking equipment. It still produced $10.1 billion of quarterly free cash flow. Meta spent $19.8 billion and generated $12.4 billion of free cash flow. Microsoft produced $127.5 billion of operating cash during nine months while adding $80.1 billion of property and equipment.
These companies can keep investing through a weaker period. They own profitable advertising, software, commerce and cloud businesses, and they can reduce buybacks or issue debt at relatively favorable rates.
Affordability is a low bar, though. Spending can remain financially survivable and still earn disappointing returns. Alphabet has already warned that higher depreciation and energy costs will pressure future profits. Microsoft’s cloud margin has also declined as AI infrastructure costs rise.
If you want more recent data on this point, please see our latest data center market report.
This chart, featured in our data center market deck, shows how Equinix is capturing share in data centers
Are data center revenues catching up with spending?
Data center revenues are catching up inside mature property businesses, while the broader AI infrastructure buildout still consumes cash at a punishing rate.
Equinix’s latest quarterly revenue grew by 10%, monthly recurring revenue rose by 12%, and its stabilized developments continued producing cash-on-cash returns around 26%. Digital Realty signed leases representing $707 million of annualized rent at full ownership share and ended the quarter with $1.8 billion of signed annual rent waiting to commence. Existing, well-located facilities can still be excellent businesses.
The lag between investment and revenue is becoming uncomfortable elsewhere. Digital Realty’s newly signed leases take an average of 19 months to start. During that gap, the company must finish construction and carry the financing.
Oracle shows the pressure at hyperscale. Cloud infrastructure revenue grew by 77% over its latest financial year, yet trailing capital expenditure reached $55.7 billion against $32 billion of operating cash flow. Its reported free cash flow was negative $23.7 billion.
There are two very different economics here. Scarce, operational facilities are producing strong rents. The next wave requires so much upfront capital that even fast revenue growth may take years to catch up.
Do giant AI backlogs make the boom safe?
Giant AI backlogs prove that customers are serious, but they leave timing, profitability and concentration wide open.
Oracle recently reported $638 billion of remaining performance obligations. Google Cloud ended its latest quarter with a $462 billion backlog, and Microsoft’s commercial remaining performance obligations reached $627 billion. CoreWeave separately reported $99.4 billion of revenue backlog.
Those figures measure different things. They include various combinations of cloud services, software, hardware and future capacity, recognized over different periods. Adding them together would produce a meaningless total.
The underlying demand can also appear at several points in the same chain. An AI developer signs a cloud agreement. The cloud provider reserves capacity from a neocloud. The neocloud leases a data center, and the data center owner supports a new power project. Every contract is real, although several ultimately depend on the same group of AI users generating enough revenue.
A backlog can disappoint without disappearing. Construction may run late, equipment may cost more, and the customer may consume the service at a lower margin than expected. CoreWeave’s $99.4 billion backlog is impressive, but one newly signed Meta commitment alone represented $21 billion. Large contracts bring visibility and serious concentration at the same time.
Backlogs kill the weakest bubble argument—that nobody wants the capacity. They offer much less comfort about the returns.
If you want more recent data on this point, please see our latest data center market report.
This chart, featured in our data center market deck, illustrates yearly funding for data center startups
Is there already too much data center capacity anywhere?
Yes, local data center oversupply is already visible in several markets, even while the largest hubs remain painfully tight.
Querétaro’s inventory expanded by 450% in one year, and vacancy jumped from 0.9% to 10.6%. Hong Kong has 18% vacancy, while Bogotá stands at 18.7%. Those figures sit far above the levels seen in Northern Virginia, Atlanta or Singapore.
This unevenness is easy to overlook when the industry talks in global gigawatts. Capacity in Bogotá cannot solve a shortage in Virginia. Customers care about latency, regulation, cloud availability zones, network connections, power reliability and where their existing systems operate.
Technical differences create another divide. A conventional facility may have space available but lack the rack density, cooling systems or electrical design needed for modern AI clusters. The industry can have vacant buildings and a shortage of usable AI capacity at the same time.
Local oversupply will probably become more common because 64% of North American construction has shifted into frontier markets. Some of those places will become major hubs. Others are being built around demand forecasts that have never been tested at that scale.
| Market | Latest vacancy | Our reading |
|---|---|---|
| Northern Virginia | 0.3% | Severe operating shortage |
| Singapore | 2.0% | Scarcity reinforced by regulatory limits |
| Querétaro | 10.6% | New supply arrived faster than local absorption |
| Hong Kong | 18.0% | Substantial capacity remains available |
| Bogotá | 18.7% | Demand has yet to justify the existing footprint |
Can all the announced data centers actually be built?
No, a meaningful share of the announced data center pipeline will be delayed, resized or abandoned.
Electricity has become the main obstacle. Developers can acquire land and publish campus plans long before they secure generation, transmission capacity, transformers and a firm grid connection. In established markets, the wait for usable power can be longer than the construction itself. CBRE now describes power availability and grid infrastructure as the main factors delaying projects in several major hubs.
The shift into frontier markets shows how far developers will go. JLL reports that North America has more than 35 gigawatts under construction, with 64% located in places such as West Texas, Tennessee, Wisconsin and Ohio rather than the traditional core hubs.
These bottlenecks provide some protection against a sudden global glut. Supply cannot appear as quickly as a spreadsheet suggests, giving demand more time to grow.
They also create a separate pocket of speculation. Developers spend money on land, engineering, deposits and interconnection studies before knowing exactly when power will arrive or what it will cost. A proposed gigawatt campus can lose much of its value after a delayed grid connection, even while completed facilities nearby keep raising rents.
Announced capacity should be treated as a probability-weighted pipeline, not future supply that already exists.
This chart, featured in our data center market deck, compares the main business model options for hyperscale data center operators
Will cheaper AI reduce the need for data centers?
Cheaper AI will probably increase total computing demand while crushing the value of inefficient hardware and poorly designed facilities.
The cost of running a fixed AI task keeps falling as chips improve, models become smaller and software uses hardware more efficiently. Microsoft, for example, reported a 40% improvement in inference throughput for its most-used models after hardware and software optimization.
That improvement allows the same server fleet to answer more requests. A provider that assumed each unit of computing would retain today’s price could see revenue per GPU fall quickly.
Lower prices also unlock uses that were previously uneconomic. Companies can add AI to customer support, search, advertising, coding, video and autonomous workflows. Consumers use more of a service when each interaction becomes faster and cheaper.
The International Energy Agency still expects data center electricity consumption to rise sharply, even after accounting for efficiency improvements. Its central outlook has overall data center use roughly doubling by 2030, with AI-focused facilities growing faster.
Total demand can keep rising while older equipment loses value. Flexible buildings with abundant power can host several generations of chips. Facilities built around one cooling design, one customer or one hardware generation face a harsher future.
Will GPUs age faster than the debt financing them?
Yes, GPU economics are changing faster than the financing structures built around them.
CoreWeave offers the clearest public example. The company generated $2.08 billion of revenue in its latest quarter, more than double the previous year. It also recorded $1.15 billion of depreciation and amortization and $536 million of net interest expense. Together, those two costs equaled roughly 81% of revenue.
CoreWeave spent another $7.7 billion on property and equipment during the quarter. Its non-current debt reached $17.3 billion, alongside substantial lease liabilities. Revenue is soaring, yet the company still reported a $740 million net loss.
These figures do not show a failing business. They show how little room remains for mistakes. The company needs high utilization, reliable customer payments, timely construction and competitive rental rates across several hardware generations.
A GPU can continue functioning long after a newer chip makes it uneconomic for premium workloads. The provider then has to lower the rental price, move the equipment to cheaper tasks or accept a write-down. Debt repayments continue on the original schedule.
This is the most bubble-like part of the data center industry today. Buildings age slowly. Computing margins can collapse within a product cycle.
If you want more recent data on this point, please see our latest data center market report.
This chart, featured in our data center market deck, shows the revenue mix across customer segments in the data center market
Who gets hurt first if AI demand slows?
Leveraged neoclouds, speculative developers and private-credit vehicles would take the first serious losses.
Amazon, Alphabet, Microsoft and Meta can move computing capacity between cloud customers, advertising systems, consumer products and internal model development. They can tolerate temporary underuse and keep financing projects from other business lines.
A specialized operator has fewer exits. Its revenue may depend on several large AI customers, while its debt and leases continue regardless of how much computing those customers consume. A developer waiting for power faces a similar problem because an unfinished site produces no rent.
The financing chain has also become harder to see. The Bank for International Settlements found that hyperscalers increasingly use joint ventures and special-purpose entities that borrow from private-credit funds and institutional investors. The technology company may hold a minority stake, sign a long lease and provide guarantees while most of the debt remains outside its main balance sheet.
The BIS calls these arrangements economically similar to borrowing. Gross bond issuance by major hyperscalers exceeded $100 billion in 2025, and private-credit exposure is growing alongside it.
When demand slows, the building may keep operating under new ownership. The first equity investors, junior lenders and highly leveraged operators would absorb the damage. Big Tech would mainly suffer lower returns and higher depreciation.
Does the telecom bubble tell us what happens next?
The telecom bubble offers the right warning: infrastructure can become essential while early investors still lose badly.
During the late-1990s telecom boom, internet usage was genuinely growing and fiber-optic technology was genuinely transformative. Investment and valuations still ran far ahead of the revenue that networks could produce.
The Federal Reserve Bank of Richmond later found that forecasts became highly inaccurate as regulatory change, rapid technological progress and easy financing arrived together. By 2001, long-distance fiber had been massively overbuilt, prices collapsed and heavily indebted operators entered bankruptcy.
The similarity with AI lies in improving efficiency. Fiber technology allowed each cable to carry far more data than developers had expected. Today, better chips and software let each data center produce more AI output.
There are clear differences. Current vacancy in the largest data center hubs is extremely low, and much of the construction pipeline has already been committed. The largest buyers are also among the world’s most profitable companies.
The telecom lesson still holds. Fast-growing traffic does not guarantee fast-growing profits. When supply expands, technology improves and prices fall together, the infrastructure can flourish while owners struggle to repay what they borrowed to build it.
This chart, featured in our data center market deck, shows how hyperscale AI-ready campus technology has evolved over time
What would show that the data center bubble is bursting?
The data center bubble would be visibly bursting once weakness spreads from speculative projects into operating capacity, customer contracts and credit markets.
One delayed campus tells us little. A stronger warning would involve vacancy rising across several important hubs while rental prices fall. We would also expect customers to postpone signed capacity, hyperscalers to cut spending together and lenders to demand much more protection against GPU and development risk.
The first failures may look less dramatic. A neocloud could struggle to refinance older GPUs. A frontier market could fill with vacant capacity. A developer could surrender a site after years of failed power negotiations. Private-credit investors could extend loan maturities rather than recognize an immediate loss.
Those events can occur while global data center revenue keeps growing. The stress will probably surface in the weakest capital structures before it appears in aggregate demand.
Today, the broad operating indicators still look healthy. Vacancy is low, cloud revenue is accelerating, leases are being signed and the largest buyers continue raising their investment plans. Credit exposure, depreciation and local oversupply are moving in a less reassuring direction.
The decisive warning would be a sustained fall in capacity prices alongside rising utilization problems. At that point, the industry would have more usable computing infrastructure than customers were willing to fund at profitable rates.
Are data centers the new bubble?
Partly yes: operational data centers are still supported by real scarcity, while speculative development, leveraged GPU fleets and aggressive financing have entered bubble territory.
Calling the entire industry a bubble goes too far. Global supply rose by one-quarter in the latest CBRE survey, and vacancy still declined. The strongest hubs are nearly full, cloud infrastructure revenue is growing rapidly, and most North American capacity under construction has already been committed.
Yet today’s shortage has encouraged investors to assume that scarcity, premium rents and high GPU prices will last for many years. That assumption is weak. Construction is moving into less-tested markets, hardware is depreciating rapidly, and spending has grown beyond the cash generated by several major participants.
The safest assets are powered, operational facilities in constrained locations with financially strong tenants and designs that can support future hardware. The danger rises sharply around unpowered land, single-customer campuses, expensive frontier developments and GPU operators carrying heavy debt.
We expect a financial shakeout rather than rows of permanently empty buildings. Some projects will be cancelled. Others will open late, refinance their debt or change owners at lower valuations. Useful capacity will survive after investors accept losses.
Real demand has never guaranteed sensible prices. The data center boom is proving that again.
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 treated the bubble question as an investment question rather than a prediction of whether data centers will remain useful. The analysis separates operating demand from project economics, financing risk and the prices investors are paying for future capacity.
We looked at current supply, vacancy, rents, construction and committed capacity to distinguish a genuine operating shortage from speculative pipeline announcements. CBRE’s global data center research and JLL’s data center outlook were the main sources for market-level comparisons.
For capital spending, cash generation, cloud growth and backlogs, we used recent company results, investor materials and regulatory filings from Amazon, Alphabet, Microsoft, Meta, Oracle, Equinix, Digital Realty and CoreWeave. We compared these figures within the job each metric actually performs rather than adding unlike backlogs or treating every capital-expenditure dollar as a data center investment.
We used operating facilities, signed leases and reported cash returns as the clearest evidence that existing assets can earn attractive returns. We treated announced campuses and unpowered sites as a probability-weighted pipeline because land, permits and public plans do not guarantee a usable grid connection.
Hardware risk was assessed through depreciation, interest expense, capital expenditure, debt and utilization requirements, with CoreWeave used as the clearest public example of a leveraged GPU operator. Energy demand and efficiency assumptions were checked against the International Energy Agency’s Energy and AI work and company disclosures on computing efficiency.
For financing structures, we relied on the Bank for International Settlements’ analysis of hyperscaler joint ventures, special-purpose entities, bond issuance and private credit. The telecom comparison draws on Federal Reserve Bank of Richmond research because it shows how essential infrastructure can still produce poor investment returns when technology improves faster than pricing assumptions.
We prioritized recent operating and financial evidence over distant forecasts. Key sources include CBRE’s Global Data Center Trends, JLL’s Global Data Center Outlook, Amazon Investor Relations, Alphabet Investor Relations, Microsoft Investor Relations, Meta Investor Relations, Oracle Investor Relations, Equinix Investor Relations, Digital Realty Investor Relations, CoreWeave Investor Relations, the International Energy Agency, the Bank for International Settlements, the Federal Reserve Bank of Richmond, and SEC EDGAR filings.
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 we already building too many AI data centers?
- Is the AI infrastructure boom a bubble?
- Which startup is setting the pace in AI data centers?
