Is the AI infrastructure boom a bubble?

Why is everyone suddenly worried about an AI infrastructure bubble?

The AI infrastructure boom is starting to look speculative because technology companies are spending hundreds of billions of dollars on data centers and computing equipment without knowing how much profit those investments will eventually produce.

The scale explains why the question has become so important. According to the International Energy Agency's 2026 report on energy and AI, capital expenditure by five major technology companies exceeded $400 billion in 2025 and was expected to increase by approximately 75% in 2026. That puts their combined annual spending at roughly $700 billion, including infrastructure investments beyond AI. When spending jumps about 75% in a year, a lot has changed; we cover the past year in our report on AI infrastructure.

The IEA also found that the computing capacity of advanced AI data centers had more than tripled in just 18 months. We're seeing an industrial expansion that would have been difficult to imagine a few years ago.

The concern is that the companies selling AI infrastructure are already collecting revenue, while the businesses ultimately expected to pay for all this computing power are still figuring out how much money AI can make.

A market can grow enormously and still attract too much investment. The telecom industry demonstrated that during the dot-com era, when demand for internet connectivity exploded but many companies financing fiber-optic networks lost their money.

How much are Microsoft, Amazon, Google and Meta spending on AI infrastructure?

The four largest American technology companies are currently committing roughly three-quarters of a trillion dollars annually to capital expenditure, much of it driven by AI infrastructure.

Amazon has indicated approximately $220 billion in annual capital expenditure, up from its earlier $200 billion plan. Alphabet raised its guidance to $195 billion–$205 billion, while Meta expects $130 billion–$145 billion. Microsoft's calendar-year spending has been estimated at approximately $190 billion, although its reporting periods and treatment of leases differ from those of its competitors.

To understand the acceleration, consider Alphabet. The company spent approximately $31 billion on capital expenditure in 2022. Its latest annual guidance is more than six times higher.

Amazon's first-half cash capital expenditure increased from $55.6 billion in 2025 to $96.3 billion in 2026, an increase of approximately 73%. Companies that once spent heavily to build cloud computing businesses are now investing at an entirely different scale.

These amounts include conventional cloud infrastructure, logistics and other capital investments, depending on the company. We cannot attribute every dollar to AI, and the different accounting definitions make the totals approximate.

Still, the direction is unmistakable. The world's biggest technology companies are undertaking one of the most expensive infrastructure expansions in corporate history.

Company	Annual capital spending indication	Main exposure
Amazon	Approximately $220B	AWS, AI infrastructure, logistics
Alphabet	$195B–$205B	AI computing, servers, data centers
Microsoft	Approximately $190B	Azure, AI capacity, enterprise cloud
Meta	$130B–$145B	AI training, inference, data centers
Combined	Approximately $735B–$760B	Mixed capex definitions
Alphabet in 2022	Approximately $31B	Historical baseline
Amazon, H1 2025	$55.6B	Cash capital expenditure
Amazon, H1 2026	$96.3B	Up approximately 73%

Are AI data centers actually filling up, or are companies building empty facilities?

AI data centers are filling up faster than operators can build them, which makes the current infrastructure boom considerably more credible than the speculative construction cycles of the past.

CBRE's August 2026 North American data center report found that capacity across the eight largest markets increased 33.7% year over year to 10.9 gigawatts. Even after that expansion, vacancy fell from 1.6% to 1.4%.

Demand also increased. Net absorption reached 1,456 megawatts, nearly 12% higher than the previous year. Meanwhile, asking rents rose across every major size category, with the sharpest increase reaching 8.3%.

Northern Virginia offers an even more extreme example. The region had approximately 4,497 megawatts of operational data center capacity and just 0.2% vacancy. Atlanta, meanwhile, had 2,882 megawatts under construction, reflecting how developers are expanding into markets where they can secure electricity and land.

Those numbers describe a market with very little spare capacity. Developers are finding customers, tenants are paying higher rents, and new facilities are being absorbed.

There's little evidence that the present AI infrastructure boom consists mainly of unwanted buildings. Whether the next wave of facilities will be equally successful is another question entirely.

Is cloud revenue growing fast enough to support the AI boom?

Microsoft, Amazon and Google are reporting exceptional cloud growth, but their AI infrastructure spending is increasing so quickly that we still cannot assume the investment returns will be attractive.

Amazon Web Services generated $42.2 billion in revenue in the second quarter of 2026, up 37% year over year. That was AWS's fastest percentage growth in 18 quarters. Its operating income also increased 64% to $16.6 billion. Whether cloud revenue can pay for all this comes down to business models, so we compare which ones win in our AI infrastructure market report.

Google Cloud grew even faster. Alphabet reported $24.8 billion in quarterly cloud revenue, up 82%, with operating profit reaching approximately $8.8 billion.

Microsoft's Azure and other cloud services revenue increased 43% in its fiscal fourth quarter. Its broader Intelligent Cloud division generated $39.3 billion, up 32%.

These results show why the largest technology companies are investing so aggressively. All three are selling significantly more cloud services, and the acceleration has been especially strong at Google and Amazon.

But their reported cloud revenue includes conventional databases, storage, networking, enterprise applications and traditional computing workloads. We cannot extract a clean AI infrastructure return from those figures.

The scale of investment also complicates the comparison. Alphabet's first-half capital expenditure roughly doubled while Google Cloud revenue increased 82%. Amazon's capital spending rose 73% in the first half, compared with 33% AWS revenue growth over the same period.

Spending is currently growing faster than established cloud revenue at both companies. That may be justified if the infrastructure generates years of future income, but it leaves little room for disappointing demand.

Are AI companies making enough money to pay for all this computing power?

AI companies are generating substantial revenue, but the industry has yet to demonstrate that profitable end-user spending can support the full cost of the infrastructure being constructed.

We can see the distinction in Microsoft's results. The company reported $214.4 billion in fiscal-year cloud revenue, with nearly 90% coming from customers outside frontier AI model developers. It also disclosed more than 30 million paid Microsoft 365 Copilot seats.

Those figures demonstrate that AI-related services are reaching ordinary businesses, rather than depending entirely on a handful of well-funded AI laboratories.

However, Microsoft Cloud revenue includes large established software businesses, and a paid Copilot subscription does not tell us the profitability of delivering that service. The company reported a 65% cloud gross margin in its fiscal fourth quarter, down year over year partly because of AI infrastructure investment and growing usage.

A similar problem appears across the industry. We know how much cloud providers charge customers, but considerably less about how much of that spending comes from profitable AI applications rather than companies consuming venture capital or strategic funding.

For example, an AI developer can spend billions on GPUs and cloud services while still losing money. The infrastructure provider records legitimate revenue, even though the developer may need another funding round to keep purchasing those services.

That's the weak spot in the boom. Eventually, enough consumers and businesses must be willing to pay for AI-generated work to cover the full cost of delivering it. Impressive infrastructure sales alone don't prove we've reached that point.

Is Nvidia's growth proof that the AI infrastructure boom is healthy?

Nvidia's latest financial results show that AI computing demand is extraordinarily strong, although the company's profits tell us much more about the economics of selling GPUs than about the returns earned by GPU buyers.

In August 2026, Nvidia reported $96.2 billion in quarterly revenue, up 106% year over year. Its Data Center division generated $89 billion, an increase of 117%.

The company's gross margin reached 75%, while quarterly operating income climbed to $63.7 billion, more than double the previous year's figure.

Few industrial businesses have ever expanded at that speed while maintaining such high margins. Nvidia has become one of the clearest financial beneficiaries of the AI boom because virtually every major cloud provider needs advanced accelerators for training and inference.

The difference between Nvidia and its customers is important. Nvidia receives payment when computing equipment is sold. Microsoft, CoreWeave, Oracle and other buyers then have to recover the cost by operating that equipment over several years.

Nvidia can therefore report extraordinary earnings even if some infrastructure operators eventually discover that their investments were too expensive.

Its financial success confirms the strength of the current hardware investment cycle. It doesn't settle whether the entire AI infrastructure ecosystem will generate attractive long-term returns.

Is the AI boom being inflated by companies investing in their own customers?

Some AI infrastructure spending is supported by circular financial relationships in which technology companies fund AI developers that subsequently purchase computing services from the same ecosystem.

Amazon provides a particularly revealing example. Its 2026 SEC filings show a $28.7 billion investment in OpenAI preferred stock during the first half of the year, followed by an additional $21.3 billion funding commitment. Billions flowing from a supplier into its own customers raise the question of who is funding whom, which we follow in our AI infrastructure report.

The company also invested $10 billion in Anthropic nonvoting preferred stock during the second quarter. Meanwhile, Amazon's public communications described an OpenAI infrastructure commitment exceeding $100 billion.

These arrangements can make commercial sense. AI developers need capital and computing power, while cloud providers want large customers willing to commit to long-term contracts.

But where is the demand really coming from?

Imagine a cloud company investing $10 billion in an AI developer, which then commits a substantial portion of that money to purchasing cloud services. The cloud company receives real revenue, but the arrangement doesn't independently demonstrate that consumers or businesses are paying enough for the developer's products.

We should distinguish three amounts: money invested in AI companies, money those companies commit to infrastructure, and revenue earned from independent end customers.

All three can grow rapidly without being economically equivalent. The more tightly connected those flows become, the harder it is to judge whether infrastructure demand can eventually support itself.

Are AI infrastructure companies like CoreWeave taking on too much debt?

CoreWeave's financial results show how quickly a successful AI infrastructure business can become exposed to debt, depreciation and financing costs.

The company reported $2.575 billion in second-quarter 2026 revenue, compared with $1.212 billion one year earlier. Sales more than doubled, reflecting the enormous appetite for GPU computing capacity.

However, CoreWeave also reported a $626 million quarterly net loss and $640 million in net interest expense. Its interest burden alone represented almost 25% of revenue.

According to its quarterly SEC filing, CoreWeave had approximately $35.6 billion in total debt at the end of June 2026. Its property and equipment totaled approximately $46.7 billion.

The business therefore resembles a capital-intensive infrastructure operator much more than a conventional software company. Its growth depends on buying expensive equipment, securing data center capacity and financing those investments before collecting the revenue.

CoreWeave reported $1.51 billion in adjusted EBITDA for the quarter, showing that its operations generate substantial earnings before interest, depreciation and certain other expenses.

But depreciation and amortization reached approximately $1.39 billion, while interest costs consumed another $640 million. Those expenses help explain why rapid revenue growth has yet to translate into net profitability.

CoreWeave may ultimately build a highly profitable business if utilization remains strong and refinancing stays available. Right now, though, its accounts show how much financial pressure an infrastructure provider can face even while customer demand is booming.

Don't long-term AI contracts protect companies from a bubble?

Long-term AI infrastructure contracts provide valuable protection against weak demand, but their strength depends heavily on the financial health of the customers signing them.

CoreWeave reported approximately $104 billion in contracted revenue backlog in its second-quarter results, before including more than $25 billion in additional customer commitments announced for the beginning of the following quarter.

Microsoft reported $678 billion in commercial remaining performance obligations, up 84% year over year. This includes a broad range of cloud and enterprise services, so it should not be interpreted as exclusively AI infrastructure revenue.

According to Microsoft's fiscal-year earnings commentary, roughly 30% of those obligations were expected to become revenue within the following 12 months. The weighted average contract duration was approximately 2.3 years.

Physical data center contracts also show strong demand. CBRE found that 80.4% of capacity under construction in North America's eight largest markets was already preleased, compared with 74.3% one year earlier.

These agreements make infrastructure investment considerably less speculative than building facilities without identified tenants.

Still, backlog represents expected future revenue subject to contractual terms and the provider's ability to deliver. A commitment from a profitable multinational company carries a different level of risk from one made by a developer dependent on future fundraising.

What we really want to know is how much contracted AI demand comes from customers capable of paying throughout the contract, even if financial markets become less generous.

Could AI GPUs become outdated before their owners make their money back?

AI GPUs can lose economic value surprisingly quickly because each new generation of accelerators changes how cheaply companies can train and run models.

Nvidia has moved through its Hopper and Blackwell architectures while bringing Vera Rubin into production deployments. Each generation aims to deliver more computing performance and better efficiency.

That creates an unusual problem for infrastructure investors. A data center building may remain useful for decades, but the accelerators inside it face a much shorter technological cycle. Chips that wear out faster than the buildings around them are one of the unsolved problems we dig into in our AI infrastructure reality check.

Older GPUs can continue operating and serving valuable workloads. Their economic problem arises when newer systems deliver the same work at lower cost, forcing operators to cut prices or accept lower utilization.

CoreWeave's financial results offer a concrete illustration of the burden. Quarterly depreciation and amortization increased from $559 million to $1.393 billion year over year, an increase of approximately 149%.

That reflects the rapid expansion of the company's asset base rather than proving that existing GPUs are already obsolete. Nevertheless, depreciation is becoming a major expense while the business is still scaling.

The risk becomes particularly serious when expensive hardware is financed with debt. Operators must recover their equipment costs and interest payments before technological improvements undermine the prices they can charge.

Could cheaper AI models hurt the companies building AI infrastructure?

Falling AI inference costs could put pressure on infrastructure providers if computing prices decline faster than customers increase usage.

Model developers are making AI cheaper through smaller specialized models, improved inference software, quantization and more efficient hardware.

Lower costs are generally good for adoption. Businesses can automate customer support, process documents or use coding assistants more frequently when every request becomes cheaper.

But the companies providing computing power still need enough revenue to cover their servers, buildings, electricity and financing costs.

Consider a provider that charges $1 for a standardized unit of AI computing work. If competitive pressure drives the price down to $0.40, the company needs to process 2.5 times as much work merely to maintain its previous revenue, assuming its service mix remains unchanged.

If workload volume triples, revenue grows 20%. If workload only doubles, revenue falls 20%.

The calculations are illustrative rather than forecasts, but they show why enormous growth in AI usage doesn't automatically guarantee equally strong infrastructure revenue.

Greater efficiency may ultimately help application developers and customers more than infrastructure owners. And that's a real possibility, even in a world where AI becomes much more widely used.

Is the world building too many AI data centers?

The world's AI data center pipeline is enormous, but the latest occupancy figures suggest that the main risk is future overbuilding rather than excess operational capacity today.

CBRE recorded 7,481 megawatts under construction across eight major North American markets in the first half of 2026, an increase of 24.8% from the preceding period.

JLL's broader North American report described approximately 66 gigawatts under construction across its market coverage, with around 95% precommitted. The two figures cover different geographies and categories, so combining or directly comparing them would be misleading.

The distinction between construction and delivery is especially important. Projects under development can take years to become operational, and some may be delayed or canceled.

For developers, securing a tenant before construction reduces exposure. However, the financial strength of those tenants and the cost of completing facilities remain important.

The biggest danger would emerge if several years of planned capacity became available just as AI companies reduced spending, computing prices fell or financing became more expensive.

At present, developers are still struggling to deliver enough usable infrastructure. That's a much better position than operating a market already suffering from empty buildings, but it doesn't guarantee that every project in the pipeline will make money.

Could electricity shortages slow down the AI infrastructure boom?

Electricity shortages are already restricting AI data center growth, and the problem is becoming large enough to affect where infrastructure companies build their next facilities.

According to the International Energy Agency's 2026 projections, global data center electricity consumption could rise from approximately 485 terawatt-hours in 2025 to 950 terawatt-hours by 2030.

That represents almost a doubling in five years. Electricity consumption from AI-focused facilities is expected to triple over the same period.

The shortage is particularly visible in established data center hubs. CBRE reported severe power-delivery constraints in Northern Virginia and delays around Chicago that could push some grid connections into 2032 or later.

The consequences go beyond waiting longer for electricity. Developers may need to pay for grid connections, power generation, cooling systems and other equipment that substantially increases the total cost of a project.

Infrastructure companies that already control reliable power access may gain a valuable advantage. Meanwhile, businesses unable to secure electricity could be forced to delay investments regardless of customer demand.

Indicator	Latest figure	What it shows
Global data center electricity use, 2025	485 TWh	Current demand baseline
Projected electricity use, 2030	950 TWh	Approximately 96% higher
Projected AI-focused electricity growth	Approximately 3×	Faster than overall data centers
Data center share of global electricity demand, 2030	Approximately 3%	Large, but still a minority
North American primary-market vacancy	1.4%	Operational capacity remains scarce
Northern Virginia vacancy	0.2%	Almost no available space
Primary-market capacity under construction	7.48 GW	Major development pipeline
Share of construction already preleased	80.4%	Strong demand commitments

Is the AI infrastructure boom repeating the dot-com bubble?

The AI infrastructure boom has a clear resemblance to the telecommunications investment bubble of the late 1990s, particularly in how quickly companies are building capacity based on expectations of future demand.

During the internet boom, telecom operators spent heavily constructing fiber-optic networks. They correctly anticipated that internet traffic would grow enormously, but many overestimated how quickly customers would pay enough to justify the investment. Being right about demand but wrong about timing hurts; we look at who the likely future winners are in our AI infrastructure market report.

When financing conditions deteriorated, heavily indebted companies struggled. Several went bankrupt, even as internet adoption continued expanding.

AI infrastructure could follow a similar path. Cloud providers, chip manufacturers and data center developers are investing in a technology that may become essential to economic activity. That doesn't guarantee attractive returns for every company financing its expansion.

There is, however, a major difference in who is funding the boom.

Microsoft generated approximately $155 billion in fiscal-year 2026 operating income. Alphabet reported $119.8 billion in quarterly revenue with a 34% operating margin. Amazon generated $27.5 billion in quarterly operating income.

These businesses have large, established sources of profit beyond AI. They can absorb investment disappointments that would overwhelm a smaller infrastructure company.

So the dot-com comparison is most convincing for highly leveraged providers and developers whose economics depend heavily on a single infrastructure cycle. The technology could become essential, just as the internet did, while some of the companies financing it end up losing everything.

Who would suffer most if the AI infrastructure boom crashed?

Highly indebted AI cloud providers and data center developers would face the greatest financial pressure in an infrastructure downturn, while diversified technology companies would have more room to absorb losses.

For specialized GPU cloud operators, the danger comes from several costs arriving together. Equipment prices may fall, customer spending may weaken, and lenders may demand higher interest rates just as expensive computing assets need replacement.

CoreWeave's combination of rapid sales growth, substantial debt and significant interest expenses illustrates that exposure.

Data center developers could face another problem. Projects involving land, electricity connections and construction commitments may be difficult to abandon once work has started. Falling demand could leave developers holding expensive unfinished facilities or negotiating contracts with weaker tenants.

Equipment suppliers would experience a different kind of downturn. Nvidia's current financial results are extraordinary, but a slowdown in GPU orders could sharply reduce its revenue growth and put pressure on margins.

Microsoft, Amazon, Alphabet and Meta would also face consequences. Their earnings could suffer from depreciation, weaker cloud returns and reduced free cash flow. Their existing businesses, however, make them considerably less dependent on the success of any single infrastructure project.

The likely damage would be concentrated among businesses with expensive assets, heavy borrowing and limited financial flexibility. Some could face severe losses even if overall AI demand continued growing.

What would prove that the AI infrastructure boom has turned into a bubble?

A sustained rise in unused data center capacity, falling computing prices and deteriorating infrastructure profits would provide convincing evidence that the AI investment cycle has gone too far.

We would first look at actual occupancy. CBRE currently reports exceptionally low vacancy across major North American markets. If that figure began rising consistently while new facilities continued opening, it would indicate that supply was catching up with or exceeding demand.

Pricing would provide another test. Today, CBRE reports higher asking rents across major deployment sizes. A reversal lasting several quarters, especially alongside weaker occupancy, would be more concerning than a temporary slowdown in construction.

For GPU cloud companies, the critical evidence would come from revenue per unit of deployed capacity, utilization, depreciation and financing costs. Falling rental prices are manageable when hardware and operating costs decline faster. They become dangerous when infrastructure owners cannot generate enough cash to service debt.

We would also examine whether major customers continue honoring their long-term contracts. A wave of renegotiations, cancellations or financing difficulties among AI developers would challenge the quality of current demand.

The combination matters more than any individual figure. An AI stock market correction might happen without an infrastructure collapse. Conversely, financial trouble among heavily indebted operators could develop while overall AI usage continues growing.

Right now, occupancy and revenue data remain strong. The financial risks are becoming more visible, but the industry-wide deterioration needed to confirm a full infrastructure bubble has yet to appear.

So, is the AI infrastructure boom actually a bubble?

The AI infrastructure boom is partly a financial bubble, but the underlying demand for computing power is real. The strongest evidence points toward overinvestment risk in certain companies and projects rather than an industry-wide market built on imaginary customers.

We can already establish several things with reasonable confidence.

Operational data center capacity is scarce, cloud revenue is growing quickly, and Nvidia is generating extraordinary profits from the equipment being deployed. Those are substantial commercial results, backed by company filings and property-market data.

The less reassuring evidence comes from the cost of achieving that growth. Capital expenditure is increasing faster than established cloud revenue at some major companies. Specialized infrastructure providers are borrowing heavily. GPU depreciation can become expensive, and some customer commitments depend on businesses that have yet to prove their long-term profitability.

The most revealing comparison may be between Nvidia and CoreWeave. Nvidia generated $63.7 billion in quarterly operating income selling computing equipment. CoreWeave more than doubled its quarterly revenue but still lost $626 million after accounting for its operating expenses, financing and other items.

Both companies are benefiting from the same AI investment cycle, yet the financial outcomes are radically different. Strong demand for a technology clearly doesn't guarantee strong returns for every company providing it.

We expect the biggest risk to sit with expensive, debt-financed infrastructure whose returns depend on consistently high utilization and favorable computing prices. The largest hyperscalers can afford more mistakes, although their shareholders may still pay for disappointing investment returns.

Calling the entire AI infrastructure boom a bubble goes too far. Calling it financially safe would be equally misleading.

Our conclusion is that AI infrastructure is going through a genuine industrial expansion with speculative levels of spending in parts of the market. The technology can succeed, demand can keep growing, and investors can still lose enormous amounts of money by financing that growth at the wrong price.