Is the AI infrastructure boom a bubble?

Why does the AI infrastructure boom look bubbly now?

The AI infrastructure boom looks bubbly now because investment is racing ahead of the broad economic payoff.

The latest International Energy Agency assessment says capital spending by five large technology companies exceeded $400 billion in 2025 and could rise another 75% in 2026. That pushes the annual pace into a range few private investment cycles have ever reached. JLL separately expects close to 100 gigawatts of new data center capacity by 2030, requiring as much as $3 trillion once buildings, power systems, chips and networking equipment are included.

Meanwhile, the Federal Reserve’s latest review says the economic effects of AI remain concentrated in a limited part of the economy. Company adoption is rising, but usage inside many adopters remains shallow, and broad productivity data still show little evidence of an economy-wide transformation.

A recent BIS working paper sharpened the concern. Its model, calibrated with company balance sheets and deal data, estimates that competitive pressure could push AI investment to roughly 1.5 times the efficient level, and as high as three times when demand reacts weakly to lower prices. It is a model rather than a direct measurement, but it captures the central problem: every major player fears that spending too little today could cost it the market tomorrow.

If you want more recent data on this point, please see our latest AI infrastructure market report.

What would an AI infrastructure bubble actually mean?

For AI infrastructure, a bubble means building assets whose future cash flows cannot justify their price, timing or financing.

AI can become a major technology and still produce a painful infrastructure bust. Fiber networks were essential to the internet, yet many telecom investors lost money because they installed too much capacity too early and financed it badly. The useful test is whether data centers, power contracts, chips and cloud capacity can earn acceptable returns under realistic demand and pricing.

There are three different risks. A capacity bubble would leave facilities or servers underused. A return bubble would keep the equipment busy but force prices so low that owners earn weak returns. A financing bubble would expose companies that must keep paying interest, leases and suppliers even when customers delay spending.

Current evidence weighs against a broad capacity bubble in delivered data centers. Future returns and debt-heavy financing are much more worrying. That is how strong chip sales and record-low vacancy can coexist with a serious bubble debate.

How big is AI infrastructure spending now?

AI infrastructure spending is now large enough for a modest forecasting error to wipe out tens of billions of dollars.

The expected annual capital expenditure of the major technology companies has climbed beyond $700 billion, according to the IEA. Alphabet currently expects $180 billion to $190 billion of capital spending for the year. Meta expects $125 billion to $145 billion. Microsoft added $30.9 billion of property and equipment in its latest reported quarter, while Amazon says its trailing free cash flow fell to $1.2 billion mainly because property and equipment purchases increased by $59.3 billion, primarily for AI.

The scale becomes clearer beside entire industries. Alphabet and Meta together may spend more than $300 billion in one year, before counting Microsoft, Amazon, Oracle, CoreWeave or the data center developers supporting them. A 10% mistake across that annual investment base equals roughly $70 billion of capital arriving too early, in the wrong place or in hardware that earns less than expected.

The spending surge is still accelerating. The BIS describes the current buildout as one of the largest technology-driven investment booms in US history. The description fits: companies are committing the equivalent of several major national infrastructure programs every year, largely on the assumption that AI usage will keep compounding.

Company or group	Latest spending indicator	What it tells us
Five large technology companies	Spending could rise 75% from above $400B	The boom has reached an economy-shaping scale
Alphabet	$180B to $190B expected	One company may spend close to $500M per day
Meta	$125B to $145B expected	Advertising profits are funding a huge AI capacity bet
Microsoft	$30.9B of property and equipment added in one quarter	Infrastructure spending has become a recurring quarterly commitment
Amazon	Property and equipment purchases rose $59.3B over twelve months	AI investment has absorbed almost all trailing free cash flow

Is demand for AI chips and cloud capacity real today?

Yes, demand for AI infrastructure is very real today, with revenue surging across chips, foundries and cloud platforms.

Nvidia’s latest quarter produced $75.2 billion of data center revenue, 92% more than a year earlier. TSMC’s latest quarterly revenue reached $40.2 billion, up 33.7% in US dollars, while its gross margin climbed to 67.7%. The rush clearly reaches beyond one chip designer: the foundry manufacturing the most advanced chips is also running at exceptional profitability.

The cloud results tell the same story from another angle. Microsoft’s AI business passed a $37 billion annual revenue run rate and grew 123% year over year. Azure revenue rose 40%. AWS reached $37.6 billion of quarterly sales, up 28%, and generated $14.2 billion of operating income. Google Cloud crossed $20 billion of quarterly revenue, grew 63% and reported a backlog above $460 billion.

Those results do not prove that every cloud customer has a profitable AI product, and backlog can stretch over several years. Still, this is paid demand across several layers of the supply chain, backed by rapidly growing revenue rather than conference-stage promises.

Infrastructure layer	Latest result	Why it is important
Nvidia data center	$75.2B quarterly revenue, up 92%	Demand for leading AI accelerators remains extraordinary
TSMC	$40.2B quarterly revenue, 67.7% gross margin	The boom is flowing into advanced chip manufacturing
Microsoft AI and Azure	$37B AI annual run rate; Azure up 40%	AI capacity is already producing material service revenue
AWS	$37.6B quarterly revenue; $14.2B operating income	Large-scale cloud infrastructure remains highly profitable
Google Cloud	More than $20B quarterly revenue; backlog above $460B	Customers are making large, long-term cloud commitments

Can the hyperscalers afford this spending race?

The biggest hyperscalers can afford the AI infrastructure race today, although affordability says little about the eventual return.

Microsoft reported $38.4 billion of quarterly operating income. Meta produced $32.2 billion of operating cash flow while keeping a 41% operating margin. Alphabet generated $45.8 billion of operating cash flow and held $126.8 billion in cash and marketable securities. Together, those three businesses generated nearly $125 billion of operating cash flow in one quarter.

Their existing products are also benefiting from AI. Meta’s latest revenue increased 33% as ad impressions rose 19% and the average price per ad rose 12%. Google says AI is supporting Search engagement, subscriptions and cloud demand. Google Cloud’s operating margin expanded sharply while revenue accelerated. Microsoft is selling AI through Azure, GitHub, Microsoft 365 and its broader cloud stack.

Amazon shows the trade-off more clearly. AWS is growing at its fastest rate in 15 quarters, yet Amazon’s infrastructure purchases have pushed trailing free cash flow close to zero. That may be sensible during a rare platform shift, but shareholders are giving up cash today for profits that may arrive years later.

The top of the market has financial staying power. A poor return would hurt margins and shareholder value long before it threatened the survival of Alphabet, Microsoft, Meta or Amazon. Suppliers and leveraged capacity providers would feel a spending slowdown much faster.

Is business AI adoption deep enough to support the boom?

Business AI adoption is growing quickly, but it is still too shallow to justify the full infrastructure pipeline on its own.

Federal Reserve analysis of Census Bureau data found that about 18% of US firms had adopted AI by the end of 2025. A worker survey put job-related generative AI use near 41%, while an executive survey found that 78% of workers were employed by companies reporting some AI adoption. These figures describe different units, but they reveal the same pattern: many people now work inside an “AI-using” company even when only part of that company uses the technology regularly.

The Fed’s newest assessment is especially useful here. Headline adoption does not measure intensity, and available surveys suggest that usage remains shallow in many firms. Productivity experiments often show employees completing selected tasks faster, yet those gains have barely appeared in aggregate productivity data. Integrating AI into an entire workflow is harder than giving employees access to a chatbot.

Infrastructure investors need depth more than logos. A company testing an assistant in one department creates limited demand. Continuous agents, customer-facing inference, automated software development and AI embedded in millions of transactions create a very different load.

Deeper use is likely because model capability is improving and prices are falling. For now, though, the infrastructure market is building for that next stage before the broader economy has reached it.

If you want more recent data on this point, please see our latest AI infrastructure market report.

Does record-low data center vacancy prove there is no bubble?

Record-low data center vacancy proves that available capacity is scarce today, although it cannot settle what happens to projects opening several years from now.

CBRE’s latest global review tracked 16 major markets. Their combined supply grew 25% in one year to 16 gigawatts, while vacancy still fell from 8.3% to 6.7%. Northern Virginia reached 0.3% vacancy, Atlanta 1%, Dallas–Fort Worth 1.8% and Chicago 2.2%.

The absorption data are even stronger. The four main North American markets absorbed 2,236 megawatts in one quarter, 34% more than a year earlier. Northern Virginia added 1,136 megawatts of inventory and absorbed 1,148 megawatts. Dallas–Fort Worth had 717 megawatts under construction, with 88% already preleased.

Empty speculative buildings would produce rising vacancy and discounting. Current facilities instead show tight availability, high preleasing and continued rent increases. Anyone calling the whole operating data center market a bubble has to explain those numbers.

The uncertainty sits further out. Construction takes years, leases may depend on a few large tenants, and power constraints hold back supply. Today’s shortage can persist while developers quietly create too much capacity for the early 2030s.

Could today’s data center shortage turn into a glut?

A future data center glut is plausible because the planned expansion assumes that AI will create most of the industry’s growth.

JLL expects global capacity to rise from roughly 103 gigawatts to 200 gigawatts by 2030. It also expects AI to grow from about 25% of workloads to 50%. Combining those assumptions gives a useful result: AI-related capacity would rise from around 26 gigawatts to 100 gigawatts, while every other workload combined would move from roughly 77 gigawatts to 100 gigawatts.

That leaves AI responsible for about 74 of the 97 additional gigawatts, or more than three-quarters of total net growth. It is a much tougher bet than simply expecting normal internet and cloud demand to continue.

Power shortages will slow the delivery. JLL says grid connections in major markets now take more than four years on average, and the IEA reports tightening supplies of transformers, turbines and electrical equipment. Slow delivery reduces the chance of a sudden wall of new capacity, but developers can still lose money on land, permits, equipment orders and financing before a delayed facility earns revenue.

The riskiest projects are distant campuses marketed around future gigawatts rather than powered buildings with signed customers. They may eventually open into a larger AI market, but many currently depend on demand, power access and financing all arriving on schedule. That is a lot to get right at once.

Approximate global workload capacity	2025	2030 forecast	Increase
Total data center capacity	103 GW	200 GW	97 GW
AI workloads	26 GW	100 GW	74 GW
Other workloads	77 GW	100 GW	23 GW
AI share of capacity	25%	50%	About 76% of net growth comes from AI

Will AI inference grow enough to fill the new capacity?

AI inference should become the main source of compute demand, but nobody can currently show that it will fill every campus being planned.

Training a frontier model creates a large but occasional burst of demand. Inference repeats whenever a user asks a question, a coding agent runs, an advertisement is selected or software completes an automated task. That recurring pattern gives inference a much larger long-term ceiling.

JLL expects inference to overtake training as the main AI workload around 2027. It also says the applications needed to support its 2030 demand forecast remain far below the required scale. That deserves more attention than the headline growth estimate. The industry is building for continuous, widely distributed AI use that remains early today.

The IEA offers encouraging evidence. Data center electricity consumption increased 17% in 2025, far above the 3% rise in global electricity demand. It expects total data center power use to double by 2030 and AI-focused consumption to triple, even as energy required per task falls quickly.

Inference should absorb a great deal of infrastructure, especially inside search, advertising, coding, consumer assistants and enterprise software. The harder question is distribution. A few platforms may generate enormous volumes while specialized clouds, remote campuses and older GPU fleets struggle to win enough of that traffic.

Do falling AI costs strengthen or weaken the infrastructure boom?

Cheaper AI strengthens total usage while making the economics of individual infrastructure assets much harsher.

Epoch AI estimates that global AI computing capacity has grown about 3.3 times per year since 2022, equivalent to doubling every seven months. It also finds that chip performance per dollar has improved around 37% annually across more than 20 accelerators released since 2012. Model inference prices have fallen even faster for many comparable capability levels.

Lower prices invite more consumption. Developers use longer contexts, run several reasoning attempts, add agents and serve users who were previously too expensive. Electricity use rising despite huge efficiency gains suggests that this rebound effect is already powerful.

Asset owners face the other side. A server can remain busy while earning far less per hour. New chips complete more work with less power, smaller models take over routine tasks, and cloud platforms keep optimizing their software. An owner who borrowed against today’s rental rates can miss its return target even with respectable utilization.

Falling prices are bullish for AI adoption and potentially brutal for undifferentiated compute sellers. The biggest winners may be the companies using cheap intelligence, rather than every company that financed the hardware.

If you want more recent data on this point, please see our latest AI infrastructure market report.

Are AI infrastructure deals becoming too circular?

Yes, several AI infrastructure deals are now circular enough to hide how concentrated the underlying demand really is.

The BIS recently described a financing chain in which hyperscalers take minority stakes in infrastructure vehicles, commit to long-term leases or capacity purchases, and help those vehicles borrow from private credit funds. The obligations may sit outside the hyperscaler’s reported debt even though its lease or purchase commitment ultimately supports the financing.

CoreWeave offers a clear public example. Nvidia supplies its key chips and invested $2 billion in the company. CoreWeave then uses debt and customer commitments to expand its infrastructure, while large cloud companies and AI laboratories buy the resulting capacity. CoreWeave reported almost $100 billion of revenue backlog, which looks highly diversified at first glance.

Its SEC filing gives a different picture. The top two customers generated 65% of quarterly revenue, with the largest accounting for 45%. CoreWeave also expects OpenAI and Meta to remain significant customers under multiyear agreements. When a supplier, investor, customer and competitor appear repeatedly inside the same network, several transactions can ultimately depend on the same few sources of cash.

These contracts are genuine and often include strong commitments. Circularity becomes dangerous when investors treat each link as independent demand. A pullback by one major AI laboratory or hyperscaler could reduce cloud purchases, chip orders, data center leases and lenders’ willingness to refinance at the same time.

Is debt making the AI infrastructure boom dangerous?

Debt is currently the clearest dividing line between a costly technology bet and a potential financial accident.

CoreWeave ended its latest quarter with $25.1 billion of debt and $10.1 billion of operating lease liabilities. It generated $2.08 billion of revenue, recorded a $144 million operating loss and paid $536 million of net interest expense. Interest alone equaled about 26% of revenue.

The company also reported $1.16 billion of adjusted EBITDA, but depreciation and amortization on property and equipment reached $1.1 billion in the same quarter. That expense reflects an unavoidable economic reality: GPUs, networking hardware and data center equipment wear out or lose value while debt remains due.

Its backlog offers meaningful protection, though much of the cash sits years away. CoreWeave expects only 36% of its $98.8 billion in remaining performance obligations during the first 24 months, another 39% during months 25 to 48, and the rest later. It also disclosed $40.7 billion of future lease payments tied to leases that had not started.

The wider market is moving in the same direction. The BIS calls some special-purpose data center structures “shadow borrowing” because they shift debt into vehicles supported by hyperscaler leases and guarantees. Banks, insurers and private credit funds then become part of the same chain.

Debt makes sense for long-lived infrastructure with dependable cash flows. AI hardware changes too quickly for lenders and investors to assume that every long contract will remain equally valuable throughout its life.

If you want more recent data on this point, please see our latest AI infrastructure market report.

Will expensive GPUs become outdated before they pay for themselves?

Some expensive GPUs will lose their premium pricing well before their financing ends, even though the machines will keep doing useful work.

Epoch AI’s 37% annual improvement in chip performance per dollar compounds quickly. At that pace, a new accelerator four years later offers roughly 3.5 times as much performance for the same dollar. Nvidia’s rapid product cycle adds pressure because customers keep moving toward newer systems with more memory, faster networking and better energy efficiency.

CoreWeave shows the financial effect in real time. Its quarterly depreciation and amortization on property and equipment rose from $443 million to $1.1 billion in one year. Meanwhile, the company is already preparing to deploy Nvidia’s Rubin platform while previous generations remain on its balance sheet.

Older chips rarely become worthless overnight. They can serve smaller models, batch jobs, fine-tuning and cheaper inference. Their position simply moves down the price ladder. The newest fleet earns a premium, the previous generation handles mainstream work, and older hardware competes for increasingly price-sensitive jobs.

Hyperscalers can soften that problem by shifting older chips across internal products. A specialist GPU lessor must keep external customers willing to pay enough. Hardware aging is therefore a much bigger threat to leveraged neoclouds than to companies running search engines, advertising systems and software platforms.

Is this the telecom fiber bubble happening again?

The telecom fiber boom is the closest historical comparison, although today’s hyperscalers enter the cycle with much stronger businesses.

During the late-1990s telecom expansion, internet traffic was genuinely growing and fiber was clearly useful. Operators still built too much too early, relied on optimistic demand forecasts and financed networks that could not earn enough once prices fell. Bankruptcies followed, while the surviving fiber later became essential infrastructure.

AI shares that awkward combination of technological truth and investment risk. Compute demand is rising, unit costs are falling, and companies fear losing a winner-take-most race. The BIS’s newest work argues that this competitive pressure can cause firms to overcommit because each company benefits from winning, while the cost of industry-wide overbuilding falls on everyone.

Today’s strongest builders have several advantages. Microsoft, Alphabet, Amazon and Meta produce enormous cash flows, already operate profitable digital platforms and can reuse capacity across many services. Current data center vacancy is also extremely low, whereas the telecom bust exposed large amounts of excess capacity.

The comparison becomes much tighter among heavily indebted clouds and speculative developments. Their returns depend on future traffic arriving quickly, prices staying high enough and a small group of customers honoring very large commitments. The internet eventually justified huge fiber networks, but it did not rescue every company that financed them.

Who would get hurt first if AI infrastructure spending slows?

Leveraged GPU clouds, speculative campuses and owners of older hardware would feel an AI spending slowdown first.

The hyperscalers could cut orders and still operate profitable cloud, advertising, retail and software businesses. A neocloud with fixed interest and lease payments has far less room. A developer holding land and partial power rights may have no operating revenue at all. An older GPU fleet can remain technically useful while suffering a sharp fall in rental income.

The earliest warning signs would probably appear in financing and contracts before they appear in headline AI usage. We would watch for repeated capex cuts by several hyperscalers, weaker preleasing, contract renegotiations, lower GPU rental rates, rising data center vacancy, falling utilization and widening credit spreads among infrastructure borrowers.

None of those indicators currently point in the same negative direction. Chip revenue, cloud sales, electricity use, absorption and preleasing remain strong. A slowdown could still begin with only one or two weak links, especially where customers, suppliers and lenders overlap.

Participant	Exposure in a slowdown	Likely problem
Speculative campus developers	Very high	Power, tenants or refinancing arrive too late
Leveraged neoclouds	High	Rental prices and utilization fail to cover debt and leases
Older GPU fleets	High	New hardware pushes them into lower-priced workloads
Data centers with strong long-term tenants	Moderate	Customer concentration or contract renegotiation
Chip and equipment suppliers	Moderate	Orders fall after the largest construction phase
Cash-rich hyperscalers	Lower survival risk, meaningful return risk	Margins and free cash flow weaken
Powered operating sites in tight markets	Lowest near-term risk	Scarcity supports occupancy and residual value

So, is the AI infrastructure boom a bubble?

Partly yes: the AI infrastructure boom is real at its core, while the spending race around it has already become excessive in some places.

Today’s demand is too strong to dismiss. Nvidia and TSMC are posting extraordinary growth, cloud platforms are selling billions of dollars of AI services, electricity use is climbing, and available data center capacity remains scarce. The biggest technology companies also have enough cash flow to keep building through a period of disappointing returns.

The bubble sits mainly in the assumptions wrapped around future capacity. The industry plans to nearly double global data center capacity by 2030, with AI responsible for more than three-quarters of the increase. That requires inference applications to scale far beyond their current footprint. At the same time, chip performance improves quickly, compute prices fall and financing is spreading through debt, leases and special-purpose vehicles.

A selective correction is more likely than a universal one. Powered data centers with strong tenants, modern hardware and flexible workloads should keep their value. Some remote megaprojects, debt-heavy GPU clouds and older fleets will probably discover that fast-growing demand still falls short of the price they paid for it.

AI infrastructure has entered the dangerous stage of a genuine boom: the technology is working, customers are spending and too many investors are treating those facts as a guarantee of high returns.

If you want more recent data on this point, please see our latest AI infrastructure market report.

Who is the author of this content?

NEW MARKET PITCH TEAM

We track new markets so founders and investors can move faster

We build living "market pitch" documents for emerging markets: AI, synthetic biology, new proteins, and more. Instead of outdated PDFs or hallucinated LLM answers, our clients get a clean, visual, always-updated view of what's really happening: key players, deals, regulations, and signals that matter. Learn more about us.