Is the AI Data Center Boom a Bubble? Stocks, ETFs & Risk

Short answer: the AI data-center boom shows several classic warning signs of a financial bubble, but it also has real, structural differences from past ones. Five hyperscalers plan to spend roughly $660 to $690 billion on AI infrastructure in 2026 alone, according to Futurum Group's 2026 capex tracking. Much of that spending is now tangled up in "circular deals" between chipmakers, cloud providers, and AI labs, plus tens of billions in debt moved off company balance sheets. Whether that adds up to a bubble or a normal, if unusually large, infrastructure cycle is the live debate this page walks through.

This is a financial and investor-focused explainer. If you're looking for how data centers affect water, electricity bills, and nearby communities, see our data center impact explainer and the interactive data center map. Both connect to the wider AI backlash this site tracks, but for different reasons than the ones below.

What's actually being built

The scale of construction is historically large. Microsoft, Alphabet, Amazon, Meta, and Oracle have collectively committed to between $660 billion and $690 billion in AI and cloud capital expenditure for 2026, nearly doubling 2025 levels, according to Futurum Group. Goldman Sachs separately revised its own 2026 compute capex estimate upward, from $465 billion to $527 billion, in its "Tracking Trillions" research note.

That money is buying land, power contracts, cooling systems, and enormous numbers of GPUs, packed into single campuses that can draw as much electricity as a small city. The pace is what makes this cycle unusual. Multiple hyperscalers are trying to build multi-gigawatt sites at the same time, competing for the same scarce transformers, land, and grid connections.

Circular deals, explained

A circular deal happens when a chipmaker or cloud provider invests in an AI lab, and that lab then spends the money buying products from the same investor. Nvidia agreed to invest up to $100 billion in OpenAI, while OpenAI separately committed to purchasing roughly $250 billion in cloud services from Microsoft and billions more in chips from AMD, according to reporting from Bloomberg's tracking of AI circular deals and TechCrunch's coverage of the infrastructure deals.

CoreWeave sits at the center of a similar web. Nvidia holds a stake in CoreWeave, which has a cloud deal with Oracle and is also one of Microsoft's largest customers, with additional multi-billion-dollar contracts tied to OpenAI and Meta. Money keeps circulating through the same small group of companies rather than flowing in from genuinely new, external customers.

Why this worries analysts

The concern is straightforward: circular deals can make demand look bigger than it really is. If Nvidia's investment in OpenAI mostly comes back to Nvidia as chip purchases, that revenue is not proof of independent market demand — it is the same capital passing through the loop. Bloomberg's reporting notes this structure creates skewed incentives and can magnify losses across the whole group if AI revenue growth ever slows.

The off-balance-sheet debt problem

More than $120 billion in AI data-center debt has been shifted off corporate balance sheets and into special-purpose vehicles, or SPVs, according to financial reporting compiled by outlets including Techerati and Cryptopolitan. In this structure, the SPV — not the tech company — legally owns the land, buildings, and chips, while the tech company simply leases the finished data center.

Meta's roughly $30 billion Hyperion data center project in Louisiana is a widely cited example. The financing vehicle raised about $27 billion in loans from asset managers including Pimco, BlackRock, and Apollo, plus roughly $3 billion in equity from Blue Owl Capital. Oracle used a similar approach for multiple sites: a $13 billion package from Blue Owl and JPMorgan backed its Abilene, Texas facility, alongside a separate $38 billion debt package covering sites in Texas and Wisconsin and an $18 billion loan for a New Mexico location.

Why it matters

Keeping this debt off the parent company's balance sheet protects its credit rating and keeps leverage ratios looking healthy. But it also makes the true scale of AI-related debt harder for investors to see. If AI demand slows and lease payments stop covering the debt, the risk lands first on the private-credit lenders — but a wave of defaults across dozens of SPVs could still ripple back into the broader credit market.

The depreciation fight

Investor Michael Burry has made the most detailed public case that hyperscaler profits are overstated. His argument centers on depreciation: how quickly a company writes off the cost of its GPUs. Burry contends that Meta, Amazon, Microsoft, Google, and Oracle depreciate Nvidia chips over five to six years, when the real useful life is closer to two and a half years given how fast each new chip generation makes the last one obsolete, according to reporting from Yahoo Finance.

If Burry's estimate holds, it could understate combined industry depreciation by roughly $176 billion between 2026 and 2028, artificially inflating reported profits by around 20 percent across these companies. Burry claims the effect is largest at Oracle, where he estimates profits are overstated by 48 to 62 percent. Nvidia disputed the analysis directly, taking the unusual step of sending sell-side analysts a memo pushing back on Burry's stock-based compensation and depreciation math — itself a sign of how seriously Wall Street is taking the critique.

Bubble buildout vs. a normal capex cycle

Not every large capex cycle is a bubble. The table below compares features of a typical infrastructure buildout against the specific warning signs analysts are flagging in the current AI data-center cycle.

Are data centers actually getting canceled?

Reports on this question conflict sharply. Tom's Hardware and NBC News both reported that roughly half of planned U.S. data-center builds face delay or cancellation in 2026, and that opposition groups had blocked or delayed projects worth about $130 billion in the first months of the year, driven largely by shortages of grid components like transformers and switchgear rather than a lack of capital or chips.

SemiAnalysis pushed back hard on that framing, arguing that its own hyperscaler forecast for year-end 2026 moved by only about 1 percent over six months. Its analysts say most of the "canceled" megawatts never had site control, financing, or grid interconnection secured in the first place, and that new regulatory filters — like Texas's site-control and curtailment requirements — are simply weeding out speculative filings that were never going to get built.

Both things can be true at once: the serious, well-financed hyperscaler pipeline may be largely intact, while a much larger number of speculative or under-funded proposals fail. That distinction matters for anyone reading capex headlines — a canceled speculative filing is not the same signal as a canceled Microsoft or Google project.

AI data-center stocks and ETFs

Investors looking for exposure to this theme generally sort options into a few buckets: chipmakers (Nvidia, AMD, Broadcom), data-center REITs (Equinix, Digital Realty), power and cooling suppliers (Vertiv, Constellation Energy, Vistra), and neocloud operators that rent out GPU compute (CoreWeave, Applied Digital). Diversified options exist too — the Global X Data Center & Digital Infrastructure ETF (DTCR), for example, holds a basket of data-center REITs rather than betting on a single company.

None of these categories are insulated from the risks described above. A REIT's tenant might be a hyperscaler funded partly through off-balance-sheet debt; a chipmaker's revenue might include circular-deal dollars; a power supplier's growth assumes data-center demand keeps compounding. Anyone considering this sector should read a company's customer concentration and financing structure, not just its growth headline, before deciding it's a safe bet. This page is informational, not financial advice.

How this connects to the AI backlash

The data-center bubble debate is one thread inside a much larger public reaction against unchecked AI expansion, which we track across the AI backlash pillar. Financial skepticism and community opposition are reinforcing each other: the same power-grid shortages slowing construction are also the reason nearby residents are fighting new projects over electricity costs, covered in depth on our data center impact explainer.

If hyperscaler capex ever pulls back sharply — whether from a funding crunch, a depreciation reckoning, or simply weaker AI product revenue than promised — it would validate a core argument of AI critics: that the industry's growth was propped up by circular financing and investor momentum rather than by proven, durable demand. You can see where specific proposed and operating facilities stand today on our interactive data center map.