Is There an AI Data Center Bubble? The Bull and Bear Case

Is there an AI data center bubble? It is the most contested question in technology investing in late 2026. This article, part of the Data Center Economy hub, lays out the bull case and the bear case side by side, using reported figures rather than slogans, and identifies the indicators that will decide who is right. For the spending numbers in detail, see the AI capex supercycle; for the long-range demand scenarios, read our data center market forecast to 2030.

How much is being spent on AI data centers in 2026?

The four largest US hyperscalers guide to roughly $720–745bn of capital spending in 2026, most of it on servers and data centers. Alphabet has signalled that its 2027 capex will exceed $300bn, so the curve is still steepening rather than flattening.

Company guidance compiled in October 2026 shows Amazon at about $220bn, Alphabet at $195–205bn, Microsoft at about $175bn for calendar 2026 and Meta at $135–145bn. Measures differ — Meta includes finance-lease principal, and Microsoft’s figure dropped from about $190bn after a lease reclassification — so the total is approximate. The Financial Times put the combined four-company figure 77% above 2025’s record $410bn when guidance stood at about $725bn in April.

Quarterly numbers confirm the pace. Microsoft spent $39.4bn on property and equipment in the quarter to September 2026, roughly double a year earlier. Meta’s capex reached $34.91bn in the third quarter, up 80%. Alphabet’s capex was $51.4bn, according to an AI Weekly summary of its filing. Dell’Oro Group expects total data center capex to exceed $1tn in 2026.

What is the bull case for AI data centers?

Bulls argue that demand is real, measurable and broadening: cloud revenue is accelerating, contracted backlogs are enormous, capacity is sold out and the biggest spenders are profitable companies with strong balance sheets rather than dot-com start-ups.

Revenue is accelerating, not fading

Synergy Research Group estimated third-quarter 2026 cloud infrastructure revenue at $167.7bn, up 55% year on year — a growth rate it described as surging to well over 50%. Microsoft reported Azure growth of 48% and Microsoft Cloud revenue of $63.9bn. Google Cloud revenue rose to $30.7bn, more than doubling, according to AI Weekly’s summary of Alphabet’s results. Nvidia’s data center revenue reached $89.0bn in the quarter to July 2026, up 117%.

Demand is contracted years ahead

Microsoft’s commercial remaining performance obligation reached $677bn, up 73%. CoreWeave reported a backlog of about $104bn at mid-2026, plus more than $25bn of new commitments early in the third quarter. Amazon CEO Andy Jassy said AWS capacity was sold out for 2026, largely sold out for 2027, and that a substantial portion of 2028 was committed, according to IG’s earnings preview. JLL put North American data center vacancy at about 1% for a third straight year.

Valuations and balance sheets look nothing like 2000

In December 2025, Bloomberg data showed the Nasdaq 100 trading at about 26 times projected profits, versus more than 80 times at the dot-com peak. Supporters such as Janus Henderson’s Alison Porter told AI Business that hyperscalers have multiple revenue streams and that risks at individual firms like Oracle are not systemic. Revenue at AI labs is also growing fast: OpenAI’s CFO put 2025 revenue at about $20bn, up from roughly $2bn in 2023.

What is the bear case for an AI data center bubble?

Bears argue that spending has far outrun the revenue it can plausibly generate, that much of the demand is financed by suppliers and debt, and that accounting choices flatter profits. They see classic late-cycle signals: circular deals, rising leverage and falling free cash flow.

The revenue gap

Sequoia Capital’s David Cahn framed the problem in 2024 as AI’s “$600B question”: roughly four times Nvidia’s run-rate revenue would need to be earned by end users to justify the hardware, once data center costs and customer margins are included. Bain & Company’s 2025 technology report estimated that about $2tn of annual revenue would be needed by 2030 to fund expected compute demand, leaving roughly $800bn unaccounted for even after AI-driven savings.

Cash flow and debt

Meta’s capex exceeded its operating cash flow in the third quarter of 2026, producing free cash flow of -$2.12bn versus $10.63bn a year earlier, and it bought back no shares. Meta also disclosed $454.5bn of non-cancellable contractual commitments, according to FourWeekMBA’s review of its 10-Q. CNBC, using S&P Global data, reported that data center-related debt issuance nearly doubled to $182bn in 2025. Microsoft was reported in late October 2026 to be preparing its first bond sale since 2017. Oracle, which signed a $300bn compute deal with OpenAI in September 2025, was cited by AI Business as a source of credit concern.

Profitability at the AI labs

Harvard Business School’s Andy Wu told AI Business that most AI companies lose enormous amounts of money. An MIT study reported in August 2025 found most organisations saw no measurable return from generative AI, though an Omdia survey of 350 enterprises reached the opposite conclusion. If AI labs cannot grow revenue fast enough, the contracts underpinning data center leases would come under pressure.

What is circular financing and why does it worry investors?

Circular financing describes deals where a supplier invests in, lends to or guarantees a customer that then buys the supplier’s products. It can accelerate genuine build-outs, but it also makes demand look stronger than it would be on independent funding alone.

The most cited example is Nvidia’s relationship with OpenAI. In August 2026 Nvidia signed a partnership guaranteeing up to $105bn of obligations for an OpenAI data center campus in Pike County, Ohio, being built by SoftBank-backed SB Energy, according to Fortune. Earlier reports had suggested a backstop of up to $250bn, and Nvidia shares fell about 4.5% intraday when that figure surfaced. Nvidia also invested $1.5bn in SB Energy. The campus could reach 8 GW, with the first 800 MW expected in 2028, and Nvidia’s materials put the cost at roughly $50–60bn per gigawatt.

Other links reinforce the pattern. Nvidia bought $2bn of CoreWeave stock in January 2026. Microsoft, Nvidia and xAI are members of the partnership that bought Aligned Data Centers for about $40bn. Defenders, including Nvidia CEO Jensen Huang, dispute that the arrangements are circular and argue they unlock capacity customers genuinely need. Our coverage of the neoclouds and the biggest data center deals details these relationships.

Are GPUs being depreciated too slowly?

The depreciation debate asks whether hyperscalers’ five-to-six-year useful lives for servers match the economic life of AI chips. If chips become obsolete faster, reported profits are overstated today and write-downs could follow.

Michael Burry argued in November 2025 that Meta, Amazon, Microsoft, Alphabet and Oracle were depreciating Nvidia hardware over five to six years when its economic life is closer to two to three, overstating profits by about $176bn between 2026 and 2028. Practices differ: Amazon shortened the life of a subset of servers from six to five years in early 2025, citing faster AI technology development, while Meta extended its server life to 5.5 years, reducing depreciation expense by about $2.9bn, according to an analysis by Dave Friedman.

The counter-argument is that GPUs cascade from training to inference to batch work. Nvidia’s CFO reported six-year-old A100s running at full utilisation, CoreWeave has an A100 rental contract running to 2029, and Google’s AI infrastructure lead said seven- and eight-year-old TPUs were fully used. Rental prices are mixed evidence: H100 rates fell from more than $8 an hour in 2023 to $1–2 by late 2024, but an index from Ornn Exchange put them around $3.17 an hour in September 2026. Bloomberg Intelligence noted depreciation at Alphabet, Microsoft and Meta had already risen from about $10bn in Q4 2023 to nearly $22bn in the September 2025 quarter.

How does today’s build-out compare with the dot-com and telecom bubbles?

The closest parallel is the late-1990s fibre build-out, when carriers borrowed heavily to lay capacity that took years to fill. The key differences today are who is spending — mostly cash-generative giants — and how much capacity is pre-leased.

In 2000, many builders were young, debt-funded companies with little revenue. Today, the biggest spenders earn hundreds of billions in operating cash flow, and Cushman & Wakefield reported that 91.7% of the record 37.7 GW under construction in the Americas was precommitted in H1 2026. That reduces the risk of empty buildings.

But the parallels are growing. Debt is playing a larger role, through bonds, private credit and off-balance-sheet joint ventures, as explained in how data centers are financed. A fast-falling cost of compute could, like falling bandwidth prices in 2001, help users while hurting owners of older capacity. And the physical constraint of power, discussed in our power bottleneck analysis, may slow supply enough to prevent a glut — or strand projects that cannot get grid connections.

What would a correction look like for data center investors?

A correction would most likely hit the most leveraged and least-contracted players first — speculative developers, neoclouds and suppliers with concentrated order books — before reaching investment-grade landlords with long hyperscaler leases.

S&P Global Market Intelligence described neoclouds, whose funding tripled to more than $35bn in 2025, as being in an unenviable position as chip supply improves, and warned that a demand or supply shock could trigger consolidation. Supplier stocks tend to fall first because orders are discretionary. REITs with 15-year leases to AA-rated tenants, such as those in Digital Realty’s 2026 Northern Virginia acquisition, would be better protected, though their valuations could still compress.

Some analysts expect a rotation rather than a crash. Value Point Capital’s Sameer Bhasin and BlackRock’s Tony DeSpirito both argued in late 2025 that AI valuations were not at dot-com levels. Investors weighing exposure can compare options in our guide to investing in data centers.

Which indicators should investors watch?

The most useful signals are hyperscaler capex guidance, AI company revenue growth, GPU rental prices, data center vacancy and pre-leasing, credit spreads on data center debt and the pace of grid connections.

<

p style=”font-size:0.85em;color:#6b7280″>Source: Synergy Research Group; Microsoft FY27 Q1 release; CoreWeave Q2 2026 results; JLL; Meta Q3 2026 release; CNBC/S&P Global; Bloomberg via Tiger Brokers; Dave Friedman analysis.

A sudden cut to capex guidance from one of the big four would be the clearest warning. Rising vacancy in core markets such as Northern Virginia, Dallas or Frankfurt, falling rental rates for recent-generation GPUs, or widening spreads on data center ABS would also signal stress. Conversely, continued revenue growth above 40–50% at the cloud platforms and steady AI lab revenue gains would support the bull case.

Frequently Asked Questions