INTRODUCTION
This article is Part III of a three-part series on AI Data Centers. These articles are a response to a reader who asked these three questions:
1. Which public issues are impacting AI Data Centers?
2. A few of us believe their expansion is limited by physics. Is that so?
3. And how long will this last? Is an AI bubble forming?
In response, I am answering in three parts, as follows:
· Part III: AI Bubble
COMPLIANCE SOLUTIONS
AI POLICY PROGRAM FOR MORTGAGE BANKING™
Policies and procedures for artificial intelligence related to mortgage banking compliance administration, mortgage loan origination compliance, mortgage servicing compliance, and industry best practices—including, but not limited to, applicable federal and state banking and consumer lending laws. Currently, it consists of nine policies, with more planned.
AI FOR MORTGAGE LOAN ORIGINATION - SALES MANUAL
The Sales Manual provides 9 chapters covering the full AI-powered origination funnel in a 90-day implementation roadmap, with clear team roles and fair lending, FCRA, and advertising compliance guidance built into every chapter. It includes sample AI prompts, an ROI tracking framework, and a compliance glossary.
RESPONSE
AI BUBBLE: Promises, Panics, and Economics
I want to commence this discussion with probably the most famous quote describing the Dutch tulip bubble (also known as the Tulip Mania). The quote comes from Scottish journalist Charles Mackay in his 1841 book, Extraordinary Popular Delusions and the Madness of Crowds:
"A golden bait
hung temptingly out before the people, and one after the other,
they rushed to the tulip-marts, like flies around a honey-pot."
Parts of this article are technical. If that's not for you, skip ahead to the What to Watch and Conclusion sections at the end.
Are data centers a symptom of an AI bubble?
Short answer: data centers show the classic financing signs of a bubble – spending racing ahead of revenue, a shift from cash to debt, off-balance-sheet structures, and circular deals – but not yet the classic demand sign of one, because the capacity being built is still being used. In layperson's terms, the industry is building the plane while flying it.
Or, a better metaphor: this is a case where the grasp exceeds the reach!
The poet Robert Browning wrote, "a man's reach should exceed his grasp" – we should aim beyond what we can already hold. Here, it is the reverse. The industry has already grabbed – committed, signed, and borrowed for – more than its revenue can reach today.
So, whether this becomes a bubble depends on one question: does AI revenue close the gap before the debt comes due?
The spending has outrun the spenders' own cash
The five hyperscalers – Microsoft, Amazon, Alphabet, Meta, and Oracle – are the giant cloud companies building most of this capacity. Their capital expenditure (capex), meaning spending on buildings, chips, and equipment, rose from about $256 billion in 2024 to $443 billion in 2025, and is projected at $602 billion in 2026. Later estimates run higher: the Swiss private bank Union Bancaire PrivĂ©e (UBP) puts 2026 near $820 billion and 2027 at $1.0–1.3 trillion.
Five years ago, these firms spent 10–15% of their revenue on capex. Today, Amazon spends about a quarter, Alphabet and Microsoft around half, Meta close to 60%, and Oracle more than 80%.
The problem: their operating cash flow no longer covers it. The grasp exceeds the reach!
The build has moved into the credit markets
Hyperscaler bond issuance went from about $17 billion in 2024 to $109 billion in 2025, and $194 billion in the first half of 2026 alone. Across the wider AI industry, Goldman Sachs tracks nearly $500 billion of AI-related debt issued this year. Morgan Stanley sees a roughly $1.5 trillion gap through 2028 between data center spending and what the hyperscalers can fund themselves. Private credit, securitizations, and project finance are filling it.
Much of the commitment doesn't show up as debt. According to Axis Intelligence Research, in its article "The Five Largest Tech Companies Now Owe Nearly $1 Trillion in Data Center Lease Commitments," the five hyperscalers
"held $969 billion in total undiscounted future data center lease commitments as of year-end 2025, of which $662 billion represented leases not yet commenced and therefore sitting entirely off their reported balance sheets. That $662 billion figure equals 113% of the same five companies' combined adjusted on-balance-sheet debt – meaning their shadow infrastructure obligations now exceed what their own credit profiles show investors."
In plain English: the leases they have signed but not yet started are larger than all the debt they report.
Markets are starting to price the risk
By September 2026, bonds of AI-related companies paid about 1.15 percentage points (115 basis points) more than U.S. Treasuries, versus about 0.78 points for investment-grade companies generally. A wider gap means lenders see more risk. Apollo Global also warned that credit default swaps – the cost of insuring hyperscaler bonds against default – are getting pricier for reasons beyond routine hedging.
Cash flow tells the same story. Oracle's free cash flow was minus $24 billion in its 2026 fiscal year, and UBP expects several hyperscalers to report negative free cash flow next year, with capex above $1 trillion against about $900 billion of operating cash flow. Microsoft is the only one still funding its build-out from its own cash flow.
Let me be clear about what that means. Negative free cash flow means a company spends more on its operations and investments than it brings in, so it must cover the shortfall from cash reserves, new debt, or new shares.
As UBP puts it, their "capital spending is expected to exceed operating cash flow, with investment projected to reach $5.6 trillion by 2030, raising fundamental questions about how this unprecedented expansion will be funded."
Much of the leverage is hard to see
Leverage simply means borrowing. Meta, for example, put data center debt in a special purpose vehicle (SPV) – a separate company that borrows the money and is repaid from Meta's long-term lease. The debt stays off Meta's balance sheet.
The four largest cloud providers also report about $2.3 trillion of contracted future revenue (their backlog). Roughly 40% of it is owed by two private AI labs, OpenAI and Anthropic, and much of it is for capacity not yet built.
Some of the money goes in circles
Here is how the circle works: Nvidia invests in AI labs; the labs commit to buy computing power; and the cloud providers buy Nvidia chips to supply it. In August 2026, Nvidia even disclosed that it would guarantee OpenAI's lease payments on a roughly 4.25 GW campus, with OpenAI agreeing to repay Nvidia for anything it pays. We know this from Nvidia's Form 8-K filing.
Critics call this "round-tripping"; defenders call it "ordinary vendor financing."
The real risk is timing. According to UBP, lease and contract payments start on a schedule centered on 2027–28, whether or not AI lab revenue has caught up. When suppliers are also investors in and guarantors of the same customers, true demand is harder to see. Even a slowdown in growth – not an outright decline – could strain the most heavily borrowed projects.
Depreciation is another concern. Many operators write off GPUs over five to six years, while new generations arrive every year or two. If the chips lose value faster than the books assume, today's reported earnings look better than they really are.
The case against an "AI bubble"
So, do these concerns prove there is a bubble? Not yet.
Unlike the 1990s fiber-optic glut, when miles of cable sat unused, today's capacity is largely leased before it is even switched on. Customers have committed to pay for it before the power is connected.
In Part II, I showed that physical limits actually restrain overbuilding. And demand is real: Nvidia reported $89.0 billion of data center revenue in its July 2026 quarter, up 117% from a year earlier. AI lab revenue is compounding fast; Anthropic, for example, recently turned profitable, according to UBP. And the biggest spenders are among the most profitable companies in history. So, a correction would more likely hit stock prices and heavily borrowed developers than cause a banking crisis.
Consider this table, which lists my bubble and non-bubble readings.
Bubble versus Non-Bubble
|
Signal |
Bubble reading |
Non-bubble reading |
|
Capital expenditures above cash flow |
Spending on faith, funded by debt |
Normal for an infrastructure build-out, like railroads or telecom |
|
Debt and SPVs (special purpose vehicles) |
Hidden leverage, like 2000s telecom or 2007 SIVs |
Leases backed by investment-grade tenants |
|
Circular deals |
Revenue that is really the seller's own money |
Vendor financing to secure scarce supply |
|
Customer concentration |
Two private labs, one still burning cash, owe about 40% of the backlog |
AI lab revenue is growing several-fold a year |
|
Utilization |
— |
Capacity is "sold out": power – not demand – is the
constraint |
As best as I can determine, data centers are less a symptom of a bubble than the place where a bubble would do its damage. If AI revenue keeps compounding, today's debt will look like prudent infrastructure finance. If it plateaus, the stranded assets are concrete, turbines, and grid upgrades – and, as I discussed in Part I, ratepayers may already be carrying part of that bill.
What to watch
The three questions of this series resolve together: physics sets the pace, the public sets the price of permission, and the credit markets decide who absorbs the loss if revenue disappoints. Five indicators will tell the story over the next 12–18 months:
1. Large-load tariffs: Whether states make data centers pay the full cost of their grid upgrades, which would shift cost from households to developers and slow marginal projects.
2. Energized versus announced gigawatts: The gap between capacity announced and capacity actually connected is the truest measure of the physical bottleneck.
3. AI credit spreads and credit default swaps: Continued widening would mean lenders are repricing the risk before stock investors do.
4. Lab revenue versus compute commitments: The backlog is only as good as the two or three customers that owe most of it.
5. Midterm politics: Data center siting and power bills are now ballot issues, and the November 3 elections will show how much public permission the industry still has.
CONCLUSION
So, how long will this last, and is an AI bubble forming?
My answer: a bubble is forming in the financing, but not yet in the demand.
Recall Mackay's tulips. A tulip bulb produced nothing but a flower, and when buyers stopped showing up in 1637, prices collapsed within weeks. AI Data Centers differ in one important way: the computing power they produce is being used, and paid for, as fast as it comes online. That is why I don't call this a classic bubble today.
But the way it is being paid for has the features of one: borrowed money, hidden leases, circular deals, and a handful of customers – two of them private labs – carrying much of the risk. The grasp exceeds the reach. Bubbles rarely burst when the excitement fades; they burst when the bills come due.
How long will this last? I see three stages:
· Through 2026 and into 2027, the build-out continues. Contracts are signed, and capacity is sold out. As I explained in Part II, physics and the grid slow construction more than any money shortage does.
· In 2027–2028 comes the test. Lease and contract payments start, and the AI labs must pay from their own revenue rather than from fresh investment. This is when we learn whether the reach can catch up with the grasp.
· After 2028, there are two paths. If AI revenue keeps compounding, today's spending will look like the railroads or the power grid: costly, but foundational. If revenue growth stalls, expect a painful shakeout – falling stock prices, defaults among heavily borrowed developers, and half-finished campuses – though probably not a banking crisis, because the biggest spenders can absorb large losses.
In short, the AI build-out will last as long as AI revenue keeps pace with the debt, and the next two years will tell us whether it can. Watch the five indicators above. If credit spreads keep widening while lab revenue falls behind its commitments, the bubble is real – and, like the tulip mania of 1637, the crowd will leave much faster than it arrived.
___________________________
This article, AI Data Centers: AI Bubble – Part III, published on October 15, 2026, is authored by Jonathan Foxx, PhD, MBA, the Chairman & Managing Director of Lenders Compliance Group, founded in 2006, the first and only full-service mortgage risk management firm in the United States, specializing exclusively in residential mortgage compliance.
