Artificial intelligence was supposed to be a software revolution.
Instead, it is rapidly becoming one of the largest infrastructure projects in modern economic history.
Behind ChatGPT, Gemini, Claude, Copilot and the hundreds of AI services appearing across the internet sits an increasingly physical economy made of GPUs, data centers, electrical substations, cooling systems, fiber networks and power contracts.
And all of it has to be financed.
A Nikkei Asia analysis published in July estimated that five major US technology companies — Alphabet, Microsoft, Amazon, Meta and Oracle — had accumulated approximately $1.65 trillion in off-balance-sheet commitments, up roughly eightfold in four years. Nikkei compared that figure with around $1.35 trillion of debt already recognized on their balance sheets.
The headline number is spectacular.
It is also easy to misunderstand.
Big Tech does not literally have another $1.65 trillion of secret bank loans hidden somewhere in its accounts. Much of the figure consists of long-term leases, data center contracts, GPU and infrastructure purchasing commitments and other contractual obligations that accounting rules do not necessarily require companies to recognize as conventional debt today.
But dismissing those commitments because they are not technically classified as debt would be equally misleading.
They represent something very real:
hundreds of billions of dollars that some of the world’s largest technology companies have already promised to spend on the infrastructure required to keep the AI boom running.
And the numbers are still growing.
The $1.65 Trillion Number May Already Be Outdated
The most remarkable part of the story is that the Nikkei estimate was based on financial information available before some of the latest corporate filings.
Those newer filings show how quickly commitments are expanding.
Microsoft disclosed that as of June 30, 2026 it had $329.1 billion of additional leases, primarily for data centers, that had not yet commenced. Those agreements are expected to begin between fiscal 2027 and 2033 and can run for as long as 20 years.
Microsoft also reported roughly $194 billion in purchase commitments, primarily related to data centers, including purchase orders and take-or-pay agreements.
Meta’s numbers are equally striking.
As of June 30, the company reported approximately $278.99 billion in operating and finance leases that had not yet commenced, largely involving data centers, colocation facilities and network infrastructure.
Then, in July, Meta signed another $68 billion of data center leases, expected to commence in 2027 and 2028.
Separately, Meta disclosed approximately $349.31 billion in non-cancelable contractual commitments, largely related to cloud capacity, servers, networking infrastructure and data centers.
Oracle may be the most extreme case.
Its latest annual filing shows $260 billion of additional lease commitments, almost entirely associated with data centers. Those contracts generally run for 15 to 19 years and had not yet appeared on Oracle’s balance sheet as lease liabilities because the leases had not commenced.
Alphabet reported another $85.2 billion of future lease payments, primarily connected with data centers, under leases that had not yet begun as of June 30.
These numbers cannot simply be added together — companies report commitments differently and several categories have different accounting definitions.
But they reveal the scale of what is happening.
AI is no longer being financed purely through annual technology budgets.
The industry is signing contracts extending well into the 2030s and 2040s.
“Hidden Debt” Isn’t Quite Debt — But It Isn’t Imaginary Either
The phrase hidden debt is effective.
Accounting purists will understandably dislike it.
Imagine a technology company signs a contract today to lease a data center beginning in 2028 for the following 20 years.
Until certain conditions are met and the lease officially commences, that future obligation may not appear on the balance sheet in exactly the same way as a bond or bank loan.
Yet economically the company has still made a significant commitment.
The same principle applies to arrangements such as:
- long-term data center leases;
- GPU supply agreements;
- cloud capacity contracts;
- power purchase agreements;
- take-or-pay infrastructure contracts;
- construction commitments;
- project financing structures.
This is why looking exclusively at traditional corporate debt provides an incomplete picture of the AI infrastructure race.
The obligation may have a different accounting treatment.
The cash will still eventually have to come from somewhere.
AI Has Turned Big Tech Into Heavy Industry
This is perhaps the biggest structural change produced by generative AI.
Traditional software economics are extraordinary because distributing another copy of a piece of software costs almost nothing.
AI economics are different.
Every sufficiently large AI service requires enormous amounts of computing capacity.
That requires chips.
The chips require servers.
The servers require data centers.
The data centers require cooling.
And the entire system requires extraordinary amounts of electricity.
The consequence is that companies famous for asset-light, high-margin digital businesses are beginning to behave more like infrastructure companies.
J.P. Morgan estimates that capital expenditure among the five largest US hyperscalers could reach approximately $697 billion in 2026 alone, an increase of $173 billion from estimates made earlier in the year.
For perspective, that is not the total investment expected over an entire decade.
That is one year.
And the infrastructure race is moving beyond corporate cash reserves.
Technology companies are increasingly tapping public bond markets, project finance, private credit and leasing structures to keep construction moving.
More than $220 billion of AI-related debt issuance had already reached the market during 2026 by late August, according to data reported by Reuters.
AI is therefore becoming something unusual in the history of Silicon Valley:
a technology revolution whose success increasingly depends on the plumbing of global capital markets.
Big Tech Is Even Borrowing From Europe
The financing requirements have become so large that American technology companies are increasingly issuing debt outside the United States.
Reuters reported on September 3 that US hyperscalers have issued around €40 billion in euro-denominated bonds.
Big Tech now represents close to 10% of new non-financial corporate bond issuance denominated in euros, according to figures cited by Reuters from the European Central Bank.
That creates an interesting second-order effect.
The AI infrastructure boom may eventually influence borrowing costs for companies that have absolutely nothing to do with AI.
Investors only have a finite amount of capital to allocate.
If highly rated US technology companies absorb a growing percentage of European bond demand, governments and European corporations could eventually have to offer higher yields to attract investors.
The AI race is therefore no longer just affecting semiconductor prices or electricity demand.
It may increasingly affect the price of capital itself.
So Where Is the Bubble Risk?
None of this automatically proves that AI is a bubble.
But it does create several vulnerabilities.
1. The infrastructure lasts longer than the technology
Oracle’s future data center leases can last 15 to 19 years.
Meta has contracts extending as far as 30 years.
Microsoft has committed to leases that can extend for two decades.
That creates an obvious question.
What does the AI market look like in 2040?
Nobody knows.
A data center building may remain useful for decades, but the hardware inside it does not.
GPU generations move extremely quickly.
The infrastructure industry is therefore making very long financial commitments around technology that can become obsolete remarkably quickly.
2. AI demand could be overestimated
The biggest danger is not that artificial intelligence disappears.
It almost certainly won’t.
The danger is that companies collectively build more capacity than customers are eventually willing to pay for.
That distinction matters.
Railways transformed the nineteenth-century economy.
Railway investors still experienced spectacular financial bubbles.
Telecommunications transformed society.
The telecom infrastructure boom still produced enormous bankruptcies.
The internet changed civilization.
The dot-com bubble still happened.
A technology can be revolutionary and simultaneously attract too much capital.
3. Data center demand may contain duplication
Another warning sign is emerging from electricity grids.
Reuters recently reported concerns in Texas over so-called “ghost” data center demand, where multiple developers can effectively request electrical capacity for projects that may overlap, be speculative or never ultimately get built.
The uncertainty has become serious enough to influence how utilities and grid operators evaluate new connections.
If projected electricity demand exaggerates the number of data centers that will actually operate, infrastructure forecasts could be distorted.
And if infrastructure forecasts are distorted, financing assumptions can follow.
4. The business model requires enormous utilization
A GPU sitting idle is expensive.
A half-empty data center is even worse.
The economics of AI infrastructure depend heavily on maintaining high utilization while charging customers enough to cover:
- electricity;
- hardware depreciation;
- financing costs;
- cooling;
- networking;
- personnel;
- maintenance;
- property;
- future hardware upgrades.
That is relatively easy when computing capacity is scarce.
It becomes harder if the industry eventually moves from shortage to oversupply.
Oracle Shows What Happens When the Balance Sheet Gets Stretched
The five hyperscalers should not be treated as if they face identical financial risks.
Alphabet, Microsoft, Amazon and Meta remain extraordinarily profitable companies with substantial cash-generation capacity.
Oracle is different.
Reuters reported in August that Oracle carried approximately $129.5 billion in debt, while leverage had climbed to roughly 4.3 times EBITDA.
S&P downgraded Oracle to BBB-, just one notch above speculative-grade status.
At the same time, Oracle is pursuing approximately $95 billion in capital expenditure during fiscal 2027 while supporting enormous long-term data center commitments.
The company also has an enormous backlog that could ultimately justify the spending.
But Oracle demonstrates precisely why investors are beginning to look beyond revenue growth.
The question is increasingly:
Who is carrying the financial risk if AI infrastructure demand fails to develop exactly as expected?
Why This Is Not the Dot-Com Bubble — At Least Not Yet
There is another side to the argument.
Comparing every technology boom to 2000 is intellectually lazy.
Microsoft, Alphabet, Amazon and Meta are not Pets.com.
They are among the most profitable corporations ever created.
Their cloud businesses already generate huge revenues.
Microsoft recently disclosed that Azure alone generated $101.9 billion in revenue during the fiscal year ending June 30, 2026.
Demand for AI computing capacity also remains extremely strong.
Companies are still struggling to secure power, GPUs and suitable data center locations.
J.P. Morgan argues that despite the extraordinary scale of financing, credit investors continue to perform disciplined risk analysis rather than indiscriminately funding every project.
Those are important differences from classic speculative bubbles.
The AI boom is producing real infrastructure for real products with real customers.
The uncomfortable question is whether the amount being spent will produce an adequate return on capital.
That’s a much harder question.
The Real AI Bubble Test
Forget Nvidia’s share price for a moment.
Forget ChatGPT user numbers.
Forget benchmark scores.
The most important number in AI may eventually become utilization.
If trillions of dollars of infrastructure are built and that capacity remains heavily utilized, today’s extraordinary investment boom could eventually look rational.
AI revenues could grow into the infrastructure.
Productivity could rise.
Data centers could become the factories of a new digital economy.
But if compute becomes abundant, model efficiency improves faster than expected, customers resist rising AI prices or enterprise adoption disappoints, the economics can change quickly.
The contractual obligations will not disappear with them.
Companies could find themselves paying for 15- or 20-year facilities designed around demand forecasts made during the most aggressive phase of the AI boom.
That is when today’s off-balance-sheet commitments become tomorrow’s very visible problem.
The AI Bubble May Not Be Where Everyone Is Looking
There is a tendency to frame the AI bubble debate around one question:
Are AI companies overvalued?
That might be the wrong question.
The more interesting risk could be hiding deeper inside the infrastructure stack.
What if AI itself succeeds, but investors dramatically overestimate how much computing infrastructure is required to deliver it?
That scenario would not mean artificial intelligence had failed.
It would mean the capital cycle had overshot the technology cycle.
History has seen this pattern before.
Railroads were transformative.
Fiber-optic networks were transformative.
The internet was transformative.
Every one of them experienced periods where infrastructure investment ran substantially ahead of economically sustainable demand.
AI could do the same.
A $1.65 Trillion Warning, Not a $1.65 Trillion Time Bomb
Calling the entire $1.65 trillion figure “debt” would therefore be misleading.
Calling it irrelevant because much of it sits outside conventional debt accounting would be even worse.
The number should instead be interpreted as a measure of conviction.
Big Tech is effectively making one of the largest financial bets in corporate history:
that global demand for artificial intelligence will grow fast enough, for long enough, to justify an infrastructure system costing trillions of dollars.
Perhaps it will.
There are compelling reasons to believe AI will become a foundational technology.
But revolutionary technologies do not repeal the laws of finance.
At some point, every GPU, every data center and every megawatt of electricity must generate an economic return.
That may ultimately be the defining question of the AI boom.
Not whether artificial intelligence works.
But whether it can possibly become profitable enough to justify everything being built around it.
