JPMorgan has not predicted that the AI bubble will burst in the next few months.
What the bank is saying is more complicated—and arguably more important.
One of JPMorgan’s technical strategists has identified a divergence in the AI trade that resembles a pattern seen around the end of the dot-com boom: companies selling the infrastructure are continuing to outperform even as some of the companies spending enormous amounts of money to buy that infrastructure struggle.
At the same time, other parts of JPMorgan remain bullish on artificial intelligence, are helping arrange billions of dollars in AI infrastructure financing, and project that trillions more may be spent before the decade is over.
There is no contradiction in saying AI could transform the economy while also questioning whether investors are spending too much money, too quickly, at valuations that assume an extraordinary amount of future profit.
The internet survived the dot-com bubble.
Amazon survived.
Fiber-optic networks survived.
The technology was real.
A tremendous amount of the capital invested around it still disappeared.
That is the distinction worth understanding as Wall Street begins asking harder questions about AI.
What Did JPMorgan Actually Warn About?
The viral version of this story is something like: JPMorgan says AI looks exactly like the dot-com bubble and investors should get out before it crashes.
That overstates what happened.
JPMorgan technical strategist Jason Hunter has been warning about several technical conditions in the U.S. stock market ahead of September. He said major indexes were still trending bullishly and had not produced an unambiguous sell signal, but recommended investors be prepared to reduce long exposure if the deterioration continued.
One of the patterns behind that concern involves AI.
JPMorgan Private Bank has separately discussed Hunter’s analysis comparing today’s AI market with the late-1990s communications boom.
Near the end of that earlier cycle, communications-equipment companies continued climbing even as communications-service companies—the businesses purchasing much of that equipment—began flattening or declining.
JPMorgan sees a partial parallel today.
Its hyperscaler basket, including companies such as Microsoft, Amazon, Alphabet, Meta and Oracle, has lagged the Philadelphia Semiconductor Index while the semiconductor and infrastructure side of the AI trade continued climbing.
JPMorgan Private Bank described hyperscaler stock prices as stagnating and their free cash flow as declining while semiconductor, optical-networking and related infrastructure providers continued appreciating.
That is the warning.
It is not a prediction of the date the AI bubble will burst.
And JPMorgan as an institution certainly has not abandoned the market. On August 10, the bank raised its 2026 year-end S&P 500 target from 7,800 to 8,000, specifically citing stronger earnings and increasing confidence that AI investment could translate into greater revenue growth.
The interesting story is therefore not that JPMorgan suddenly believes AI is doomed.
It is that one of the financial institutions most deeply involved in the AI investment boom is now documenting stresses inside that boom.
The Companies Selling AI Infrastructure Are Winning. The Companies Buying It Still Have to Prove the Economics.
Think about the current AI economy as a chain.
Nvidia and other semiconductor companies sell chips.
Equipment manufacturers sell networking, memory, cooling and electrical systems.
Construction companies build data centers.
Banks arrange financing.
Utilities build or expand power infrastructure.
Data-center operators lease capacity.
Then, eventually, the businesses purchasing all of this infrastructure must earn enough additional money from AI to justify what everybody upstream has already been paid.
That distinction matters.
Selling $10 billion worth of shovels proves there is enormous demand for shovels.
It does not prove the people buying the shovels will find $20 billion worth of gold.
Goldman Sachs Research has independently identified essentially this same problem.
In June, Goldman research chief Jim Covello argued that semiconductor companies have been capturing enormous economic benefits while companies higher in the AI chain absorb the cost.
Normally, chip companies thrive because their customers are thriving. In this cycle, Covello argues, the semiconductor companies have in some cases been thriving economically at the expense of their customers.
Eventually, something has to change: either the companies buying AI infrastructure begin producing much greater profits from it, or they eventually reduce what they spend on that infrastructure.
That may be the simplest way to understand the entire AI bubble debate.
The question isn’t whether people want GPUs.
Clearly they do.
The question is whether the trillions of dollars flowing into AI infrastructure ultimately create enough additional profit downstream to justify the trillions flowing upstream.
AI Can Be Revolutionary and Still Be a Bubble
One of the worst arguments in the AI debate is that the technology cannot be a bubble because AI is genuinely useful.
That misunderstands what a bubble is.
A financial bubble does not require the underlying technology to be fraudulent.
It does not require zero revenue.
It does not require every company involved to fail.
And it certainly does not require the technology to disappear afterward.
The dot-com bubble provides the obvious example.
The central technological prediction behind the boom—that the internet would radically transform commerce, media, communication and everyday life—was spectacularly correct.
The financial conclusions investors drew from that prediction were often spectacularly wrong.
Companies were valued as though technological transformation automatically meant unlimited profits.
Telecommunications infrastructure was built far ahead of economically sustainable demand.
Companies that eventually became enormously successful still suffered staggering market losses.
The internet won.
That did not mean every investor financing the internet won with it.
The same possibility exists with artificial intelligence.
AI may become one of the most important general-purpose technologies developed in generations.
That tells us remarkably little about whether a particular AI company deserves a $300 billion valuation, whether another $50 billion data center makes economic sense, or whether the next trillion dollars of infrastructure investment will earn an adequate return.
Technology and valuation are different questions.
The AI Industry Has a $650 Billion Revenue Question
JPMorgan itself has attempted to quantify the hurdle.
In January, JPMorgan Asset Management estimated that achieving a 10% return on then-current AI investment could require approximately $650 billion in annual AI revenue.
Its conclusion was not that AI monetization did not exist.
Quite the opposite.
JPMorgan found that monetization was already occurring, particularly in infrastructure and cloud services.
But it remained much less established at the application layer, where businesses must ultimately demonstrate that AI either creates additional revenue or reduces costs sufficiently to justify what they are paying for it.
The spending has only continued accelerating.
JPMorgan now says hyperscaler capital expenditure surpassed $400 billion in 2025 and is approaching $800 billion in 2026—roughly ten times the level of 2019.
That spending could absorb approximately 90% to 100% of hyperscaler operating cash flow this year, compared with around 60% in 2025.
That does not mean the investment will fail.
There is substantial evidence supporting the bull case.
JPMorgan says more than 70% of data-center capacity currently under construction is already leased. Major cloud platforms averaged approximately 44% year-over-year revenue growth during the first quarter of 2026. The four largest hyperscalers entered the year with extraordinarily strong balance sheets.
Those facts make today’s market importantly different from the worst excesses of the late-1990s telecommunications buildout.
They do not eliminate the return problem.
Demand can be real and still be overpriced.
Follow the Money: Who Gets Paid Before AI Proves Itself?
This is where the current boom becomes especially interesting.
The financial rewards arrive at different times for different participants.
A chipmaker generally recognizes revenue when its products are sold.
A construction company gets paid to construct the facility.
A lender earns interest.
An investment bank can earn fees for arranging financing.
A developer can monetize a long-term lease.
But the ultimate economic justification for the entire chain may take years to establish.
That creates an important asymmetry.
The people financing and supplying an infrastructure boom can make substantial money during the boom without guaranteeing that every project they facilitate will ultimately generate an adequate return for whoever owns it a decade later.
JPMorgan offers an unusually clear window into this system because it is operating on multiple sides of it.
On August 10, JPMorgan’s investment-banking division called AI infrastructure financing one of the defining capital-deployment themes of the current market and estimated that five major hyperscalers alone would spend approximately $697 billion during 2026.
The bank has also participated directly in some enormous transactions.
JPMorgan says it originated $9.6 billion across two construction loans for the Stargate AI infrastructure campus in Abilene, Texas.
It acted as lead-left bookrunner on a $4.25 billion bond offering financing Hut 8’s Beacon Point data center.
And in April, it served as lead-left and active bookrunner on CoreWeave’s $5.25 billion combined debt and convertible-note offering.
There is no evidence that JPMorgan is secretly financing projects it knows will fail so it can dump them on unsuspecting investors.
That claim would go well beyond the evidence.
But the incentive structure is worth examining without inventing a conspiracy.
Wall Street does not need the AI boom to produce perfect long-term capital allocation in order to make money from the process of allocating the capital.
Loans must be arranged.
Bonds must be sold.
Companies raise equity.
Projects get structured.
Assets change hands.
Risks get hedged.
Money moving through the financial system creates business for financial institutions.
Whether the ultimate project earns the return originally imagined is a different question.
AI’s Next Phase Is Increasingly Being Financed With Debt
This may be the most important change in the AI boom.
For its first several years, much of the infrastructure spending could be financed from the enormous cash flows generated by companies such as Microsoft, Alphabet, Amazon and Meta.
That cushion is becoming less sufficient as spending accelerates.
JPMorgan Asset Management said in August that the AI capital-expenditure cycle has begun migrating “from balance sheet sourced to capital markets” because hyperscaler free cash flow can no longer finance the entire buildout.
The numbers are striking.
Hyperscaler corporate bond issuance was approximately:
- $17 billion in 2024
- $109 billion in 2025
- $194 billion in the first six months of 2026
JPMorgan estimates the broader data-center buildout could require roughly $5 trillion through 2030, with around $2 trillion financed through investment-grade credit markets.
Another JPMorgan analysis projects that cumulative investment-grade issuance associated with the buildout could reach approximately $2.1 trillion by 2030 while helping fund an estimated $5.5 trillion in AI-related capital expenditure.
This does not mean an AI debt crisis is imminent.
Microsoft, Alphabet, Amazon and Meta are not dot-com startups borrowing money with no functioning businesses behind them.
They are among the largest and most profitable corporations in history.
But debt changes the nature of the risk.
Capital spending funded from excess cash can be reduced when management decides the return no longer looks attractive.
Debt creates contractual obligations.
And as more AI financing enters bond markets, securitization, private credit and infrastructure funds, exposure to the AI capital cycle spreads beyond people consciously purchasing AI stocks.
That is how a technology boom gradually becomes a financial-system story.
Your Retirement Account May Already Have Significant AI Exposure
This does not mean people should panic and liquidate their retirement accounts.
That would be replacing one unsupported prediction with another.
It does mean investors should understand what they own.
As of July 31, information technology alone represented 36.8% of the S&P 500.
Nvidia was the index’s largest constituent, followed near the top by companies including Microsoft, Amazon, Alphabet, Broadcom and Meta—all major beneficiaries, purchasers or participants in the AI buildout.
So someone who has never deliberately purchased an “AI stock” can still have substantial indirect exposure through a broad index fund.
The same phenomenon is beginning to appear in bonds.
JPMorgan warns that AI financing is creating a new form of concentration in investment-grade debt indexes. Unlike stock indexes, where larger market capitalizations generally produce larger weights, bond indexes can give larger weights to companies simply because they have issued more debt.
As the largest AI spenders issue increasingly large quantities of bonds, passive bond investors acquire more exposure to them automatically.
The rational conclusion is not “sell everything connected to AI.”
It is much simpler:
Know your concentration.
A diversified retirement portfolio and a concentrated bet on speculative AI companies are not the same thing. But neither should investors assume that an index fund means they have no meaningful exposure to one extraordinarily dominant investment theme.
The Bigger Question for Communities: Who Pays If the Data-Center Forecasts Are Wrong?
The AI boom is no longer confined to brokerage accounts.
It is reshaping power grids, municipal development policy, tax incentives and land-use decisions across the United States.
That makes the data-center side of the boom particularly important.
A data center can remain economically useful even if AI stocks fall.
And today’s market is not yet showing the widespread unused capacity that characterized the telecom crash.
JPMorgan says more than 70% of data-center capacity under construction is already leased, while another analysis put pre-leasing even higher in major markets.
But data centers create unusually long-lived commitments.
Utilities may build generation, transmission and substations.
Governments may approve tax exemptions.
Roads and water systems may be expanded.
Communities may make planning decisions around projected employment and tax revenue.
Developers may sign financing agreements extending 10 or 15 years into the future.
The right question therefore isn’t:
Will every AI data center become abandoned after the bubble bursts?
There is no evidence for that.
The better question is:
If today’s demand projections are wrong, who has contractually agreed to absorb the loss?
Is it the developer?
The lender?
The hyperscaler?
The utility?
Existing electricity customers?
Municipal taxpayers?
Or some combination of them?
Federal regulators are already treating that question seriously. In June, the Federal Energy Regulatory Commission ordered all six regional grid operators under its jurisdiction to justify or reform rules governing the connection of data centers and other massive new electricity loads, explicitly pairing faster connections with protections for existing ratepayers.
That concern would make little sense if the distribution of infrastructure risk were purely hypothetical.
Governments Are Already Reconsidering the Data-Center Bargain
At least 38 states offer tax incentives specifically targeting data centers, according to a 2026 review by Washington state’s Joint Legislative Audit and Review Committee.
Washington provides a useful example of why those agreements deserve scrutiny.
Its legislative auditor found that beneficiaries of one urban data-center tax preference received an estimated $42.4 million in tax savings between 2023 and 2026.
Those beneficiaries reported creating 53 family-wage jobs and supporting nearly 300 temporary construction jobs.
But the auditor also found that all of the participating data centers predated the preference, making it impossible to determine whether the incentive actually caused the economic activity it was designed to encourage. No new urban data center had been built using the preference.
That does not mean the projects produced no public benefit. The same review found increased property-tax payments following refurbishment and likely increases in public-utility taxes.
It means the government could not establish how much of that activity taxpayers actually needed to subsidize.
Fulton County, Georgia, went further this month.
Its Board of Commissioners formally opposed additional local tax abatements and other financial incentives for data centers.
The county said state incentives were estimated to cost Georgia county and municipal governments $1.1 billion in sales-tax revenue in 2026, while its local development authority had granted approximately $150 million in data-center tax abatements since 2020.
Again, this does not prove data centers are bad investments.
It demonstrates why communities should treat an AI development proposal as an infrastructure agreement—not as a referendum on whether artificial intelligence is exciting.
The relevant questions are contractual:
What is being subsidized?
How many permanent jobs are guaranteed?
Who pays for new electrical infrastructure?
Are minimum power payments required if demand disappears?
What happens if the tenant leaves?
Are there clawbacks if employment promises are missed?
Who owns stranded infrastructure?
And how long are local taxpayers or ratepayers exposed?
Those questions still matter if AI becomes the most important technology of the century.
What Would an Actual AI Bubble Burst Look Like?
A bubble does not usually announce itself with one dramatic event.
If the AI capital cycle is becoming unsustainable, there are several developments that would matter far more than another viral prediction that Nvidia is about to collapse.
Hyperscalers stop being rewarded for spending more
During the expansionary phase of an investment cycle, announcing another enormous capital budget can signal confidence and growth.
Later, the same announcement can begin frightening shareholders because investors start asking when the spending will generate cash.
That change in market reaction matters.
AI revenue fails to catch capital expenditure
Cloud revenue and AI subscriptions are growing rapidly today.
The bearish thesis becomes more credible if infrastructure spending continues accelerating while incremental AI revenue, margins and free cash flow cannot keep pace.
Hyperscalers cut capital-expenditure forecasts
This would be one of the clearest signals.
A reduction in spending by Microsoft, Amazon, Alphabet or Meta would flow directly into chips, memory, networking, cooling, power equipment and data-center construction.
That is exactly why the current divergence between AI suppliers and their customers deserves attention.
Semiconductor demand weakens
Goldman’s Covello essentially describes this as the eventual economic test: companies above the semiconductor layer either begin earning enough money from AI to justify the spending, or semiconductor spending eventually has to adjust.
Financing becomes more expensive
As increasingly large portions of the AI buildout move into debt markets, widening credit spreads, weaker demand for project debt or difficulty refinancing highly leveraged facilities would become important warning signs.
Marginal data-center projects get cancelled
Projects without firm tenants, secure power, favorable financing or compelling economics would likely be the first casualties—not necessarily the giant campuses operated by the strongest companies.
Only after several of those conditions begin appearing together would the argument for a genuine capital-cycle reversal become substantially stronger.
What Would Prove the AI Bubble Skeptics Wrong?
A serious analysis also needs a way to be wrong.
The bubble thesis weakens considerably if AI revenue grows rapidly enough to catch the spending.
If enterprise customers begin demonstrating large and repeatable productivity gains, businesses will have stronger reasons to keep buying AI services.
If hyperscaler free cash flow recovers despite enormous capital expenditures, concern over the spending burden declines.
If inference costs continue falling while total usage rises rapidly enough to offset those declines, AI providers could generate far better economics.
If data-center utilization remains high as supply expands, fears of a telecom-style infrastructure glut become less compelling.
And if companies can slow capital-expenditure growth without crushing semiconductor and infrastructure earnings, the industry may simply transition from a frantic construction phase into a more mature growth cycle.
There are already reasons for bulls to believe this can happen.
Goldman Sachs, while publishing Covello’s skepticism, has other researchers projecting enormous growth in AI-agent usage and improving hyperscaler economics as inference becomes cheaper and enterprise adoption expands.
JPMorgan likewise says current demand for compute is exceeding supply and emphasizes that today’s hyperscalers have real revenue, enormous cash flows and much stronger balance sheets than the speculative companies associated with the dot-com era.
Those are not trivial distinctions.
They are why declaring the crash inevitable would be premature.
The Question Isn’t Whether AI Works. It’s Whether We Paid Too Much for the Future Too Early.
Artificial intelligence does not need to be a scam for investors to lose staggering amounts of money investing in it.
That is the lesson the dot-com analogy is actually useful for.
The skeptics who insist AI is worthless may ultimately look as foolish as anyone who dismissed the internet in 1999.
But the opposite argument—that a transformative technology automatically justifies virtually unlimited spending—is no more sophisticated.
AI can change medicine, software development, scientific research, education, logistics and almost every knowledge industry.
Nvidia can build an extraordinary business.
Microsoft can generate tens of billions of dollars from AI.
Data centers can remain essential infrastructure.
And the AI investment boom can still become a bubble.
All of those things can be true at once.
JPMorgan’s current position inadvertently captures that tension better than almost anyone.
One part of the institution is helping arrange some of the largest AI infrastructure financings ever attempted.
Another is warning that hyperscaler cash flow can no longer finance the buildout alone.
Another warns that AI-related borrowing is becoming an increasingly important concentration inside bond indexes.
Another has identified a market divergence resembling one that appeared around the end of the dot-com boom.
And JPMorgan still believes AI demand is real enough to help propel the S&P 500 higher.
That isn’t necessarily hypocrisy.
It is a snapshot of an investment boom reaching the point where the easy question—“Is AI important?”—has already been answered.
The hard question is what comes next:
How much is that importance actually worth?
Goldman Sachs estimates today’s trajectory could require approximately $7 trillion to $8 trillion of AI investment, while acknowledging that the economic payoff would require the technology to create entirely new pools of economic activity beyond simply redistributing existing profits.
That is the bet now being made.
Not that AI will work.
We already know it works.
The bet is that it will create enough additional profit to justify several trillion dollars being spent before anyone can know exactly where that profit will come from.
The internet eventually justified building an enormous amount of infrastructure.
That did not save the investors who paid the wrong price to build too much of it too early.
The technology can win while the capital allocation loses.
That is the AI bubble risk worth watching.
References and Further Reading
JPMorgan — Primary Analysis
J.P. Morgan Private Bank — “Patchmageddon” and the semiconductor/hyperscaler divergence JPMorgan’s own discussion of Jason Hunter’s comparison between today’s AI market structure and the late-1990s communications boom.
J.P. Morgan Asset Management — Financing the AI Buildout Current estimates for hyperscaler spending, cash-flow pressure, cloud growth, data-center leasing and differences between today’s market and the dot-com infrastructure boom.
J.P. Morgan Asset Management — Can Credit Markets Absorb the AI Buildout? Documents the rapid growth in hyperscaler bond issuance and JPMorgan’s expectation that roughly $2 trillion of the buildout could enter investment-grade credit markets.
J.P. Morgan Asset Management — AI: The New Bond Giant Explains how AI borrowing is creating new concentration risk inside passive bond indexes.
J.P. Morgan Asset Management — How Is AI Being Monetized? Source for JPMorgan’s estimate that a 10% return on AI investment could require roughly $650 billion in annual revenue.
J.P. Morgan — Financing AI Infrastructure and U.S. Data Centers Details JPMorgan’s role in financing Stargate, Hut 8, CoreWeave and the broader data-center capital cycle.
Independent Wall Street Analysis
Goldman Sachs — The AI Investment Boom: When Will It Pay Off? Jim Covello and other Goldman researchers debate whether profits farther up the AI supply chain can justify enormous semiconductor and infrastructure spending.
Goldman Sachs — Will the Corporate Investment in AI Pay Off? Explores the imbalance between semiconductor profitability and the still-developing economics of enterprise AI.
Goldman Sachs — Tracking Trillions: The Assumptions Shaping the Scale of the AI Build-Out Models approximately $7.6 trillion in AI-related compute, data-center and power investment from 2026 through 2031.
Market Concentration
S&P Dow Jones Indices — S&P 500 Constituents and Sector Weightings Current index composition showing the extraordinary weight of information technology and major AI-related companies.
Data Centers, Taxpayers and Ratepayers
Washington State JLARC — 2026 Tax Preference Review: Data Centers in Urban Counties Examines $42.4 million in estimated tax savings, reported employment benefits and uncertainty over whether the incentive actually caused the investment.
Fulton County, Georgia — Fulton County Opposes Local Incentives for Data Centers Documents the county’s August 2026 decision opposing additional local data-center subsidies.
Federal Energy Regulatory Commission — Action on Large-Load and Data-Center Grid Integration Explains FERC’s 2026 action requiring regional grid operators to address large-load connections while protecting existing ratepayers.
Current Reporting
Business Insider — JPMorgan Says Be Ready to Sell Stocks as Technical Warnings Pile Up Heading Into September Reporting on Jason Hunter’s latest technical-market warning and the important qualification that major indexes remain in bullish trends.
Financial Times — The Multiplying Risks of Financing Data Centres Examines leverage, obsolescence, insurance, demand uncertainty and stranded-asset risks emerging around the enormous data-center financing boom.
Editorial currency note: Market prices, index weights, capital-expenditure forecasts, bond issuance, data-center projects, tax incentives and utility regulations can change rapidly. Figures and policy descriptions in this article reflect information available through August 27, 2026.
