Artificial intelligence is changing the world. But with hundreds of billions pouring into data centers, rising debt, and investors pricing in extraordinary future profits, the biggest technology revolution of our lifetime may also become one of the biggest financial corrections. The uncomfortable question isn’t whether AI works. It’s whether anyone can make enough money from it to justify what we’re spending.
Welcome to the Most Expensive Game of Musical Chairs in History
Imagine walking into a casino where everyone is winning.
Nvidia is selling chips at extraordinary margins. Microsoft, Google, Amazon, and Meta are building enormous data centers. AI startups are raising money at valuations that would have sounded like comedy sketches five years ago. Investors are pouring money into anything remotely associated with artificial intelligence.
The casino is booming.
There’s just one slightly awkward question:
Who is ultimately paying for all of this?
Not who is financing it. We know that answer: shareholders, bondholders, private-credit funds, venture capitalists, and the enormous cash flows of Big Tech.
I mean who is paying for the actual economic value created by AI?
Businesses? Consumers? Governments? Advertisers?
And are they paying enough to generate an acceptable return on hundreds of billions of dollars of infrastructure?
That question sits at the center of the AI bubble debate.
In August 2026, Nvidia reported quarterly revenue of $96.2 billion, up 106% from the previous year. Its data center business alone generated $89 billion.
CEO Jensen Huang declared:
“Now, compute is revenue.”
He’s right. For Nvidia, compute is revenue. Very impressive revenue.
But for Nvidia’s customers, compute is an expense that must eventually earn its keep.
And that’s where this story gets interesting.
My thesis is straightforward: AI is a genuine technological revolution, but parts of its financial ecosystem are displaying classic bubble characteristics.
I don’t believe AI is going away. Quite the opposite.
I believe AI will become so important that we may spend an extraordinary amount of money building it too quickly, at the wrong prices, and with the wrong expectations.
History has a rather unpleasant habit of punishing that combination.
1. The Numbers Have Become Almost Absurd
Let’s begin with the spending.
According to Charles Fitzgerald’s analysis of hyperscaler capital expenditures, Amazon, Google, Meta, and Microsoft were collectively planning approximately $735 billion to $750 billion in 2026 capital expenditure, depending on accounting treatment.
That’s not a typo.
Three-quarters of a trillion dollars.
And that’s just four companies.
Not all this money goes exclusively toward AI. These businesses also invest in conventional cloud capacity, networking, warehouses, and other infrastructure. But AI has become a major driver of the acceleration.
Consider some of their disclosed plans:
Company
2026 capital expenditure outlook
Amazon
Approximately $220B
Alphabet/Google
$195B–$205B
Meta
$130B–$145B
Microsoft
Approximately $175B reported, with lease-accounting changes affecting comparability
These are company forecasts and analysts’ aggregations, not audited full-year spending totals.
The scale is nevertheless breathtaking.
In the second quarter of 2026 alone, the four companies spent approximately $172 billion on capital expenditures, including relevant finance leases. That exceeded their combined spending throughout all of 2023.
Think about that.
The AI infrastructure arms race has reached the point where three months of spending exceeds what these companies spent in an entire year only three years earlier.
Meanwhile, Stanford’s 2026 AI Index reports that global corporate AI investment reached approximately $582 billion in 2025, more than double the previous year’s level.
That figure includes investment categories such as private funding and acquisitions, so it shouldn’t be added directly to hyperscaler capital expenditure. Still, it illustrates the same phenomenon.
Capital is racing into AI at a remarkable speed.
Investment at this scale isn’t automatically irrational.
The internet required enormous spending on networks, servers, and fiber-optic cables. Electricity required power plants and distribution networks. Railroads required tracks and stations.
Transformative technologies need infrastructure.
But here’s the important distinction:
The amount of infrastructure society ultimately needs is not the same as the amount investors can profitably build today.
Railroads changed civilization.
They also bankrupted investors.
The two outcomes are perfectly compatible.
2. The Dot-Com Comparison Is Both Right and Wrong
Whenever someone mentions an AI bubble, somebody inevitably responds:
“This is just like the dot-com bubble.”
And somebody else immediately says:
“No, this time the companies have actual revenues!”
Both arguments contain some truth.
During the late 1990s, investors became convinced that the internet would transform virtually every industry.
They were correct.
Unfortunately, they also concluded that nearly every company with a website deserved an astronomical valuation.
That part didn’t work out so well.
According to Goldman Sachs’ historical review, the Nasdaq Composite peaked at 5,048 in March 2000. By October 2002, it had declined roughly 78%.
Pets.com disappeared.
Amazon survived.
Google became a giant.
The internet kept growing.
Investors who bought the wrong businesses at the wrong valuations lost fortunes even though their broad technological thesis was correct.
But an even more relevant comparison is the telecommunications infrastructure boom.
A Federal Reserve Bank of New York analysis concluded that excessive optimism about future technology profits contributed to unsustainable investment, particularly in communications infrastructure, and helped drive the subsequent downturn.
Companies built networks for demand that hadn’t yet materialized.
When expectations collapsed, the financing collapsed with them.
Yet the physical infrastructure remained.
The world eventually benefited from networks financed by investors who never recovered their capital.
There’s an important difference today: Microsoft, Alphabet, Amazon, Meta, and Nvidia are not Pets.com.
These are large, established, profitable businesses.
Microsoft, for example, reported annual revenue of $331.8 billion for fiscal 2026, with operating income of $155.2 billion.
That’s real money.
Nobody needs to invent an imaginary customer to explain Microsoft’s business.
But a profitable company can still make a terrible capital allocation decision.
And investors can overpay for an excellent company.
The technology, the business, and the stock are three different things.
Confusing them is how bubbles become expensive.
3. AI Is Creating Real Value. Let’s Stop Pretending Otherwise.
Before going further, we should challenge the strongest version of the bubble argument.
The claim that AI is mostly smoke and mirrors doesn’t survive contact with the evidence.
AI is already helping people write code, analyze documents, support customers, create content, conduct research, and automate increasingly complex tasks.
Stanford’s 2026 AI Index found that 88% of surveyed organizations reported using AI in at least one business function in 2025.
The same report estimated approximately $172 billion in annual consumer surplus from generative AI in the United States by early 2026.
Consumer surplus is the economic value people receive beyond what they pay. It is not equivalent to company revenue, which is precisely why the distinction matters.
There are also meaningful productivity gains.
Stanford summarizes studies showing gains of roughly 14%–15% in customer support, 26% in software development, and substantial improvements in certain marketing tasks.
Microsoft provides a compelling commercial example.
In its July 2026 earnings release, the company reported more than 30 million paid Microsoft 365 Copilot seats.
Satya Nadella framed Microsoft’s ambition as helping:
“every customer can turn tokens into business results.”
That’s the right ambition.
And Microsoft’s Azure business grew 43% year over year in its June 2026 quarter, although that growth includes more than AI workloads.
The evidence for commercial demand is real.
But here’s what makes the situation complicated.
Value creation doesn’t automatically translate into value capture.
A customer might receive $100 worth of productivity from an AI tool while paying $20 for it.
That’s fantastic for the customer.
Whether it’s fantastic for the model developer depends on what delivering that service costs.
Likewise, AI can make employees more productive without delivering equivalent financial savings to their employers. Workers may simply accomplish more, review more work, or absorb additional responsibilities.
And an enterprise can spend millions on AI initiatives without ever successfully reorganizing the workflows required to capture the benefits.
Which brings us to the uncomfortable productivity gap.
4. The Productivity Paradox: Everybody Uses AI, But Where Are the Profits?
One of the most revealing pieces of research comes from a February 2026 National Bureau of Economic Research working paper.
Researchers surveyed nearly 6,000 senior business executives across the United States, United Kingdom, Germany, and Australia.
They found that 69% of firms actively used AI.
That sounds encouraging.
But executives who regularly used AI averaged only about 1.5 hours of usage per week.
The sample, methodology, and definition of AI use differ from Stanford’s organizational adoption surveys. Those numbers are not contradictory.
Together, however, they reveal a critical distinction:
Deploying AI is not the same as transforming a business with AI.
An organization might purchase hundreds of AI licenses and declare itself transformed.
Meanwhile, employees are still copying data between spreadsheets, approving invoices through seven different systems, and attending meetings to discuss why nobody can find the latest PowerPoint.
AI has arrived.
The bureaucracy remains undefeated.
Even experimental evidence challenges some of the more enthusiastic productivity claims.
In a 2025 randomized study by METR, experienced open-source developers working on familiar repositories took 19% longer to complete assigned tasks when using the AI tools available at the time.
That doesn’t mean AI makes developers slower in general.
The study examined a particular group, task environment, and generation of tools. AI capabilities have advanced considerably since then.
But it does demonstrate something investors sometimes forget.
A compelling product demonstration is not the same thing as a proven productivity improvement in a real production environment.
And benchmark scores are not cash flow.
This is especially important for autonomous AI agents.
An agent successfully completing a task in a controlled demonstration is impressive.
An agent reliably performing thousands of financial transactions, navigating exceptions, maintaining an audit trail, respecting compliance requirements, and avoiding costly mistakes is a different proposition.
In regulated industries, the last few percentage points of reliability can be more expensive than the first ninety.
From a product management perspective, this is the gap between feature adoption and measurable business outcomes.
It’s also where the difference between a transformative AI company and an expensive science project becomes painfully obvious.
5. The $2 Trillion Question Nobody Can Avoid
Now let’s examine the economics.
In September 2025, Bain & Company published an analysis with an extraordinary conclusion.
Under its projected high-growth AI infrastructure scenario, supporting the required computing capacity by 2030 could involve approximately $500 billion in annual data center capital expenditure.
Bain estimated that sustaining that investment at conventional cloud-provider capital-intensity ratios would require about $2 trillion in annual revenue.
Even after modeling substantial enterprise IT spending shifts and productivity savings, its scenario still contained an approximately $800 billion annual revenue gap.
These figures are projections based on assumptions about compute demand, efficiency, and spending. They are not an audited shortfall in today’s AI economy.
But the underlying question is unavoidable.
Where does the money come from?
Consider a simplified example.
Suppose an AI infrastructure provider invests $100 billion building computing capacity.
Assume the hardware has an economic life of five years.
Ignoring financing costs, maintenance, buildings, energy, and other expenses, that already implies roughly $20 billion of annual depreciation or capital recovery requirements under a simplified straight-line model.
Now add operating costs, financing, hardware replacement, and the return investors expect.
The revenue hurdle becomes enormous.
And GPUs have an awkward characteristic compared with railway tracks or fiber-optic cables.
They can become technologically outdated quickly.
A railway built in 1900 could remain useful decades later.
A premium AI accelerator purchased in 2026 may face significantly more capable and economically efficient competition just a few years afterward.
The building might last twenty-five years.
The economics of the machines inside it can change much faster.
That creates a brutal financial tension.
AI companies need more computing capacity to grow.
But growing their infrastructure increases the amount of revenue they must generate just to justify the investment.
At some point, demand must catch up with capacity.
If it doesn’t, prices fall, margins compress, and the least-efficient providers discover that depreciation is not just an accounting concept.
It’s reality sending an invoice.
6. The Most Dangerous Part of the Bubble Is the Financing
The AI boom began with technology companies spending their own enormous profits.
That is a relatively resilient model.
More recently, however, its infrastructure financing has become increasingly complex.
Debt markets, private credit, equipment-backed loans, financing vehicles, and long-term capacity commitments now play larger roles.
And one company illustrates both the opportunity and the danger particularly well.
CoreWeave: The Canary in the GPU Mine
CoreWeave is a specialized AI cloud provider that has grown rapidly by supplying compute infrastructure to major technology companies and AI developers.
Its customer demand is substantial.
According to CoreWeave’s second-quarter 2026 results, the company reported approximately $104 billion in revenue backlog as of June 30.
Impressive.
But its SEC filing for the same quarter disclosed $35.6 billion in total indebtedness.
Its growth requires tremendous upfront infrastructure expenditure.
And backlog is not cash in the bank.
Contracts still need to be fulfilled. Capacity must become operational. Customers must continue paying. Operating and financing costs must remain manageable.
By October 15, 2026, Reuters reported that CoreWeave’s debt-related exposure had reached roughly $50 billion, and that credit-default-swap spreads had risen above 1,100 basis points.
That reflected sharply increasing market concern about creditworthiness.
It did not mean bankruptcy was inevitable. Market-implied default probabilities depend on pricing assumptions and can change rapidly.
But the signal was unmistakable.
Investors were demanding much more compensation for taking AI infrastructure credit risk.
And CoreWeave wasn’t the only company under pressure.
On October 15, Reuters documented postponed or abandoned data center financing and listing plans as rising borrowing costs and falling market valuations disrupted the sector.
This is how a financial bubble often begins to unwind.
Not with a dramatic announcement that the technology is useless.
With a lender saying:
“Wonderful technology. We’d like a higher interest rate.”
Or, even worse:
“No, thank you.”
The Circular Financing Problem
There’s another feature of the AI economy that deserves scrutiny.
Some of the same companies are simultaneously investors, suppliers, customers, and strategic partners.
A chipmaker invests in an infrastructure company.
That infrastructure company buys chips.
A major AI developer commits to purchasing computing capacity.
The developer itself receives investment from technology companies that supply its infrastructure.
These arrangements can be commercially legitimate and strategically sensible.
They can also make demand look more independent than it really is.
A supplier’s revenue might ultimately depend on the financial viability of customers whose growth is supported by the supplier’s broader investment ecosystem.
That doesn’t automatically constitute fraud or artificial revenue.
But it does increase interconnectedness.
If one heavily financed participant runs into trouble, several others may discover that their supposedly separate growth forecasts were relying on the same underlying funding.
The IMF’s October 2026 Global Financial Stability Report explicitly warned that AI-market corrections could be amplified by leverage, concentration, and circular financing.
The risk is not simply that AI demand falls.
It’s that the financial system has increasingly placed correlated bets on that demand.
A useful analogy is a neighborhood where everyone keeps buying each other’s houses at higher prices using newly borrowed money.
Everybody has impressive transaction volume.
Everybody’s property valuation is rising.
But eventually, somebody needs to pay the mortgage with income earned outside the neighborhood.
7. Even the Winners May Be Vulnerable
One argument against an AI bubble is that leading technology companies continue reporting extraordinary results.
This is true.
Nvidia’s revenue growth is real.
Microsoft’s cloud revenue is real.
Amazon’s AWS revenue grew 37% year over year in the second quarter of 2026.
Oracle’s infrastructure business is also experiencing exceptional demand.
In June 2026, Oracle reported $638 billion in remaining performance obligations, reflecting enormous future contract commitments.
But Oracle also reported negative free cash flow of $23.7 billion for fiscal 2026 while financing major infrastructure expansion.
Importantly, Oracle disclosed that customers had prepaid or supplied $75 billion worth of GPU-related hardware under large AI contracts, reducing part of its financing burden.
That’s a meaningful mitigating factor.
The point isn’t that Oracle’s backlog is fictional or that every infrastructure company is financially fragile.
The point is that different companies occupy radically different positions within the same AI economy.
Nvidia gets paid when infrastructure is purchased.
A cloud provider must earn a return from operating that infrastructure.
An AI model developer must monetize the computing services it consumes.
An enterprise must convert AI expenditure into genuine business results.
Every layer depends on the next layer eventually creating enough value to support the economics of the one above it.
Today, the financial benefits are concentrated heavily in parts of the infrastructure supply chain.
The ultimate economic returns are still being distributed and tested.
That’s the mismatch investors should watch.
8. AI Has Another Problem: It Keeps Getting Cheaper
Here’s a possibility that doesn’t receive enough attention.
What if AI becomes dramatically better and cheaper, exactly as its advocates predict?
That would be wonderful for users.
But it could complicate the economics for infrastructure owners.
Imagine a company expects to sell AI computation at a certain price for the next five years.
It invests billions based on that assumption.
Then more efficient models, specialized chips, better algorithms, and increased competition cause the price of useful computation to collapse.
Demand may increase enormously.
Yet revenue per unit falls.
This is familiar in technology.
Storage became cheaper.
Bandwidth became cheaper.
Computing became cheaper.
Consumers benefited spectacularly.
Not every supplier earned spectacular returns.
Economists would recognize a possible Jevons effect: greater efficiency can expand total consumption.
That could save the AI infrastructure boom if falling costs unlock enough additional demand.
But it’s not guaranteed.
If unit prices fall 80% while usage triples, revenue still declines 40%.
That’s illustrative arithmetic, not an AI market forecast.
And the infrastructure owner’s debt doesn’t conveniently decline alongside token prices.
There is a second complication.
As high-quality models become more widely available, a growing share of AI functionality may become commoditized.
A startup built around access to one model, one prompting technique, or one narrow workflow can suddenly find its supposed competitive advantage appearing inside a major platform.
The model improves.
The customer wins.
The startup’s valuation has a difficult conversation with reality.
This is why I suspect many AI startups will struggle even if enterprise AI adoption continues accelerating.
The market can grow while individual companies become less valuable.
9. So When Will the AI Bubble Burst?
Here comes the answer that financial television dislikes.
Nobody knows.
Anyone claiming to know the exact quarter when the AI bubble will burst should probably spend less time forecasting and more time explaining why they haven’t already become a billionaire trading on that extraordinary insight.
But we can identify the conditions that would make a significant correction more likely.
And we can identify a plausible window of financial vulnerability.
As of October 2026, I believe the next 12 to 36 months will be an especially important test.
That’s my assessment, not a statistically reliable crash forecast.
The reason is straightforward.
Infrastructure investment is accelerating now.
New capacity will continue entering service.
Financing costs matter increasingly as debt-funded projects expand.
And investors will demand more evidence that the investment is generating attractive returns rather than simply producing impressive announcements.
In fact, the correction may already be beginning in some parts of the market.
The October 2026 weakness in AI-linked equities, bonds, and infrastructure financing is a warning sign.
It is not yet proof of a comprehensive bubble collapse.
Three Plausible Outcomes
Scenario A: The Slow Deflation
AI adoption continues growing, but not fast enough to justify every infrastructure investment and startup valuation.
Capital expenditure growth slows.
Some data centers are delayed.
Private valuations decline.
Publicly traded AI companies experience substantial multiple compression.
Weaker businesses disappear while profitable technology leaders continue expanding.
This is probably the healthiest type of correction.
Painful for investors, but manageable for the wider economy.
Scenario B: The Credit Accident
A heavily indebted infrastructure company struggles to refinance its obligations.
Lenders become nervous.
Credit spreads widen.
Projects are postponed or canceled.
Suppliers lose expected orders, forcing them to revise growth forecasts.
Investors then begin questioning other companies whose business models depend on similar financing arrangements.
The critical shift occurs when people stop asking, “How quickly can you grow?” and start asking, “Can you pay your debts?”
That is a much more dangerous situation because credit problems can become self-reinforcing.
The latest market developments make this scenario more credible than it appeared a year ago, although its probability remains impossible to establish precisely.
Scenario C: The Boom Continues
This scenario deserves serious consideration.
AI capabilities improve rapidly.
Enterprises redesign entire workflows rather than merely buying software licenses.
Autonomous agents become dependable enough to automate meaningful economic activity.
AI-driven scientific discovery accelerates.
Demand for computing capacity grows faster than infrastructure can be supplied.
Under that scenario, current investment may eventually appear rational.
Some investors could still overpay, but the sector could grow into today’s extraordinary expectations.
The challenge is that Scenario C requires more than better models.
It requires successful commercialization at enormous scale.
The winners must produce real productivity gains, customers must be willing to pay for them, and the economics must work throughout the technology stack.
That’s possible.
It is not guaranteed.
10. Five Signals That Will Tell Us the Bubble Is Bursting
Rather than pretending to predict a date, I would watch five indicators.
Capital expenditure guidance: If Microsoft, Alphabet, Amazon, and Meta begin reducing infrastructure spending because returns are disappointing rather than because capacity is becoming more efficient, suppliers will feel the consequences quickly.
Credit spreads and refinancing: Rising yields, tougher loan terms, failed financings, and increasingly expensive default protection at AI infrastructure companies would signal that lenders are losing confidence.
Utilization and pricing: Data centers can remain busy while economics deteriorate. Watch whether compute rental prices, utilization, and gross margins support the cost of installed capacity.
Enterprise ROI: More licenses and token consumption are not enough. Look for independently demonstrated improvements in operating margins, revenue, customer retention, or the cost of delivering a service.
The quality of AI-company revenue: Annualized run rates, contract commitments, and actual recognized revenue are different things. So are revenue and cash flow.
That final point is particularly timely.
In October, Reuters reported that OpenAI’s September annualized revenue was approaching $50 billion, compared with an earlier indication approaching $70 billion.
The reported discrepancy involved differences in revenue definitions, including how cloud-partner sales were counted. It should not automatically be interpreted as a collapse in underlying sales.
But it exposes a serious analytical problem.
When private companies are valued on extraordinary future expectations, inconsistent revenue definitions can create enormous misunderstandings.
An annualized revenue run rate isn’t annual audited revenue.
Backlog isn’t cash.
Gross revenue isn’t profit.
And an investor presentation isn’t an income statement.
These distinctions become particularly important when financial conditions tighten.
11. What Would the Fallout Look Like?
Suppose the bearish scenario materializes.
What happens next?
First, AI Stocks Get Punished
Infrastructure suppliers, specialized cloud providers, and highly valued AI startups would probably face the greatest immediate pressure.
Investors would revise expected growth rates, profitability, and terminal valuations.
Some excellent businesses could experience large share-price declines.
This is what happens when the future is already priced into the present.
Even companies growing rapidly can disappoint if the market expected them to grow faster.
The damage could extend beyond specialist AI names because major technology companies occupy such large positions in equity indexes.
Retirement portfolios, pension funds, and passive investors could experience losses even if they never deliberately invested in a single AI startup.
Second, Startups Enter a Funding Winter
The venture capital industry would become much less interested in beautiful demonstrations and much more interested in revenue quality.
Companies offering little beyond generic AI functionality would struggle.
Startups with expensive inference costs and weak customer retention would face painful restructuring.
Some would be acquired for modest prices.
Others would simply disappear.
I suspect many founders would suddenly rediscover an ancient concept called profitability.
Terrifying, I know.
But unlike the collapse of a genuine technological market, a financing correction could actually strengthen the surviving ecosystem.
Resources would move from speculative experiments toward businesses delivering repeatable value.
Third, Data Center Investment Slows
A sharp spending slowdown would affect more than technology companies.
Construction firms, electrical contractors, semiconductor suppliers, utilities, networking equipment manufacturers, and local real estate markets could all experience lower demand.
The energy implications are substantial.
The International Energy Agency projects that global data center electricity consumption could more than double to approximately 945 terawatt-hours by 2030.
AI is a major driver of that forecast.
If investment contracts abruptly, some planned power projects could be delayed or reevaluated.
Communities that benefited from rapid development might discover how quickly speculative growth can reverse.
A data center doesn’t create the same employment footprint as a factory employing thousands of people.
But constructing one can support a substantial temporary ecosystem of workers and suppliers.
Fourth, the Real Economy Takes a Hit
In its October 2026 World Economic Outlook, the IMF identified a reassessment of AI expectations among the downside risks to global growth.
The accompanying financial stability report warned that a market correction could be amplified through leverage and concentrated financial exposures.
That matters because AI investment has become an important source of economic activity.
A sudden reduction would weaken demand for equipment, construction, energy, and professional services.
Falling financial markets could also reduce household wealth and consumer spending.
But we should avoid sensationalism.
An AI correction would not automatically produce another 2008 financial crisis.
The largest technology companies still have substantial operating profits and diversified businesses. The IMF also finds the global banking system broadly resilient under its adverse stress scenarios, though vulnerabilities remain.
The severity of any downturn would depend on where the losses occur, how much leverage is involved, and whether distressed companies can continue operating.
A stock-market correction is one thing.
A credit crisis is another.
Fifth, AI Could Become More Affordable
Here’s the twist.
An AI infrastructure bust might ultimately accelerate AI adoption.
If excess computing capacity enters the market and financially distressed operators need customers, prices could fall.
Startups building useful applications could access cheaper computing resources.
Enterprises might obtain better commercial terms.
Open-source and smaller specialized models could become more attractive.
Capital-intensive model providers might struggle while application companies with strong customer relationships prosper.
Think about the internet after the dot-com collapse.
The financial bubble disappeared.
The technological opportunity did not.
In fact, much of the internet’s subsequent growth benefited from infrastructure installed during the boom.
The biggest winners from an AI infrastructure bust might be companies that never had to pay the original price of building it.
That’s an intriguing possibility.
And perhaps the most important strategic insight in this entire debate.
12. What This Means for Product Leaders
From a product strategy perspective, I believe the next stage of AI will involve a significant change in how companies measure success.
Over the past several years, much of the conversation has focused on model capabilities, benchmark scores, context windows, token throughput, and adoption.
Those metrics matter.
But they’re not sufficient.
The next generation of successful AI products will have to demonstrate defensible economics.
A company building an AI customer support system should be able to show that it reduces the fully loaded cost per resolved issue without damaging customer satisfaction.
An AI coding product should improve delivery speed, code quality, or development cost in realistic environments—not merely generate impressive code during a demonstration.
An AI financial services agent should demonstrate better processing efficiency while preserving security, regulatory compliance, and customer trust.
And an AI platform should understand its unit economics.
What does it cost to acquire a customer?
What does it cost to serve one?
How much inference capacity is required?
How much human supervision remains necessary?
What happens to margins when usage increases tenfold?
Where does durable competitive advantage come from if foundation models become interchangeable?
These are not glamorous questions.
They don’t generate standing ovations at developer conferences.
But they determine whether a company can survive when venture funding becomes less generous.
A product strategy built on the assumption that computing costs will fall forever, customers will keep paying premium prices, and competitors will never catch up is not really a strategy.
It’s a collection of wishes with an architecture diagram attached.
I would rather build an AI business that survives a downturn than one that requires the boom to continue indefinitely.
That means focusing on proprietary customer insight, deeply integrated workflows, measurable outcomes, distribution advantages, and defensible customer relationships.
Models can become commodities.
Real customer value is harder to replace.
13. The Uncomfortable Truth: AI Can Change Everything and Still Destroy Trillions
Financial bubbles are rarely built around entirely imaginary ideas.
They’re often built around genuine transformations combined with unreasonable expectations about timing and profit.
The railway revolution was real.
The automobile revolution was real.
The internet revolution was real.
The housing market was real, although the financial assumptions surrounding it became dangerously detached from reality.
AI is real too.
Possibly more transformative than many of these earlier technologies.
But technological importance does not guarantee investment returns.
I think much of the market is making three dangerous assumptions.
First, that AI adoption will quickly produce enormous revenue growth.
Second, that today’s infrastructure owners will capture a durable share of that revenue.
Third, that the economics of AI infrastructure will remain sufficiently attractive to justify current capital commitments.
All three assumptions might prove correct.
But they are far from certain.
And investors are committing extraordinary amounts of money before the answers are known.
As the IMF warned in October 2026:
“Resilience is not safety.”
That’s an exceptionally good description of the current market.
Strong earnings don’t eliminate financial risk.
Rapid adoption doesn’t guarantee profitability.
And a revolutionary technology doesn’t protect investors from paying too much.
Final Verdict: Is AI a Bubble?
Yes—in parts of the market, I believe it is.
Not because AI is fake.
Not because people will stop using it.
And not because the technology has failed to demonstrate genuine value.
I see bubble-like conditions because financial commitments and valuations in parts of the ecosystem depend on extraordinary future demand, financing availability, and assumptions about profitability that remain unproven.
The most vulnerable area appears to be debt-heavy infrastructure and businesses whose economics depend on continued capital-market enthusiasm.
The strongest companies have more ways to survive. They possess profitable existing operations, distribution advantages, and the financial flexibility to adjust spending.
But even they are not immune to valuation corrections or wasted investment.
When will the bubble burst?
My preferred window to watch is late 2026 through 2029, with 2027–2028 a particularly important period for assessing whether the enormous infrastructure buildout is producing adequate cash returns.
That is a risk window, not a prediction of a crash date. The correction could arrive sooner, take years to unfold, or be avoided if demand grows into current expectations.
And the fallout?
Potentially severe for leveraged infrastructure firms, speculative startups, concentrated technology portfolios, and industries dependent on rapid data center construction.
But not necessarily catastrophic for AI itself.
In fact, there’s a reasonable chance that the eventual correction—if it comes—will help separate enduring products from expensive experiments.
The technology will continue improving.
Useful products will continue finding customers.
Businesses will keep automating workflows and redesigning operations.
And the next generation of companies may benefit from cheaper compute, more realistic valuations, and lessons paid for by the previous generation’s investors.
The internet didn’t end in 2000.
It finally grew up.
AI may eventually experience something similar.
The people most excited about its long-term potential should therefore be willing to ask the toughest questions about its current economics.
Because the truly dangerous statement in an investment bubble isn’t that a revolutionary technology will change the world.
It’s that changing the world means the price no longer matters.
The future of AI may be extraordinary. That doesn’t mean today’s investors are entitled to extraordinary returns.
And when the music finally stops, somebody is going to be left holding a very expensive GPU.
The real question isn’t whether AI will transform the economy. It’s whether the companies financing that transformation can survive long enough—and capture enough of the value—to enjoy the results.
What do you think: are we witnessing the infrastructure foundation of the next economic revolution, or the biggest capital misallocation since the dot-com boom? Perhaps both?