The AI Bubble Is Finally Starting to Crack

Executive Summary

Big Tech is no longer merely investing in artificial intelligence.

It is reorganizing the capital structure of the technology sector around it.

Amazon, Microsoft, Alphabet, and Meta are now spending at a scale that increasingly resembles a national industrial policy program rather than a normal corporate investment cycle. The money is flowing into data centers, chips, servers, networking equipment, cooling systems, power infrastructure, and long-duration AI capacity commitments.

The bullish argument is obvious: AI could become the next operating layer of the global economy.

The bearish argument is becoming harder to dismiss: the spending is arriving much faster than the proven return on investment.

McKinsey’s 2025 global AI survey found that most organizations still have not seen meaningful enterprise-wide EBIT impact from AI. Only 39% of respondents attributed any level of EBIT impact to AI, and most of those said AI accounted for less than 5% of enterprise EBIT. McKinsey classified only about 6% of respondents as AI “high performers” generating meaningful bottom-line impact. ([McKinsey & Company][1])

That gap — between infrastructure spending and realized earnings contribution — is the entire story.

AI may be real. The productivity upside may be real. The long-term platform shift may be real.

But markets do not crash because the technology is fake.

They crash when the price paid for the technology becomes untethered from the cash flows it can produce.

The New AI Trade: Spend First, Justify Later

The current AI boom has evolved from a software story into a capital intensity story.

In the early phase, the market rewarded anything connected to generative AI: chips, cloud, software, cybersecurity, automation, enterprise tools, data-center REITs, power equipment, nuclear power, cooling systems, and even utilities.

Now the question has changed.

It is no longer: Who benefits from AI?

It is: Who earns an acceptable return on the capital being deployed?

Goldman Sachs’ baseline model estimates annual AI infrastructure capital expenditures of roughly $765 billion in 2026, rising to $1.6 trillion by 2031, with approximately $7.6 trillion of cumulative capex between 2026 and 2031. That estimate includes the full physical stack required for AI: accelerators, data centers, power delivery, redundancy systems, cooling infrastructure, and supporting electrical equipment.

That is not a normal upgrade cycle.

That is a civilization-scale buildout.

And the market is now being asked to believe that this buildout will produce returns high enough, fast enough, and durable enough to justify one of the largest corporate spending waves in history.

The Hyperscaler Spending Surge Is Real

The AI buildout is not theoretical. It is already visible in the financial statements.

Alphabet reported Q1 2026 revenue of $109.9 billion, up 22% year over year, while Google Cloud revenue increased 63% to $20.0 billion, driven by enterprise AI solutions, AI infrastructure, and core Google Cloud Platform services. Alphabet also reported that Google Cloud backlog nearly doubled quarter over quarter to more than $460 billion.

But the capex intensity is rising quickly. Alphabet’s Q1 2026 purchases of property and equipment reached $35.7 billion, more than double the $17.2 billion spent in the prior-year period.

Microsoft also delivered powerful headline results. Revenue rose to $82.9 billion, up 18%, while CEO Satya Nadella said Microsoft’s AI business surpassed a $37 billion annual revenue run rate, up 123% year over year. ([Microsoft][4]) But Microsoft’s additions to property and equipment rose to approximately $30.9 billion for the quarter, compared with $16.7 billion in the prior-year period.

Meta reported Q1 2026 revenue of $56.31 billion, up 33% year over year, while capital expenditures, including principal payments on finance leases, were $19.84 billion. More importantly, Meta raised its full-year 2026 capex outlook to $125 billion to $145 billion, up from its previous range of $115 billion to $135 billion, citing higher component pricing and additional data-center costs.

Amazon’s Q1 2026 results also show the scale of the buildout. AWS sales increased 28% year over year to $37.6 billion, while Amazon disclosed that free cash flow fell to $1.2 billion for the trailing twelve months, driven primarily by a $59.3 billion year-over-year increase in purchases of property and equipment, which the company said primarily reflected AI investments.

The numbers are staggering.

These companies are not experimenting with AI.

They are building the physical rails of the next computing era.

The problem is that investors still do not have clean visibility into the actual AI return profile.

The Real Problem: AI Revenue Is Still Too Blurry

The hyperscalers are reporting cloud growth, backlog growth, AI usage growth, token volume growth, customer demand, and capacity constraints.

Those are real signals.

But they are not the same thing as transparent AI profitability.

Investors need to separate three different questions:

First, is AI driving cloud demand? Yes, clearly.

Second, are hyperscalers generating revenue from AI workloads? Yes, increasingly.

Third, are those AI workloads producing attractive returns on the massive incremental capital being deployed? That is still the unresolved question.

That distinction matters.

A company can grow revenue and still destroy shareholder value if the capital required to generate that revenue rises faster than the economic return.

This is where the AI cycle becomes dangerous.

During a mania, the market rewards revenue growth.

During the fear phase, the market starts asking about depreciation, free cash flow conversion, return on invested capital, customer payback periods, and useful asset life.

That is the point where the story can change quickly.

McKinsey’s AI ROI Data Should Make Investors Uncomfortable

The most important piece of the AI debate is not whether people are using AI.

They are.

The question is whether AI is transforming enterprise economics.

So far, the evidence is mixed.

McKinsey’s 2025 survey found that AI adoption is broadening across industries and functions, but most companies still have not integrated AI deeply across workflows. Only about one-third of all respondents said their organizations were scaling AI programs across the enterprise.

The more important finding is the earnings impact.

McKinsey found that only 39% of respondents attributed any level of EBIT impact to AI, and most of those said less than 5% of enterprise EBIT was attributable to AI. The firm identified AI high performers — companies seeing meaningful bottom-line impact — as only about 6% of respondents.

That does not mean AI is useless.

It means AI is not yet translating into broad, measurable, enterprise-wide profit uplift for most companies.

That is a problem when the infrastructure bill is already moving toward trillion-dollar annual scale.

The market is effectively pricing a future where AI becomes a broad corporate productivity engine.

But today’s data still suggests that most companies are somewhere between experimentation, partial deployment, and limited workflow integration.

That gap between expectation and realization is where bubbles form.

The Deflationary Problem: AI Gets Cheaper Over Time

There is another underappreciated risk.

AI infrastructure spending is exploding at the same time the cost of using AI is collapsing.

According to Stanford’s 2025 AI Index, the inference cost for a system performing at the level of GPT-3.5 dropped more than 280-fold between November 2022 and October 2024. Stanford also noted that AI hardware costs have declined by about 30% annually, while energy efficiency has improved by about 40% annually.

That sounds bullish for adoption.

And it is.

But it may not be bullish for every company spending hundreds of billions to build capacity.

If the price per unit of intelligence falls rapidly, then AI infrastructure can become a deflationary business faster than investors expect.

That creates a paradox.

The cheaper AI becomes, the more widely it may be used.

But the cheaper AI becomes, the harder it may be to earn monopoly-like returns on expensive infrastructure.

This is especially important because Stanford also found that open-weight models have been closing the performance gap with closed models, reducing the difference from 8% to 1.7% on certain benchmarks in a single year.

That weakens the idea that every hyperscaler has a durable technical moat.

If models converge, inference costs collapse, and customers become more price-sensitive, then the AI infrastructure layer may eventually behave less like a software monopoly and more like a capital-intensive commodity market.

That is the risk nobody wants to price.

The Depreciation Question Could Become the Next Market Obsession

Markets love capex during the buildout phase.

They hate depreciation during the payback phase.

AI infrastructure has a useful life problem.

Data centers may last a long time. Power infrastructure may last a long time. But the most valuable components — GPUs, accelerators, servers, memory, networking equipment, and specialized AI hardware — can become economically obsolete much faster than traditional infrastructure.

That matters because depreciation is not just an accounting detail.

It is the delayed recognition of yesterday’s capital spending.

If companies are spending tens or hundreds of billions on AI assets today, the income statement will eventually have to absorb the depreciation burden tomorrow.

The market should be watching five things:

1. Whether AI revenue disclosure becomes more transparent.
2. Whether AI workloads improve or dilute cloud margins.
3. Whether free cash flow conversion deteriorates as capex rises.
4. Whether depreciation expense grows faster than revenue from AI services.
5. Whether customers generate enough ROI to keep increasing AI spend.

This is where the AI thesis becomes much more fragile.

The first phase of the trade was about demand.

The second phase will be about returns.

This Does Not Mean AI Is Fake

The strongest version of the bearish argument is not that AI is fake.

That is lazy.

The stronger argument is that AI may be real, useful, and economically important — while still producing a massive capital cycle that overshoots demand.

That is exactly what happened in prior technology booms.

The internet was real in 1999.

Fiber-optic networks were real.

E-commerce was real.

Software was real.

Semiconductors were real.

But the dot-com bubble still happened because investors extrapolated real technological change into unrealistic financial assumptions.

The same logic applies here.

AI can change the world and still become overbuilt.

AI can improve productivity and still fail to justify every dollar of capex.

AI can generate enormous revenue pools and still compress returns if too much capital floods the market at once.

That is the difference between a technology revolution and an investable return.

Investors must not confuse the two.

The Bull Case Still Has Teeth

To be fair, the bull case is not weak.

Alphabet’s cloud acceleration is real. AWS growth is real. Microsoft’s AI revenue run rate is real. Meta’s ad business remains enormously profitable, and AI can potentially improve targeting, content recommendation, creative automation, and engagement. The AI supply chain also remains constrained, which can support pricing power for chips, networking, memory, cooling, and power equipment in the near term.

The hyperscalers also have balance sheets that smaller companies could never match.

They can fund the buildout.

They can absorb losses.

They can wait longer than most investors expect.

And if AI becomes the default interface for work, search, commerce, software development, advertising, customer service, cybersecurity, and enterprise automation, then the companies building the infrastructure could become even more dominant.

That is the serious bull case.

The issue is not whether there is upside.

The issue is whether the current spending path already assumes too much of that upside too soon.

The Fear Phase Comes After the Mania Phase

Every market cycle has a psychology.

First comes disbelief.

Then adoption.

Then acceleration.

Then euphoria.

Then rationalization.

Then fear.

AI has already moved through the early phases. The technology went from curiosity to boardroom mandate almost overnight. Every company needed an AI strategy. Every software product needed an AI feature. Every investor presentation needed an AI slide.

Now the market is entering a more difficult phase.

The spending is no longer small enough to ignore.

The capex is no longer theoretical.

The earnings impact is no longer assumed.

The question is becoming unavoidable:

Where is the return?

That is why this moment matters.

If investors continue to see accelerating AI revenue, stable margins, strong free cash flow, and credible payback periods, the boom can continue.

But if capex keeps rising while AI revenue disclosure remains vague, depreciation climbs, customer ROI remains inconsistent, and pricing pressure intensifies, the market will begin to re-rate the entire trade.

Not because AI failed.

Because the financial model became too aggressive.

What Investors Should Watch Next

The next phase of the AI market will not be decided by headlines.

It will be decided by financial statements.

Investors should watch AI revenue disclosure first. The more companies discuss AI in vague qualitative terms, the more skeptical the market should become. Real platform shifts eventually show up in segment revenue, margins, backlog quality, customer retention, and cash flow.

Second, watch free cash flow. Amazon’s disclosure that trailing twelve-month free cash flow fell to $1.2 billion as AI-related property and equipment investment surged is exactly the type of signal investors should track across the sector.

Third, watch depreciation. The larger the buildout becomes, the more tomorrow’s earnings must absorb yesterday’s infrastructure spending.

Fourth, watch inference pricing. If the cost per unit of intelligence keeps falling, the long-term winners may be customers and application-layer companies, not necessarily every infrastructure provider.

Fifth, watch enterprise ROI. If McKinsey’s future surveys show a major increase in companies generating meaningful EBIT from AI, the bull case strengthens. If not, the spending cycle becomes harder to justify.

Invest Daily Bottom Line

The AI boom is not over.

But the easy part of the narrative may be ending.

For the last two years, investors rewarded ambition. They rewarded scale. They rewarded capex. They rewarded every company that could credibly claim exposure to the AI buildout.

The next phase will be different.

The market will demand proof.

Proof of revenue.

Proof of margins.

Proof of free cash flow.

Proof of customer ROI.

Proof that the trillions being committed to AI infrastructure can generate returns above the cost of capital.

That is a much higher bar.

The danger is not that AI disappears.

The danger is that AI becomes so competitive, so expensive, and so deflationary that the companies building the infrastructure earn lower returns than investors expect.

That is how manias break.

Not all at once.

Slowly at first.

Then suddenly.

The AI revolution may be real.

But the price being paid for it may be insane.

And if the market begins to believe that, the fear phase will not need permission to begin.

Ready to Analyze Your Next Investment?

Get a free AI-powered fair value analysis on any stock. See intrinsic value, margin of safety, and institutional-grade risk metrics in seconds. No credit card required.

Want full access to our institutional research tools? Explore Invest Daily Pro.