AI is the most significant investment story of the decade. With hundreds of billions pouring into infrastructure annually, we are witnessing a historic gold rush. But behind the record-breaking revenues and soaring market valuations lies a critical question: is this the foundation of a new era, or are we repeating the speculative mistakes of the dot-com bubble?

To find the answer, we have to look past the binary ‘bubble or boom’ narrative and examine which layers of the AI stack are built on realized value and which are built on anticipation.

Somewhere around $760 billion will be spent on AI infrastructure in 2026 alone. Nvidia is booking more revenue in a quarter than it did in its first two decades combined. OpenAI plans to burn more than $200 billion before it makes money. Ten companies now make up 41% of the S&P 500. 

All these facts have two sides: AI is a bubble, and AI is here to stay.

The question everybody is asking: Is AI a bubble

The confusing part is that the loudest warnings about AI being a “bubble” that is going to burst soon are coming from the people spending the most on it. 

Sam Altman told reporters in August 2025 that investors were “overexcited” about AI and that “parts of it looked like a bubble.” 

Three months later, Sundar Pichai told the BBC there were “elements of irrationality” in the market and that no company, Google included, would be immune if it turned. 

The IMF flagged the same risk in October 2025, comparing valuations to the run-up before the dot-com crash.

None of that slowed the money down, thought. Global corporate AI investment more than doubled in 2025. Hyperscaler capex guidance for 2026 has been revised upward almost every quarter. Nvidia just posted its largest quarter ever.

So claiming that AI is a bubble is not a hot take, nor is hyping AI up. In fact, both of these takes are rudimentary. The real question is asking “which layer of AI is true in its form, and what price are we going to pay for it?”

Some parts of the AI economy are backed by real cash flow. Others are priced for a future that has not arrived. Treating them as one thing is where most of the people take the wrong approach.

First, what makes something a “bubble”

Economically, a bubble is not about prices rising fast. Prices rise fast in every strong growth cycle. A bubble is when asset prices detach from any plausible future cash flow, when the market is paying for an outcome that the underlying businesses cannot realistically deliver, even if everything goes right.

That definition matters because it separates two things people constantly blur: whether a technology is real and whether the assets tied to it are correctly priced. 

The internet was real, but the dot-com crash of 2000 still erased roughly $5 trillion in market value. Fiber-optic networks are real and still in use today, but the telecom overbuild of 1998 to 2001 still bankrupted WorldCom, Global Crossing, and dozens of others, because capacity was built years ahead of demand and financed with debt that could not be serviced.

Those two episodes are the benchmark. They are what analysts mean when they say “this looks like 1999.” The reason why AI is often considered a bubble is that it resembles these cases in some ways.

The Case That AI Is a Bubble

The case that AI is a bubble is built on three things: the scale of spending, the structure of the financing, and the gap between cost and revenue at the model layer.

The Numbers Are Short

According to Stanford’s 2026 AI Index, U.S. private AI investment reached $285.9 billion in 2025, more than 23 times what was invested in China. Global corporate AI investment hit $581.7 billion, up roughly 130% in a single year.

That is just the money that has already been spent. Goldman Sachs’ “Tracking Trillions” model puts annual AI capex at about $765 billion in 2026, rising to roughly $1.6 trillion by 2031, for a cumulative $7.6 trillion across chips, data centers and power. Goldman is careful to call these baseline estimates rather than forecasts, and notes that changing a single assumption about how quickly chips become obsolete moves the total by hundreds of billions.

The problem is what all that capex has to earn back. Bain & Company’s 2025 Global Technology Report estimated that sustaining roughly $500 billion a year in AI infrastructure spending requires about $2 trillion in annual AI revenue by 2030, and that on current trajectories the industry is around $800 billion short. Capex has grown since Bain published that number, so the revenue requirement has grown with it.

The Circular Financing Problem

INSEAD researchers put a name to the most underreported risk in the AI story: the difference between a flywheel and a house of cards.

The pattern works like this: A chipmaker invests in an AI lab. The lab uses the capital to rent data center capacity from a cloud provider. The cloud provider fills those data centers with the chipmaker’s hardware. The same dollars show up as revenue at the chipmaker, as funding at the lab, and as backlog at the cloud provider. Everyone reports growth. The question is whether any of it traces back to an end customer outside the loop.

The clearest example was Nvidia’s September 2025 letter of intent to invest up to $100 billion in OpenAI, alongside OpenAI’s $300 billion compute agreement with Oracle. That original $100 billion deal never closed. By early 2026, the Wall Street Journal reported that talks had stalled over internal doubts at Nvidia about OpenAI’s business model. What replaced it was a $30 billion equity stake and, in August 2026, a guarantee of up to $105 billion backing OpenAI’s Ohio data center campus. That guarantee had reportedly been discussed at $250 billion before being cut.

INSEAD’s point is that these arrangements look uncomfortably like the vendor-financing structures of the late dot-com era, when Cisco and Lucent extended credit to telecoms so they could keep buying routers and switches. Those sales looked great until the borrowers ran out of money. Bull markets forgive circular financing. Downturns do not.

The Profitability Gap Is Real

The foundation model companies are spending far more than they earn, and the gap keeps widening.

OpenAI reported roughly $13 billion in revenue for 2025 and burned about $9 billion. In September 2025, it told investors it expected cumulative cash burn of $115 billion through 2029. By February 2026, that projection had roughly doubled: The Information reported that OpenAI now expects to burn about $218 billion between 2026 and 2029 and does not expect to be cash flow positive before 2030. For scale, that is around 23 times what Tesla burned across its entire unprofitable stretch from 2007 to 2018.

Anthropic’s reported numbers are smaller but follow the same shape. Documents shared with investors and reported by the Wall Street Journal in November 2025 projected cash burn at about one-third of revenue in 2026, falling to roughly 9% in 2027, with breakeven targeted around 2028. Those are projections, not results.

Further down the stack, most AI application startups have no visible path to profitability at all. Stanford counted 1,953 newly funded AI companies in the U.S. in 2025. AI accounted for close to 87% of all U.S. venture funding by spring 2026. Many of those companies are raising at valuations that assume they will hold pricing power in a market where the underlying models are becoming cheaper and more interchangeable every quarter.

The Case That AI Is Not a Bubble

The case that AI is not a bubble is not backed by hard evidence. It requires ignoring a lot of assumptions and historical precedents.

The Revenue Is Real, At Least at the Infrastructure Layer

Nvidia reported $81.6 billion in revenue for the quarter ended April 26, 2026, up 85% from a year earlier. Data center revenue alone was $75.2 billion, up 92%. GAAP net income was $58.3 billion. Free cash flow was $48.6 billion. The company guided to $91 billion for the following quarter.

It is cash from paying customers, at gross margins near 75%. Whatever else is true about the AI cycle, the company selling the picks and shovels is booking the money.

Valuations, while high, are also not at 2000 levels. LPL Financial’s Jeff Buchbinder notes that the Nasdaq-100 is up roughly 140% since ChatGPT launched, against a 1,090% gain between Netscape’s IPO and the March 2000 peak. 

Tech trades at about 25 times forward earnings today. At the 2000 peak, it was 58 times. Expensive, not absurd.

AI Is Delivering Measurable Productivity Gains

Task-level studies have been consistent. Controlled experiments on customer success agents, writing tools, and software developers show productivity gains in the range of 15% to 55% depending on the task, with the largest gains for less experienced workers. Stanford’s AI Index estimates the value generative AI delivers to U.S. consumers at about $172 billion a year as of early 2026, up 54% in a year. Organizational adoption reached 88% in 2025.

The honest caveat is that these gains do not yet show up clearly in macro productivity statistics. They are real at the firm and task level and still hard to see at the GDP level. That was also true of computers for most of the 1980s and 1990s. But it is what separates AI from a pure speculative mania: the product works, people use it, and it changes output.

The Spending Came From Strength, At Least Until Recently

The biggest AI spenders are the most profitable companies in history. Microsoft, Alphabet, Amazon, and Meta funded the first two years of the AI buildout almost entirely from operating cash flow. In the late 1990s, telecom and dot-com companies were spending borrowed money on infrastructure for revenue that did not exist. That’s the core structural difference between investments behind dot-com and AI. 

Is AI a bubble? It depends on which layer you’re looking at

AI is not one asset class. It is a stack, and bubble risk is distributed very unevenly across it. Lumping Nvidia’s data center business in with a seed-stage AI note-taking app is how you end up either dismissing real risk or panicking about fake risk.

Here’s a table to give an overview of how to look at bubble risk for specific layers of the AI stack and why:

Risk rises as you move away from realized cash flow and toward projected cash flow. The question of whether AI is a bubble is too blunt to be useful. The real question is which part of AI, and at what price.

What Would Actually Pop the AI Bubble, and What to Watch For

Debating whether there is a bubble is less useful than knowing what a bursting one looks like in advance, so you can get your eggs in a row. Analysts and researchers have identified five risk signals across this cycle that you can watch out for. 

1. Capex outpacing free cash flow

This is the one that has moved the most in 2026. Epoch AI’s analysis of SEC filings found that aggregate cash capex across Microsoft, Amazon, Alphabet, Meta and Oracle is growing about 70% a year against operating cash flow growth of about 23%, with the two lines crossing around Q3 2026. Oracle crossed first and now funds a hyperscaler-scale buildout from a BBB-rated balance sheet with negative free cash flow. Alphabet reported its first quarter of negative free cash flow since its 2004 IPO and raised $84.75 billion in equity in June 2026, the largest equity raise by a listed company on record. Hyperscaler investment-grade bond issuance was $108 billion in 2025, about 26% of capex. Goldman expects roughly $250 billion in 2026 and $400 billion in 2027, taking debt to about 35% of capex. The “funded from earnings, not debt” argument is weakening in real time. Watch out for the bond spreading.

2. Revenue growth stalling at the infrastructure layer

Nvidia’s data center revenue is the cleanest read on whether the buildout is ahead of demand. If it plateaus or guidance gets cut, the entire stack above it reprices. Note that Nvidia’s stock has slid after four consecutive earnings beats. The market is already pricing in some deceleration.

3. Startup funding is drying up. 

With AI at close to 87% of venture dollars, a pullback in VC appetite would not be a sector correction. It would be the venture market correcting. The first sign is usually down rounds at well-known names, not failures at unknown ones.

4. Enterprise ROI disappointment at scale. 

MIT’s 2025 GenAI Divide study found that about 95% of organizations investing in generative AI were seeing no measurable return on the P&L. Stanford’s 2026 Index shows adoption at 88%, but agentic deployment is still rare. If the narrative that spending millions and billions does not guarantee profit hardens in boardrooms, enterprise AI budgets get cut, and the revenue projections underneath the model companies and cloud providers stop being achievable.

5. Rate pressure. 

The Fed held rates at 3.5% to 3.75% in July 2026, and the 10-year Treasury yield reached 4.69% in August. Bubbles historically burst after extended tightening because higher rates compress growth multiples and make debt-funded capex expensive to carry. Reuters reported in August that spreads on hyperscaler bonds had already widened meaningfully versus 2025. That is a small move. It is the direction that matters.

WorkflowFiesta Was Built for the AI Era That Survives the Hype

Whatever happens to AI valuations, the companies that come out the other side are the ones that can point to measurable results from AI today. Not the ones that bought the story.

That is the thing the bubble debate keeps circling back to. The MIT finding that 95% of enterprise AI investment produces no measurable return is not a statement about the technology. It is a statement about how it gets deployed: scattered pilots, no process ownership, no way to see what changed. The 5% that got returns did something different. They put AI inside specific workflows, measured the output, and expanded what worked.

WorkflowFiesta is built for that 5%. It is an AI orchestration platform that lets teams automate real processes end-to-end, track what each workflow actually produces, and justify AI spend with numbers rather than sentiment. If the AI cycle corrects, that is the position you want to be in: spending on what demonstrably works and being able to prove it.