Let me be upfront about my bias: I find the word “bubble” both useful and abused. Useful, because asset bubbles are real and they hurt people. Abused, because “it’s a bubble” has become the laziest critique in technology — a way to sound wise without doing analysis. So when I ask whether AI is in a bubble right now, I want to do it properly: define the term, check the evidence, separate the froth from the foundation, and give you the tools to judge for yourself.
The short version of my answer: parts of the AI economy show classic bubble warning signs, and parts show genuine durable value. The interesting question isn’t “bubble or not” — it’s which parts are which. For the historical pattern of what happens when the froth clears, see my guide to the AI winter.
Table of Contents
- What “Bubble” Actually Means
- The Case That It’s a Bubble
- The Case That It Isn’t
- The Dot-Com Mirror
- Froth vs. Foundation: A Field Guide
- Five Signals I’m Watching
- What a Pop Would Actually Look Like
- How to Think About It (Without Panicking)
- Frequently Asked Questions
What “Bubble” Actually Means
Precision matters here, because most bubble talk is sloppy. An economic bubble has a specific meaning: asset prices driven primarily by expectations of future price increases rather than by the asset’s underlying cash flows or utility. The tulip buyer wasn’t buying flowers; he was buying the expectation of selling to a greater fool. The dot-com investor wasn’t buying earnings; she was buying eyeballs that might someday become earnings.
By that definition, “is AI a bubble” decomposes into smaller questions. Are AI company valuations supported by current revenues? Is AI infrastructure spending supported by current demand? Are AI startup valuations supported by plausible paths to profitability? Some of these get uncomfortable answers. Others don’t.
There’s also a second, looser meaning people intend: “there’s too much hype and it’s going to end badly.” That’s a mood, not an analysis. Moods can be right, but they can’t be argued with. I prefer the precise version — it gives us something to actually check.
The Case That It’s a Bubble
Steel-manning the bubble case, because it deserves respect:
Valuations have detached from revenues. Multiple AI companies carry valuations that imply flawless execution of extremely ambitious plans. The math only works if AI transforms large parts of the economy within a few years. That’s possible. It’s also exactly what every bubble’s math has assumed.
The spending is circular. This is the sharpest bubble argument. Company A raises billions, spends it on chips from Company B, which books it as revenue and invests in Company C, which buys cloud capacity from Company A. Follow the money around the circle and some meaningful fraction of “AI revenue” is AI companies selling to each other. Circular revenue isn’t fake — the chips are real — but it inflates the apparent size of the market. When the music stops, the circle breaks everywhere at once.
The promises keep escalating. We’ve moved from “useful coding assistant” to “automate most knowledge work” to “transform civilization” in remarkably little time. Each escalation raises the bar the technology must clear to justify current prices. Bubbles are expectation machines; the AI boom is manufacturing expectations at industrial scale.
Concentration is extreme. A handful of companies, a handful of chip suppliers, a handful of cloud providers. The entire edifice rests on a narrow base. Narrow bases are fragile — one supply shock, one regulatory action, one spectacular failure propagates everywhere.
The history of artificial intelligence says booms end. Twice before, AI investment collapsed when promises outran delivery. The people saying “this time is different” are making the four most expensive words in finance do a lot of work.
The Case That It Isn’t
Now the other side, equally steel-manned:
The technology generates real revenue now. This is the single strongest anti-bubble fact. AI coding tools, customer service automation, search features, content generation — these aren’t demos, they’re products with paying customers and measurable ROI. The dot-com bubble’s fatal flaw was revenue-free valuations. AI’s valuations may be high, but they’re not revenue-free.
Adoption curves are real. The fastest technology adoption in history isn’t a survey artifact — it’s hundreds of millions of users choosing to use these tools. Bubbles can inflate prices, but they can’t fake that many users finding that much value. Usage is the hardest metric to manipulate.
The infrastructure spending has a physical basis. Data centers, chips, energy — this isn’t financial engineering, it’s capital expenditure on tangible assets. Even in a downturn, the assets exist and retain value. Contrast with bubbles built on derivatives of derivatives.
Productivity effects are measurable. Studies of AI-assisted work show genuine productivity gains in specific tasks — coding, writing, customer support, data analysis. The gains are narrower than the hype claims, but they’re real and they’re already in the economic data.
Every previous “AI bubble” call was early or wrong. Skeptics have been calling the top since 2023. The technology kept improving and adoption kept growing. Being early is indistinguishable from being wrong when there’s no deadline.
The honest synthesis: the bubble case is about prices; the anti-bubble case is about the technology — and both can be true at once.
The Dot-Com Mirror
The dot-com era is the inevitable comparison, so let’s make it precise rather than vibey.
Similarities: massive capital inflows, valuations detached from current earnings, every company rebranding around the technology, predictions of total economic transformation, retail enthusiasm, circular investment patterns.
Differences: the internet in 1999 had a fraction of today’s AI adoption; AI products generate revenue far earlier in the cycle; the infrastructure being built (compute, energy) has clearer durable value than pet-food delivery websites; and the technology’s capability curve is steeper — the internet didn’t get 10x better every two years.
The most useful dot-com lesson isn’t “bubbles pop” — it’s “the pop doesn’t kill the technology.” The dot-com crash destroyed trillions in paper wealth and killed thousands of companies. It did not kill the internet. Amazon, Google, and the entire digital economy were built on the infrastructure and lessons of the wreckage. If AI is in a bubble, the pop would redistribute value, not erase the technology. AI accountability work, safety research, and applied deployments all continue through corrections.
The technology’s trajectory and the market’s trajectory are related but not identical.
Froth vs. Foundation: A Field Guide
Since “bubble or not” is too coarse, here’s how I separate froth from foundation in practice:
Froth indicators: valuations requiring flawless execution of the most optimistic scenarios; revenue that’s mostly AI companies selling to each other; startups whose pitch is “AI for X” with no defensible advantage; timelines to transformative impact measured in months; anyone using the phrase “this time it’s different” unironically.
Foundation indicators: products with renewing enterprise contracts; measurable productivity gains in published studies; infrastructure spending tied to demonstrated demand; open-source models improving steadily; talent flowing toward real problems rather than hype-chasing.
Apply this to any AI company or claim and you get a quick read. The uncomfortable finding: most of the big players show both. The same company can have a rock-solid product business and a bubbly valuation. The froth sits on top of the foundation — which is why the question is hard and why simplistic answers are wrong.
My personal heuristic: the further a claim is from a paying customer, the more froth it contains. “This tool saves our support team 30% of ticket time” — foundation. “This will automate all knowledge work by 2027” — froth. Work backward from revenue and the picture clarifies fast.
Five Signals I’m Watching
If you want to monitor the bubble question yourself, these are my five gauges — concrete, checkable, and leading rather than lagging:
1. Enterprise renewal rates. Pilots are bought with enthusiasm; renewals are bought with ROI. If companies quietly decline to renew AI contracts after the pilot year, the demand story is weaker than the revenue suggests. This is the single most important number nobody publishes.
2. The circularity ratio. What fraction of AI revenue comes from AI companies buying from each other? The higher the ratio, the more fragile the edifice. Watch for disclosures, analyst estimates, and the tone of earnings calls.
3. Startup mortality at seed stage. Bubbles die from the edges inward. When early-stage AI startups can’t raise — not the famous ones, the ordinary ones — the ecosystem’s immune system is failing. The giants always raise last.
4. Capex discipline. Are the big infrastructure bets being adjusted to demand, or is spending on autopilot? Disciplined capex suggests management sees real demand curves. Autopilot spending suggests momentum investing by the companies themselves.
5. The vocabulary test. When insiders start qualifying the hype — “AI is useful for specific tasks” replacing “AI changes everything” — the culture has registered a turn. Language leads; prices follow.
None of these is a crystal ball. Together, they’re a dashboard — and right now the dashboard shows a mix of green and amber, not red.
What a Pop Would Actually Look Like
Suppose the bubble case is right. What happens? Not the apocalypse — the redistribution.
Phase one: the funding freeze. Venture funding for AI startups collapses. The famous companies raise down-rounds or don’t raise at all. This happens fast — quarters, not years.
Phase two: the consolidation. Big tech buys the survivors at fractions of their peak valuations. Talent disperses from failed startups into acquirers and new ventures. The technology doesn’t die; its ownership concentrates.
Phase three: the quiet building. With hype gone, the remaining players focus on products, unit economics, and real customers. This is historically when the most durable value gets created — Amazon’s real growth came after the dot-com crash, not before it.
What doesn’t happen: the models don’t get deleted, the knowledge doesn’t evaporate, the deployed products don’t stop working. A pop destroys paper wealth and hype-dependent companies. It doesn’t destroy capability. Anyone telling you a crash would “set AI back decades” is confusing the market with the technology.
How to Think About It (Without Panicking)
My advice for navigating the bubble question, whether you’re building, buying, or just watching:
Separate the technology from the market. The technology’s trajectory — capabilities improving, costs falling, adoption growing — is the more important and more durable story. Markets overshoot in both directions. Don’t let market volatility change your assessment of the technology, and don’t let technological enthusiasm blind you to market risk.
Work backward from customers. For any AI claim, ask: who pays, for what, and do they renew? Customer-backed value survives pops. Narrative-backed value doesn’t.
Build for the trough. If you’re building in AI, assume funding gets harder and hype gets thinner. Build products with real ROI, keep burn manageable, and don’t depend on the next round arriving on schedule. The companies built for the trough are the ones that own the recovery.
Keep the receipts. Write down the big predictions — yours and others’. Check them later. The bubble debate generates endless confident claims; the people worth listening to are the ones whose past claims survived contact with reality.
Is AI in a bubble right now? My answer: the prices show bubble warning signs, the technology doesn’t. That combination usually resolves with a correction, not a catastrophe — and the correction, when it comes, will be the market catching up to reality, not reality catching up to the market. Plan accordingly.
Frequently Asked Questions
Parts of the AI economy show classic bubble warning signs — valuations detached from revenues, circular spending between AI companies, and escalating promises. But the underlying technology generates real revenue and adoption, which distinguishes it from pure bubbles. The prices may be bubbly while the technology is real.
AI products generate revenue much earlier in the cycle, adoption is faster and broader, the infrastructure being built has clearer durable value, and the technology’s capability curve is steeper. The dot-com crash killed companies but not the internet — an AI correction would likely redistribute value rather than erase the technology.
Key warning signs include enterprises failing to renew AI contracts after pilots, a high ratio of AI companies selling to each other, early-stage AI startups unable to raise funding, undisciplined infrastructure spending, and insiders quietly tempering their public claims.
A pop would likely mean a venture funding freeze, consolidation as big companies acquire survivors cheaply, and a shift to building real products with solid economics. Paper wealth would be destroyed, but models, knowledge, and deployed products would survive — capability doesn’t evaporate with valuations.
Not much. If you use AI tools for their actual current capabilities and measurable value, a market correction changes little for you — the tools keep working. The risk is mainly for investors and for companies whose business depends on ever-increasing funding rather than customer revenue.
