Is there an AI bubble? The 2026 evidence, both sides

Searches for “AI bubble” have grown roughly fourteen-fold in a year, which tells you something before any analyst does: a lot of people holding AI exposure, in their portfolios or their budgets, are quietly asking the same question. The honest answer is that the evidence points both directions at once, and anyone who tells you the case is closed is selling a position, not an analysis.

So here is the case on both sides, in confirmed numbers, and the short list of signals that would actually settle it.

The case that this is real

The revenue is not imaginary. Nvidia booked $215.9 billion in its last fiscal year, up 65 percent, and its most recent quarter grew 85 percent year over year. Anthropic’s annualized revenue passed $30 billion this spring; OpenAI’s sits around $25 billion. The fastest AI startups now reach $100 million in revenue in 18 months, a pace the software industry has simply never produced before. And unlike 2000, the biggest spenders fund the buildout largely from profits: top technology stocks trade near 25 to 30 times earnings against roughly 55 times at the dotcom peak.

The case that it is a bubble

The gap between spending and revenue is the uncomfortable number: hyperscalers put roughly $400 billion into AI infrastructure last year against something like $100 billion in enterprise AI revenue, and 2026 spending guidance is far higher, a mismatch we broke down in our AI data center pillar. Private valuations price heroic futures: labs trade hands at 20 to 40 times revenue, and Cognition’s $26 billion valuation sits at roughly 53 times its run rate. Bloomberg has mapped a web of circular deals, chipmakers and clouds investing in labs that spend the proceeds right back on their investors’ products, that echoes the vendor financing of the telecom bust. The IMF’s chief economist has said plainly that the investment surge risks a technological bubble, and market concentration now exceeds the 2000 peak, with the top ten S&P 500 stocks near 35 percent of the index.

The synthesis most evidence supports

The most defensible reading is that this is a productive buildout wearing bubble-priced clothing. The infrastructure is real and will outlast any correction, the way dotcom-era fiber did. The risk is not that AI is fake; it is that specific valuations assume growth rates with no room for a single bad year, and that circular deal structures would amplify any demand disappointment. That is why we keep arguing that margin, not growth, is the real scoreboard: companies that own their unit economics survive repricing, and companies renting their economics do not.

For business buyers rather than investors, the practical translation is simple: negotiate contracts that survive vendor turbulence, per our agentic AI buyer’s guide, and treat any vendor whose pricing depends on continuous fundraising as a concentration risk, the theme of our capital stack explainer.

Update, July 2026: what the evidence did next

Four data points have landed since this piece was published, and they sharpen both sides of the argument rather than settling it.

On the exuberance side: startups raised a record $510 billion in the first half of 2026, and two companies took 43 percent of it. Concentration like that is what late-cycle capital looks like. One of those two, Anthropic, was repriced from $380 billion to $965 billion in roughly fourteen weeks.

On the real-demand side: the same Anthropic round came with a disclosed $47 billion revenue run rate, and TSMC, the company that manufactures nearly everyone’s AI silicon and has no reason to flatter anyone’s narrative, posted a record $39.6 billion quarter on AI demand. Foundry revenue is the hardest receipt the industry produces.

And the structural point that separates 2026 from 2000: the binding constraint has moved to physics. Memory supply and grid power, not GPUs, now cap the buildout. Bubbles typically deflate when supply finally overshoots demand; in AI infrastructure, supply cannot yet catch demand at any price. That does not make current valuations correct. It does mean the mechanism that usually ends these cycles has not engaged.

What to watch

Three signals would settle the question. A hyperscaler cutting capex guidance, since spending has only risen so far. Enterprise renewal rates at the fastest-growing startups, the difference between durable businesses and usage spikes wearing subscriptions. And the first major lab IPO, where public markets finally price what private rounds have only asserted, the same test hanging over both labs’ recent public repositioning.

Frequently asked questions