Is AI a Bubble? The Real Numbers Behind the Debate

What Does "AI Bubble" Actually Mean?

A bubble is not a judgment on whether a technology works. It describes a specific mismatch: asset prices and spending running ahead of the revenue and profit that can currently support them, driven partly by hype and fear of missing out rather than only by proven demand. The dot-com bubble did not mean the internet was fake. It meant Pets.com was valued like it would replace every pet store in America before it had the revenue to support a fraction of that.

The AI bubble debate works the same way. The useful move is to split it into two separate questions instead of treating "AI" as one homogeneous thing.

The bifurcation view

Split the AI stack into two layers and the picture gets a lot less confusing:

This split matters because when someone says "AI is a bubble," they are usually only talking about the top row. The bottom row, the actual software people pay for every month, mostly does not show bubble characteristics. It shows normal, if fast, software adoption.

The Numbers Driving the Bubble Fear

The single most cited number in the bubble debate came from Sequoia Capital partner David Cahn in 2024. He worked backward from GPU spending to estimate that the AI industry needed approximately $600 billion in annual revenue to earn a reasonable return on the infrastructure being built. At the time, he estimated actual AI revenue across the industry at around $100 billion, a gap of roughly $500 billion.

That estimate was controversial when Cahn published it. It has aged into something closer to a floor than an exaggeration.

"AI companies would need about $600 billion in annual revenue to justify current AI infrastructure investment levels." (David Cahn, Sequoia Capital, 2024)

The reason the gap matters more now than it did in 2024 is simple: the infrastructure side of the equation did not slow down. It accelerated, even as the underlying question of how AI companies actually make money remains far from settled. The next section shows exactly how much the spending side has grown.

Who Is Spending What: The 2026 Capex Numbers

Five companies account for most of the infrastructure spending that drives the bubble conversation, and their 2026 capital expenditure plans, reported between late 2025 and February 2026, make the data center buildouts of even three years ago look small.

Added together, that is $660 billion to $690 billion in a single year, from five companies, mostly on data centers, GPUs, and the power infrastructure to run them. Synergy Research Group, which independently tracks hyperscaler capital spending, has documented this same acceleration in cloud infrastructure investment over the past several years.

Nvidia CEO Jensen Huang has been the most direct about where this is headed. In comments reported in February 2026, he estimated total AI infrastructure spending would reach $3 trillion to $4 trillion by the end of the decade. Nvidia's investor relations page carries the company's own financial disclosures on the data center demand behind that estimate. Meta CEO Mark Zuckerberg has said Meta alone plans to spend $600 billion on US infrastructure through 2028.

"Between $3 trillion and $4 trillion will be spent on AI infrastructure by the end of the decade." (Jensen Huang, Nvidia CEO, Feb 2026)

One detail worth understanding if you want to know where this money actually goes: most of it buys Nvidia GPUs, the same H100 and Blackwell-generation chips priced at $25,000 to $40,000 per unit that underpin nearly every large AI training run in production today.

Revenue vs Valuation: OpenAI and Anthropic by the Numbers

This is where the bubble argument gets concrete instead of abstract. Two companies sit at the center of frontier AI, and both have reported revenue figures that can be checked against their valuations.

OpenAI reported $10 billion in annualized revenue in June 2025, rising to about $13 billion annualized by August 2025. Its valuation was reported at $300 billion in March 2025, roughly 23 times its later revenue run rate.

Anthropic reported $4 billion in annualized revenue in July 2025, rising to $7 billion annualized by August 2025. Its valuation was reported at $61.5 billion in March 2025, with later 2026 reporting describing talks around a $170 billion valuation, which would put it near 24 times revenue.

The Number Most Guides Don't Show

Here is the calculation that puts the whole debate in perspective. In 2024, Sequoia's David Cahn set $600 billion in annual revenue as the bar the entire AI industry needed to hit to justify its infrastructure spending. By 2026, five companies alone are projected to spend $660 billion to $690 billion on infrastructure in that single year, more than Cahn's entire industry-wide revenue bar, and that figure is capital expenditure, not revenue.

The gap Cahn identified did not close. It grew, and it grew before most of that new infrastructure had even finished being built and put to work. Whether that turns out to be a bubble or simply a very large, very early bet depends entirely on whether usage and revenue catch up before the depreciation clock on all those GPUs runs out.

AI vs the Dot-Com Bubble: What Is Actually Different

The comparison to the dot-com bubble comes up constantly, and it is worth checking against real numbers instead of vibes.

The real difference analysts point to is not whether the spending is large. It clearly is, larger in absolute terms than the dot-com buildout. The difference is who is funding it. Dot-com infrastructure was financed heavily by debt and speculative equity in companies with no profit. The current AI buildout is financed largely out of the operating cash flow of companies that were already making money before AI capex ramped up.

That does not make a correction impossible. It changes what a correction would look like. A dot-com-style collapse wiped out companies that never had a real business. A correction in AI infrastructure would more likely show up as write-downs and slower growth at companies that remain fundamentally profitable, closer to how hyperscalers absorbed the 2022 cloud spending slowdown than how Pets.com or Webvan disappeared entirely.

The Warning Signs Worth Watching

Three specific mechanics show up again and again in bubble-risk analysis, and they matter more than any single headline number.

• Circular deals: chipmakers, cloud providers, and model labs increasingly fund each other through intertwined contracts. Nvidia invests in a startup that buys Nvidia chips through a cloud provider Nvidia also has a stake in. Revenue can look larger than end-user demand actually justifies when money is moving in a loop between the same handful of companies.
• Vendor financing: when a supplier helps finance its own customer's purchases, reported demand can outrun real economic demand. This was a classic warning sign in the dot-com era and shows up again in some AI chip and cloud-capacity deals.
• GPU depreciation: data center GPUs are typically depreciated over 3-6 years on the assumption they stay useful that long. If newer chip generations make older GPUs obsolete faster than that, for example if a Blackwell-class chip makes a two-year-old Hopper-generation GPU commercially unattractive, the return on that capex compresses sharply, even if the company using it stays profitable overall.

Against those signals, the strongest counterargument is capacity constraint, not overbuild. Some analysts argue that power availability, chip supply, and data center construction timelines are still the bottleneck on AI growth, not lack of demand, which would mean the market is currently underbuilt rather than overbuilt. The truth is probably both are happening in different places at once: real compute scarcity in the parts of the world with power constraints, and speculative overbuild in a subset of announced projects that never get built out.

So, Is It a Bubble? What Would Actually Confirm It

The honest answer is that AI infrastructure spending shows real bubble characteristics right now: a large and growing gap between capex and revenue, valuations that assume years of future growth are already locked in, and financing structures (circular deals, vendor financing) that echo dot-com-era warning signs. AI as a technology and AI applications with real paying customers do not show the same characteristics. Confusing the two is the most common mistake in this debate.

What would actually confirm a bursting bubble, rather than just a large investment cycle, is a specific and checkable set of signals: hyperscaler capex growth slowing or reversing for two or more consecutive quarters, GPU utilization rates falling below the levels needed to justify current depreciation schedules, or a frontier AI lab needing a valuation cut of 50% or more in a funding round rather than a markup. None of those had happened as of early 2026. All three are worth watching over the next several quarters, and any one of them would be a much stronger signal than another headline comparing AI to tulip mania.

Frequently Asked Questions

Related Articles