Is AI a Bubble? Six Arguments, Examined

The case that AI is a bubble rests on six arguments. Each sounds convincing on its own. Examined against the evidence, none of them holds up as a reason to expect a 2000-style collapse.

Is AI a Bubble?

Six arguments, examined against the evidence

1. "Spending far exceeds AI revenue"

The claim. Microsoft, Alphabet, Amazon, Meta and Oracle planned $660–690 billion of capital spending for 2026, nearly double 2025. At the start of the year, OpenAI's annualized revenue was about $20 billion, roughly 3% of that figure (Futurum).Why it falls short. Infrastructure is always built ahead of revenue; railroads, power grids and telecom networks all were. The test is whether the capacity is used, and today it is sold out. Microsoft reported an $80 billion backlog of Azure orders it could not fill because of power limits, and Alphabet's cloud backlog rose 55% in one quarter to more than $240 billion. The hyperscalers are also building for their own products and enterprise customers, not only for model labs. In 2000, much of the new fiber sat unused. Here, the constraint is the opposite.

2. "Circular financing is inflating demand, just like Lucent"

The claim. Nvidia invests in and finances its own customers. In July it was reported to be weighing a $250 billion financing guarantee for an OpenAI data center and a separate deal to finance $350 billion of OpenAI chip purchases (Axios). In 1999–2000, Lucent, Nortel and Cisco lent billions to telecom carriers to buy their equipment, and many of those loans went bad (Tunguz).Why it falls short. The comparison breaks on who is being financed. Lucent lent to young, leveraged carriers with no profits; 47 of them failed between 2000 and 2003. Nvidia's largest customers are the hyperscalers, which generated about $450 billion of operating cash flow in 2024 (Tunguz) and don't need vendor financing to buy chips. Nvidia lends from a 75% profit margin and large net cash, not a stretched balance sheet. The exposure is concentrated in one customer, OpenAI, which is a real risk for Nvidia, but it is not demand manufactured across the market.

3. "The build-out is running on debt"

The claim. JPMorgan estimates $4.1 trillion of the $5.5 trillion AI build-out through 2030 will be debt-financed, with loans averaging over 85% of project costs, much of it through off-balance-sheet vehicles (Fortune).Why it falls short. Debt is how large infrastructure has always been financed, and the main borrowers can carry it. JPMorgan projects hyperscaler operating cash flow above $900 billion by 2027 and describes the borrowing as a deliberate choice to finance while credit is cheap, preserving room to deleverage later. The highly leveraged projects exist, but they sit at the edges of the market, not at its core.

4. "GPUs wear out faster than the accounting says"

The claim. Hyperscalers depreciate servers over five to six years, while Nvidia ships a new chip generation every year. If the real useful life is closer to three years, reported profits are overstated.Why it falls short. Older chips don't stop earning when a new generation ships; they move from training to inference and other lower-intensity work, which is where demand is growing fastest. And while capacity is sold out, older hardware still has paying customers. Where the assumptions were too generous, companies are correcting them: Amazon shortened the life of some servers from six years to five (Tunguz). That is ordinary accounting judgment being revised, not hidden losses. It also changes reported profit, not cash flow.

5. "Valuations are at dot-com levels"

The claim. Nvidia's market value is around $5 trillion, and a handful of AI-linked companies dominate the stock market.Why it falls short. In 2000, companies like Cisco traded at well over 100 times earnings, and many internet companies had no earnings at all. Today's AI leaders are among the most profitable companies ever built; Nvidia's profit margin is around 75% (Axios). Concentration is a portfolio risk for investors, but high valuations backed by large, growing profits are not a bubble in the 2000 sense.

6. "Companies aren't getting a return on AI"

The claim. An MIT study in 2025 found that 95% of corporate generative AI pilots delivered no measurable P&L impact.Why it falls short. The study attributed the failures mainly to poor integration with existing workflows, not to the technology, which is typical of the early phase of any enterprise technology. The trend since is clear. Among software companies building AI products, AI is projected to reach 42% of revenue in 2026, up from 32% in 2025, while average gross margin on AI products is rising from 45% to a projected 53% (ICONIQ via SaaStr). Revenue and margins are moving in the same direction, which is the opposite of a bubble.

The verdict

AI is not a bubble. Demand exceeds supply, the largest buyers are highly profitable, and revenue and margins are rising. What deserves attention is narrower: a small number of highly leveraged builders and lenders whose returns depend on the most optimistic assumptions. If some of them fail, that would be a credit event at the edges of a sound market, not the collapse of the market itself.The telecom cycle is the right reference, and it is the industry I started in. The technology won; the fiber laid in the boom became the backbone of the internet. The losses fell on those who financed capacity with too much leverage for demand that wasn't there yet. This time, the demand is already there.Three signals would change that view: falling rental prices or idle capacity for the latest GPUs; several hyperscalers cutting capex guidance in the same quarter; and widening credit spreads on AI-related debt.

Frequently asked questions

Is AI a bubble in 2026?No. Demand for AI capacity exceeds supply, the largest buyers are highly profitable, and AI revenue and margins are rising. The real risk is concentrated in a few highly leveraged infrastructure projects and lenders.How is the AI boom different from the telecom bubble?In 2000, much new fiber capacity sat unused and the buyers were unprofitable start-ups financed by their suppliers. Today, AI capacity is sold out and the main buyers are among the most profitable companies in the world.Related reading: NVIDIA vs Cisco: What the AI Cycle Means for Valuations · NVIDIA vs Hyperscalers: Who Wins the AI Cloud?