The AI Bubble Argument Is Wrong. Here Is Why.

Photo by Egor Komarov via Pexels

NVIDIA's FY2026 revenue hit $215.9 billion, up 65% year on year, with Q4 alone generating $68.1 billion and GAAP gross margins sitting at 75%. Its market cap reached approximately $4.3 trillion by February 2026, making it the most valuable company on earth. Critics call this a bubble. They are wrong, and the data makes the case clearly. When a company selling physical silicon to paying enterprise customers generates those numbers, the word "bubble" requires a far more precise definition than most commentators are applying.

The bubble argument has gained traction for understandable reasons. AI stocks have materially outperformed the broader market for several consecutive years, valuations on private foundation-model companies are eye-watering, and the circular flow of capital between hyperscalers, AI startups, and GPU vendors has drawn legitimate scrutiny. Goldman Sachs put the question formally. WIRED applied a classic bubble framework to AI across five dimensions and scored it 8/8. But framework-matching is not the same as causal analysis, and conflating market exuberance with structural instability is a category error that has misled investors before. The dot-com bubble is the wrong reference frame here, and continuing to reach for it is analytically lazy.

The most important structural difference is this: AI's economic model demands continuous compute, not one-time infrastructure spend. Pets.com needed a website and a warehouse. An enterprise deploying large language models at scale needs inference compute every single time a query runs, every hour, every day, indefinitely. That is a recurring revenue engine, not a build-and-forget capital project. Bloomberg Businessweek's January 2026 analysis put it plainly: the current AI cycle shows broad-based productivity gains already translating into higher profitability, expanding margins, and improving cash flow dynamics across adopting firms. Real productivity gains generating real cash flows are the literal opposite of bubble anatomy.

AI's economic model runs on continuous inference demand, not one-time infrastructure spend. That single structural fact makes the dot-com comparison obsolete.

The circular financing critique deserves an honest answer rather than dismissal, but it also deserves precision. Yes, hyperscalers invest in AI startups, those startups spend on cloud compute, and cloud revenue flows back to hyperscalers. That circularity exists. But as one careful analysis noted, circularity is an amplifier, not a cause. Vendor financing of this kind appeared in GE Capital, in Cisco's aggressive capital deployment during the 1990s, and across auto lending for decades. The relevant question is whether the underlying inference demand being amplified is genuine. Jerome Powell, not historically a man given to tech boosterism, acknowledged publicly that AI companies are generating real revenue. The circular structure makes a potential correction worse if demand collapses, but it is not itself evidence that demand is fictitious.

A more honest characterisation of the current market, offered by a multi-method evaluation published in May 2026, is that AI displays bubble-like characteristics in specific segments while showing substantial fundamental support in others. Highly valued private foundation-model firms with no clear monetisation path, thin-margin AI application startups trading on narrative, and speculative data-centre projects in secondary markets warrant genuine caution. But the largest public AI infrastructure companies hold bottleneck rents, realised revenue, and high margins that justify their positions on conventional valuation frameworks. The market is not monolithic. Treating every AI-adjacent asset as equivalent is the analytical mistake, not the investment itself.

For founders and investors, the actionable read is straightforward. The infrastructure layer, companies selling picks and shovels to an insatiable compute market, is not where the bubble risk lives. The risk concentrates in application-layer companies that are pre-revenue or margin-free, betting that distribution and stickiness will arrive before their runway does. The scientific uncertainty around AI scaling laws is real and should inform how far out you're projecting returns, but uncertainty about future capability does not negate present utility. Founders building on top of AI infrastructure today are not building on sand. They are building on the most capital-dense, rapidly improving platform in the history of enterprise technology.