AI Software Bubble

AI software has the awkward profile of a market that is both obviously real and easy to overpay for.

OpenAI annualized revenue has been reported crossing the $20 billion mark and later prints in the mid-twenties. Cursor went from a November 2025 raise—$2.3 billion at about a $29.3 billion valuation on more than $1 billion of revenue—to reports of $2 billion-plus ARR and chatter about a much higher next round within months. Gartner 2026 forecasts put worldwide AI spending growth around 47% and AI software near $453 billion, up from roughly $283 billion the prior year. Those are not empty slideware numbers.

The narrative wrapping them is bigger still. Agents become the interface. Software stops assisting knowledge work and starts doing it. Today AI vendors capture a meaningful slice of white-collar economic value. That leap—from revenue is exploding to these specific companies own the residual—is where bubble math lives.

Unit economics do not look like classic SaaS. Inference costs scale with usage. Price cuts show up fast when models look interchangeable; Reuters reported OpenAI cutting developer pricing on a frontier model by more than 20% in August 2026 under competitive pressure from Anthropic and Chinese labs. Gross margin quality is therefore a moving target. Investors who import SaaS multiples onto metered intelligence are making a category error unless switching costs and distribution prove sticky.

There is a second knife edge. Gartner has estimated on the order of $234 billion of enterprise application spend at risk from agentic AI by around 2030—roughly a fifth of enterprise SaaS in that framing. AI can create a huge new category while chewing through the pricing power of the old one. Value might land with model labs, application-layer startups, data owners, or customers who simply pay less for the same work. AI wins does not specify the ticker.

Adoption surveys still show the gap between demos and P&Ls. McKinsey mid-decade State of AI work found most organizations stuck in experiment or pilot mode; a minority reported enterprise-level EBIT impact from AI, and for many of those the impact was under 5% of EBIT. Markets can capitalize the end-state years before CFOs can find it in earnings. Spending and pilots are not the same as durable productivity capture.

Frontier valuations make the discount rate explicit. Reuters reporting in August 2026 described Anthropic IPO chatter tied in part to roughly $190–200 billion of 2028 revenue against a then-current run rate near $47 billion. Maybe they get there. The price still embeds a narrow path on competition, cost, share, and pricing power. When markets price one future instead of a range, disappointments do not require the technology to fail—only to land in a merely-very-good scenario.

Reflexivity does the rest. Capital funds distribution, talent wars, and price subsidies. Growth validates the narrative. The narrative raises the next round. At first that can be productive: it accelerates real product development. Later high valuations become their own evidence that the future is inevitable. When financing tightens, the same loop runs backward and sorts durable products from capital-fueled ones.

None of this requires revenue to go to zero for investors to get hurt. A classic late-cycle path is slower growth, harder price competition, softer margins, and a re-rating of the terminal share story—while AI keeps spreading through the economy. That is the dot-com lesson in a new wrapper: the technology can remake workflows while investors who paid for perpetual scarcity and winner-take-most margins still lose money.

There is a clean bull path too. If agents become the default interface for knowledge work and a few platforms keep distribution and data advantages, today revenue leaders can grow into aggressive valuations. Spending forecasts already imply a very large industry. The open question is not whether AI software matters. It is how much of the surplus accrues to vendors versus customers once competition and switching costs do their work.

The software bubble case is not that AI is fake. It is that competition, commodity pressure on models, and early ROI evidence can transfer a lot of the surplus to customers—exactly while valuations assume the opposite.