The AI Funding Bubble Has a Leak
After two years of record valuations, AI startup funding is showing the first signs of a correction. Not a collapse — a repricing of what intelligence is actually worth.
Not every AI company is worth what it was eighteen months ago.
That sentence would have been controversial in 2024. In mid-2026, it is becoming consensus — quietly, without the dramatic collapse that bears predicted, but unmistakably.
The numbers behind the shift
Global AI startup funding peaked at $112 billion in 2024. The first half of 2026 is tracking toward $67 billion annualised — still historically elevated, but down 40% from the peak.
More telling than the headline number is where the money is going. Mega-rounds for foundation model companies have slowed sharply. The action has moved downstream — to vertical AI applications with demonstrable revenue, clear retention metrics, and a path to profitability that does not depend on another funding round.
Investors are asking a question they largely skipped in 2024: what does this company look like when the model costs approach zero?
The commoditisation problem
OpenAI, Anthropic, Google, and Meta are in a race that only ends one way — with inference costs near zero and model capabilities largely commoditised at the application layer.
That is good for users. It is existential for startups whose entire moat was “we use GPT-4 better than our competitors.”
The companies that survive this transition have one of three things: proprietary data that improves with scale, a workflow so deeply embedded that switching costs are prohibitive, or a distribution advantage that pre-dates the AI wave.
Everything else is a feature, not a company.
Where the money is actually going
Enterprise AI infrastructure is still attracting serious capital. The picks-and-shovels trade — GPUs, networking, cooling, power — remains crowded and expensive.
Healthcare AI is seeing a second wave of investment, driven by FDA clearances that create defensible regulatory moats. Legal AI, similarly, benefits from a profession that is slow to change but enormous in aggregate spend.
Consumer AI is the graveyard. High acquisition costs, low retention, and a user base that switches to whatever model is newest. Building a durable consumer AI business has proven harder than almost anyone expected.
The bottom line
This is not 2001. The technology is real, the revenue is real, and the enterprise adoption is accelerating.
But valuations priced in a winner-take-all dynamic that is not materialising. The correction is a repricing — of moats, of margins, and of what it actually means to build something defensible in a world where the underlying intelligence is becoming a commodity.