The Difference Between a Bubble and Overvaluation
Not everything that deflates is a bubble. Overvaluation is common in technology markets and does not always produce a crash. Sometimes prices normalise slowly as earnings catch up to expectations. Sometimes they correct sharply and then recover. The word "bubble" implies a specific dynamic: widespread belief in a story that cannot survive contact with operational reality.
The signals below are not proof of a crash. They are evidence that the story gap, the distance between AI's current operational performance and the valuations and expectations built around it, is widening in specific, measurable ways. Each indicator on its own is ambiguous. Together they suggest a correction is more likely than it looked eighteen months ago.
Indicator One: Enterprise Renewal Rates Are Softening
The 2024 AI spending wave produced a large cohort of 12-month enterprise contracts for AI tools. Those contracts are now coming up for renewal. And the renewal rates are not matching the initial adoption enthusiasm.
Companies that signed on to AI productivity tools in 2024 are finding that actual usage is lower than projected, that the productivity gains are harder to measure than the sales pitch suggested, and that internal champions who pushed for the tools are facing procurement scepticism from finance teams who want proof of return on investment before committing to another year.
This shows up in earnings commentary. Software companies are talking more carefully about "active users" versus "seats sold" and about "expansion revenue" being slower than expected. That is corporate language for: people bought it, not everyone is using it, and fewer are paying for more.
Indicator Two: Application Layer Funding Is Collapsing
AI venture funding has not stopped. It has concentrated. The money is going to frontier model development (OpenAI, Anthropic, xAI) and to infrastructure (Nvidia, data centres, energy). The application layer, AI SaaS companies building on top of the frontier models, is seeing a significant pullback from 2024 peaks.
Series A rounds for AI application companies are down substantially from where they were at the height of the market. The logic for the pullback is straightforward: if the underlying models keep improving and their prices keep falling, the moat for any application built on top of them is thinner than it looked. Why pay a premium for an AI-powered contract review tool when the model it is built on is available via API at a fraction of the cost, and the wrapper is not that complicated?
Investors funding the application layer in 2023 and 2024 were implicitly betting that the models would plateau and that application companies would develop durable advantages. That bet is looking shakier as model improvement continues and API pricing falls.
Indicator Three: Copilot Products Are Struggling
The "copilot" product category, AI assistants embedded in existing enterprise software, was supposed to be the mass adoption mechanism. Pay a small monthly premium on your existing software subscription, get AI capabilities built in. Low friction. Broad distribution.
The reality has been more difficult. Microsoft Copilot adoption in enterprise is below internal projections. The company has not released specific figures, but public commentary from enterprise customers, analyst channel checks, and Microsoft's own shifting messaging around Copilot all point in the same direction: the product is being used by early adopters and AI enthusiasts within organisations, but broad employee adoption is not happening at the rates the product roadmap assumed.
Salesforce Einstein tells a similar story. Salesforce has made significant public commitments about AI feature adoption, including a highly publicised set of claims about "Agentforce" deployments. The gap between public announcements and independently verifiable adoption data is wide enough that it has attracted scrutiny from analysts.
The pattern across both companies: the product works. The demos are convincing. Actual daily use by typical employees is much lower than the market opportunity projections assumed.
Indicator Four: Model Prices Are in Freefall
OpenAI, Anthropic, and Google have all cut API prices by 60 to 90 percent over the last eighteen months. That rate of price decline is striking even by the standards of the semiconductor industry, which has historically followed a predictable cost reduction curve.
For developers and companies building on these models, falling prices are genuinely good news. Access to frontier AI capability is becoming cheaper. The cost per token is approaching a level where many use cases that were previously uneconomical become viable.
For investors, the picture is different. The valuations of the frontier AI companies were built on assumptions about durable platform economics: a small number of powerful models that enterprises would depend on, creating pricing power and high switching costs. If models are commodities, and the price trajectory suggests they are becoming commodities, those platform economics do not hold. A company worth $150 billion because of its position in a market where margins are collapsing is a different investment thesis than the one that supported that valuation.
The price collapse is not a sign that AI is failing. It is a sign that the market structure the valuations assumed may not be forming the way the bulls expected.
Indicator Five: AI-Narrative Stocks Are Underperforming
Companies that rebuilt their public narratives around AI in 2023 and 2024 have had a difficult time since. UiPath, which pivoted its robotic process automation story toward AI, has seen its stock price fall substantially from its AI-hype peaks. C3.ai, an enterprise AI company with a long history of aggressive revenue recognition claims, trades at a fraction of its 2021 highs. BigBear.ai, a defence AI company, has struggled to translate its positioning into durable revenue growth.
These are not the large-cap AI names. Nvidia, Microsoft, and Google are different companies with different fundamentals. But the performance of the smaller, purer-play AI stocks is worth tracking because they reflect market confidence in the AI application layer thesis without the diversification buffer of a large platform company.
When the companies most exposed to AI adoption struggle, it is a signal about where the market sees the risk concentrated.
There is also a pattern in how these companies communicate with investors. In 2023 and 2024, earnings calls were heavy with AI-forward language, forward-looking deployment stories, and customer pipeline announcements. In 2025 and 2026, the language has shifted: more caution about timelines, more emphasis on proof-of-concept rather than production deployment, and more hedging on when AI features will contribute to revenue. That shift in tone, across multiple companies, is itself a signal worth noting.
A Note on Timing
One of the challenges with bubble analysis is that early deflation signals look identical to a healthy market correction before an acceleration. Enterprise software always has a lag between initial contracts and mature adoption. Renewal rates dipping in the first cycle can reflect a product maturity curve, not a failure. Funding concentration is normal in any technology market that is consolidating around clear leaders.
The reason these indicators are worth tracking now, rather than dismissing as normal market dynamics, is their simultaneity. All five are visible at the same time, across different parts of the market, pointing in the same direction. Any one of them alone would be unremarkable. Together they suggest the market is working through the gap between 2023 and 2024 expectations and 2025 and 2026 reality.
What Is Not a Deflation Signal
The indicators above are real, but they need context. Several of the most-cited "bubble" arguments are not well-supported by the current data.
Nvidia's revenue is still growing. The company's data centre business, which is the cleanest measure of actual AI infrastructure investment, continues to expand. If the AI bubble were deflating at the infrastructure level, Nvidia would be the first place it showed up. It has not shown up there yet.
Enterprise cloud spending is still growing. Amazon Web Services, Microsoft Azure, and Google Cloud are all seeing continued growth in infrastructure revenue. Companies are still spending on the compute and storage that AI workloads require.
Model capability is still improving. The argument that the technology has hit a wall is not supported by what the labs are shipping. The models available now are meaningfully more capable than what was available eighteen months ago.
The honest picture: the deflation signals are concentrated in the application layer and in enterprise adoption rates. The infrastructure layer is still growing. The technology is still improving. This is consistent with a bubble in valuations and expectations, not a bubble in the underlying technology.
The internet bubble produced Amazon and Google. The deflation of AI application layer valuations does not mean AI is not real.
It means the current valuations priced in a future that is arriving more slowly and with lower margins than the market assumed.
That gap is worth watching carefully.