The risk of an AI bubble concerns the sustainability of investment; it does not prove that the technology is useless. For a business, the practical decision starts with the problem to solve. Measurable outcomes and sustainable costs matter more than market enthusiasm.
- What does an AI bubble mean?
- Why do revenue and adoption not tell the whole story?
- What do current levels of AI investment indicate?
- Why do cheaper models not make every project simple?
- How can a business assess an AI project?
- What should you ask before choosing a provider?
Last updated: 2 November 2026.
What does an AI bubble mean?
A technology can be useful even when the market overestimates its future returns. The risk lies in the gap between expectations and outcomes.
Imagine a company building a large warehouse before it knows how many customers it will serve. The investment works if orders grow sufficiently. If demand fails to arrive, the costs remain. Warehouses can be useful without every warehouse project being successful.
The same principle helps explain the artificial-intelligence market. Companies are financing expensive infrastructure in anticipation of substantial future demand. This imbalance deserves scrutiny, but it cannot establish whether or when a market correction will occur.
Why do revenue and adoption not tell the whole story?
Revenue measures sales. It does not automatically show what remains after expenses. A service can attract many customers while still requiring investment that exceeds the revenue it generates.
AI services involve several significant costs:
- specialised processors and cloud infrastructure;
- electricity, cooling and connectivity;
- staff, security and maintenance;
- financing and interest.
The flow of money also matters. An application provider may pay a cloud platform using revenue from its customers. The cloud operator records revenue as well. Adding each transaction does not necessarily measure demand from end users.
This does not imply wrongdoing. It simply shows why sustainability must be assessed across the entire value chain.
What do current levels of AI investment indicate?
An October 2026 forecast published by Gartner highlights the scale of infrastructure investment and the potential risk of impairment by 2030.
An impairment reduces the accounting value of an asset when its expected returns no longer support that value. A risk forecast does not describe losses that have already occurred, nor does it establish that a crash is inevitable.
The Federal Reserve Bank of Richmond compares current conditions with telecommunications investment. In that period, capacity expanded faster than demand and contributed to a crisis; some of the infrastructure nevertheless found productive uses in later years.
The comparison separates the usefulness of a technology from the financial performance of every investment. It does not prove that AI will follow the same path.
Why do cheaper models not make every project simple?
The cost of achieving a given level of model performance can decline rapidly. Epoch AI examines this trend in The plunging price of thought, using standardised evaluations.
The price of an individual model response is only one part of a project. A business must also account for the work required to:
- prepare accurate, usable data;
- integrate the service with existing systems;
- monitor quality, errors and exceptional cases;
- protect information and access;
- train users and maintain the process.
A convincing demonstration is not enough. Value emerges in daily use, where incomplete information, unexpected requests and clear accountability all matter.
How can a business assess an AI project?
The useful question is not “where can we add AI?” but “which activity consumes time without producing proportional value?”. That question supports a limited, reversible test.
Consider a hypothetical example. An office receives 100 requests a day and spends an average of three minutes assigning each one to the correct department. The process takes 300 minutes.
A system suggests a destination while staff review the results and handle uncertain cases. If the total falls to 200 minutes, the apparent benefit is 100 minutes a day. Corrections, supervision and maintenance must all be included in that figure.
Which indicators should be measured?
- Total time before and after introducing the system.
- Percentage of requests handled correctly.
- Consequences and cost of errors.
- Cost of the service, integration and oversight.
- Time required for maintenance and updates.
This example does not document a real-world result; it demonstrates a method. In some cases, conventional rule-based automation may be simpler and less expensive. The technical choice should follow the problem.
What should you ask before choosing a provider?
A useful application can become fragile when prices, terms or availability change. Before adopting it, clarify:
- the problem it must solve and how success will be measured;
- who will review incorrect outputs;
- what data the provider will receive and where it will be processed;
- how costs will change as usage grows;
- how data can be exported and the service replaced;
- what procedure will apply during an outage.
A market correction would not make every AI application useless. Equally, enthusiasm does not make every project sustainable. The business standard remains practical: measurable outcomes, defined responsibilities and controlled costs.
Which activity in your organisation would you improve, and what result would prove that the change works?
I am Matteo Papparella, a senior developer specialising in PHP, backend development and custom WordPress solutions. If you would like to assess a digital process with a practical approach, please get in touch.