# The AI bubble: how to buy software without getting stuck Author: Quadra URL: https://quadra-website.matt-madd.chatgpt.site/risorse/articoli/bolla-ai-come-comprare-software Published: 2026-09-18 Updated: 2026-09-18 Language: en Reading time: 4 min You do not need to predict whether the cycle will deflate. Your project needs to hold up either way. ## At a glance - You do not need to predict the next AI cycle. Choose a project that still makes sense if the market changes. - Assess the work the software solves, its total cost and your dependence on a single supplier. - Exportable data, documentation and an exit plan protect an investment more than a technology promise. In May 2026, the Federal Reserve published its Financial Stability Report. Risks flagged by market participants included AI-related equity valuations, increasingly debt-financed infrastructure investment, and possible labour-market effects. A year earlier, AI had been mentioned by a small minority of respondents. In the spring survey, that share had multiplied several times. Meanwhile, a question posed by a Sequoia investor keeps circulating: how much annual revenue would be needed to justify the investment made in AI infrastructure? According to those updating the calculation, the gap is widening, not closing. Two arguments usually follow. One says it is a bubble and will burst. The other says it is infrastructure investment, like fibre in the 1990s, whose effects will become visible later. For an eighty-person business considering an AI project, both arguments are unhelpful. The useful question is different: **how do you buy something today so that a market correction does not become your problem?** ## What actually happens if the cycle deflates AI does not stop working. The models remain, and probably become cheaper. What changes is the suppliers. In a correction, some startups close or are acquired, prices change, free plans disappear, and tools built for rapid growth have to start paying for themselves. Anyone who built their process inside one of those tools finds themselves renegotiating from a weak position, or rebuilding everything. It has happened before with cloud services, industry-specific business software and much of the business app market. The difference is that this time the cycle moves faster. ## Five questions to ask a supplier **What rights you have over the software.** There are three possible arrangements, and it is worth knowing which one you are signing. With a subscription, you have no ownership rights: the software belongs to the supplier and you use it as long as you pay and they exist. With full assignment, the code becomes yours and you can do what you want with it, but it is the most expensive option and not always the best. In between sits a stable licence, perpetual or multi-year: the supplier continues to develop and maintain the software, but you can keep using it even if you part ways, and you pay less than buying it outright. What matters is what you put in writing: that the licence does not expire when the relationship ends, that the data and configurations remain yours, and that you can access the code if the supplier closes. **Where the data lives, and how it gets out.** Data is where dependencies form. Ask before signing: in what format can I export it, how long will it take, what happens if I end the contract? If the answer is vague, the answer is no. **Whether the model can be replaced.** Models change every few months, with very different prices and capabilities. A well-designed system treats the model as a replaceable component, not a foundation. This is a technical choice made at the start that costs little; changing it later costs a lot. **Who wrote the system, and whether someone else can work on it.** Software built on widely used technologies can be taken over by another supplier. Software built inside a closed proprietary platform cannot. It is worth asking directly: if we stopped working together tomorrow, who could continue? **How long it takes to pay for itself.** A project that pays back in six or twelve months is a contained risk: if the context changes, you have already banked the value. A project that takes three years is betting on three years of stability, and right now that is a bet nobody knows how to make. ## Why it still makes sense to act The obvious conclusion would be to wait. It is probably the worst choice, for two reasons. First, the slow part of an AI project is not the technology. It is organising the data, understanding the process and writing down the rules nobody has written before. That work is needed anyway, takes time, and does not become obsolete when the model or supplier changes. Second, those who wait are not saving money; they are postponing learning. When tools change every quarter, being able to evaluate them is worth more than any single tool. ## A sensible position Act on focused projects with near-term returns, while retaining rights over what is built and the ability to change suppliers. If the cycle continues, you have accumulated an advantage. If it deflates, you have a system you can keep using, data you can take with you and no contract to renegotiate in a hurry. It is the same position in both scenarios. And it is the only one that does not require guessing which will happen. ## Sources and further reading **Sources:** Federal Reserve Board, Financial Stability Report, May 2026, near-term risks section and Survey of Salient Risks to Financial Stability · Federal Reserve Bank of Chicago, analysis of banking exposure to generative AI investment, 2026 · Sequoia Capital, “AI's $600B Question” and subsequent updates - [Federal Reserve · Financial Stability Report, May 2026 — Near-Term Risks](https://www.federalreserve.gov/publications/2026-may-financial-stability-report-near-term-risks.htm) — Updated 28 May 2026. Accessed: 2026-09-18. - [Sequoia Capital · AI’s $600B Question](https://sequoiacap.com/article/ais-600b-question) — Published 20 June 2024. Accessed: 2026-09-18.