Microsoft published quarterly results at the end of July showing around $90 billion in revenue, driven by cloud growth of over 40%. The market reaction was spectacular, with one of the largest single-session gains in market capitalisation ever recorded. After months of questions over the sustainability of AI investment, this is the first serious counter-argument backed by hard numbers.
We have devoted several articles to doubts about the AI economy, including the financing deal reportedly being considered between Nvidia and OpenAI. Intellectual honesty demands that we also present the data pointing the other way. Here it is, along with its limitations.
What the numbers say
The central point is not the absolute revenue figure, but its source. Growth is coming overwhelmingly from the cloud business, which hosts the AI workloads of corporate customers. In other words, this is not classic licence sales, but consumption of infrastructure directly tied to AI adoption.
That is an important distinction. You can sell a subscription to a tool nobody uses. You do not bill for compute that is not consumed. Cloud growth at this pace means companies are actually running workloads, in increasing volume.
The market reaction reflects this reasoning: the question hanging over the sector was not whether AI worked technically, but whether anyone would end up paying for its use. These results provide a partial answer.
What it does not prove
Three caveats are worth setting out, and they are substantial.
Selling shovels is not finding gold. Microsoft gets paid on infrastructure, regardless of the eventual success of its customers' projects. A company spending on compute for an experiment that never pans out still generates revenue. The profitability of the infrastructure provider does not demonstrate the profitability of the uses it serves.
The order of magnitude remains unfavourable. AI revenue across all players remains below the amounts committed to infrastructure. A single data centre project can carry a construction cost higher than the annual revenue of the leading labs. The gap may be narrowing; it is not closed.
The circuits are partly closed. A share of AI spending circulates between players linked by cross-shareholdings, supply agreements and, if the reported information is confirmed, financial guarantees. In such a network, distinguishing real external demand from internal circulation becomes difficult.
The decisive question is not cloud growth, but customer profitability. A company that spends on AI and earns more because of it will renew. A company that spends without measurable return will eventually stop. Google's ATLAS study, which we analysed in a dedicated article, showed very broad but very shallow usage, with less than 10% real automation. That tension between massive spending and still-shallow usage is the real story of the coming quarters.
The balanced reading
These results do not close the debate on the bubble, but they shift it. It is becoming harder to argue that AI generates no real revenue: infrastructure consumption is measurable and growing.
It remains just as hard to argue that the question is settled. Rapid consumption growth is compatible with two opposing scenarios: durable adoption taking hold, or a phase of mass experimentation that will contract when budgets are assessed on their results. The same figures support both readings.
What can be said without risk of error is that the debate has moved from the terrain of belief to that of accounting. For two years, the discussion was about what AI could do. It is now about what it earns. That is a sign of market maturity, and probably a better conversation, even if a less exciting one.