When the Man Selling the Shovels Mentions a Bubble

Satya Nadella recently said something that made a lot of very expensive decks wobble. He suggested that AI might be drifting into bubble territory. Not a crash tomorrow statement. Not a denial of progress. Just a raised eyebrow from someone who can see the electricity bill.

This matters because Nadella is not a man who gets paid to be nervous. He runs Microsoft, which currently has more money, more compute, and more incentive to talk up AI than almost anyone alive. If even he is clearing his throat, it is worth listening.

Here is the uncommon sense bit. Most conversations about AI are still framed as capability theatre. What it can do. What it might do. What the demo looked like. That is adolescent thinking. The grown-up question is whether the output is worth the input. Not philosophically. Commercially.

AI is not software in the old sense. It is an industrial process. It eats power, silicon, water, and people who know how to keep the whole thing upright. The cost curve does not gently slope down. It steps. And every step requires another data centre the size of Milton Keynes.

Nadella’s point, read properly, is not that AI is fake. It is that value does not automatically follow effort. We have seen this film before. Railways. Fibre. Crypto. Each time, the infrastructure was real, useful, and overbuilt. The wreckage was not the technology. It was the expectations stapled to it.

For leaders, the error is copying the loudest use cases. Chatbots everywhere. Agents bolted onto processes nobody likes. Internal tools that look clever and save nothing. This is how bubbles form. Not from fraud, but from repetition.

The better posture is dull and therefore rare. Ask where AI removes a cost that hurts. Ask where it removes a delay that blocks revenue. Ask where it replaces something you currently pay humans to tolerate rather than enjoy. If the answer is vague, stop.

AI works best when it is invisible and slightly boring. Scheduling. Matching. Forecasting. Reducing rework. Nobody applauds these things on LinkedIn. They do, however, show up in margins.

The lesson is not to slow down. It is to narrow the aperture. Treat AI like heavy machinery, not a magic trick. Measure it like a factory, not a keynote. And remember that when the people selling the shovels start warning about mud, it is usually because they have already stepped in it.