Howard Marks, co-chairman of Oaktree Capital, says this week that AI's recent stumbles have dented the sector's sense of infallibility. The same day, Kioxia's stock drops, Foxconn's chief tries to reassure on 2027, and Bloomberg asks where to take shelter should the bubble deflate. The day's dominant story fits in one word: doubt.
Two figures, published on the same day, tell another story. ChatGPT crosses a billion weekly users, a little more than one human in eight, less than four years after its release. In China, for the first time, models produce more data in service than they consumed to train. AI has entered daily work, and no trading session will pull it back out.
We observe that two questions are being conflated under the same word. The first is about valuation: will the hundreds of billions committed to chips and data centres pay off? The market is right to doubt it.
The second is about use: is AI already serving organisations, every day? The answer is yes. A correction would sink Nvidia, not your teams' skill at writing, sorting, coding or analysing with these tools.
For a French organisation, the lesson is practical. Training does not lose value when a stock falls; it costs a fraction of an infrastructure budget and stays yours whatever becomes of the bubble. Companies that use a possible cooling to train their people will come out with an asset the market cannot correct.
Those that wait for the uncertainty to pass will pay for the same licences, without the skills to make them pay off.
The bubble, if it exists, will eventually find its price. AI at work already has its own: a billion people open ChatGPT every week. Train first, the market will decide later.
The Masteria editorial team.
