Reports that AI scaling has plateaued: why that is good news if you are buying
Several outlets reported this week that simply making models bigger has stopped producing big jumps. For a buyer, a slower treadmill is a feature.
Over the past week, several reports have said the same thing: the strategy of making models dramatically bigger has stopped producing dramatically better results. Ilya Sutskever, an OpenAI co-founder now running his own lab, told Reuters that results from scaling up pre-training have plateaued, and that « the 2010s were the age of scaling » while the field is now looking for the next idea.
Coverage has treated this as bad news. For anyone buying rather than building, it is mostly the opposite.
Why a plateau helps a buyer
Decisions stop expiring. The main reason AI projects stalled in 2023 and 2024 was the reasonable fear that anything built would be obsolete in four months. If the curve flattens, a tool chosen in November is still a sensible tool next autumn.
Competition moves to price and fit. When nobody can win on raw capability alone, vendors compete on cost, reliability, support and integration — the things an SME actually experiences.
The gap to open models narrows. A frontier that advances slowly is a frontier the rest of the field catches up to, which widens your options and weakens the single-supplier problem.
What the reports do not say
They do not say progress has stopped. They say one particular lever — more data and more compute in pre-training — is yielding less. Other levers are being pulled: reasoning at answer time, as in o1, is the obvious one, and it was the story of September and October.
They are also reports, not published results. Treat the specific claims with the caution any unsourced industry story deserves.
The practical takeaway
If you have been postponing a decision because « it will all change next year », that argument just got weaker. The sensible move in a flattening market is to pick a tool that solves a named problem today, on a contract short enough to change, and get on with it.
Waiting for the curve to stop moving has already cost some companies two years. It may never stop, but it appears to be slowing, and that is enough to decide.