What happens after the AI scaling era
For a decade, capability tracked computation, and the field was a spending contest. If that correlation breaks, progress becomes a design problem — which is the kind small teams can win.
The single axis
The scaling era had a clean structure: more compute, more data, more capability. The structure was so reliable it became the industry's operating system — capital allocation, research priorities, and careers all organized around the assumption that the frontier was a single axis and the only question was who could climb it fastest.
Single axes favor the biggest spenders. They convert money into progress at a predictable rate, and predictability is exactly what large organizations are built to exploit.
What breaks the axis
No law guarantees the axis continues. The candidates for breaking it are physical (energy and silicon economics), statistical (data quality, not quantity, becoming the binding constraint), and economic (the return on each additional unit of compute falling below its cost). Which one bites first does not matter much; the consequence is the same.
When capability stops being purchasable at a fixed rate, progress moves to the systems around the model: memory that persists, tools that can be trusted with delegation, verification loops that let a system check its own work, environments that give it something to act on. These are engineering problems, not capital problems. They reward teams that can think carefully and build fast — the historical specialty of small labs.
The hedged position
The honest position on scaling is that nobody knows, including the people spending the most. So the lab's posture is deliberately hedged: build systems that are valuable if scaling continues and more valuable if it stalls — systems whose advantage comes from composition and ownership rather than raw capability.
An asymmetry like that is the only free lunch in research. You cannot know the future, but you can arrange your work so that most futures pay you.