Around 2020, researchers noticed something unusual in computing: the performance of a language model followed a predictable relationship with three quantities — its size, the volume of data, and the compute power devoted to its training. Predictable meant you could estimate in advance what a model ten times larger would deliver. This predictability justified multi-billion-dollar investments.
These relationships explain the race of the past five years, and their running out of steam explains what we are seeing today.
What these laws say
They describe an empirical regularity: when you scale up model size, data volume and compute together, the model's error decreases in a steady fashion.
Two important caveats. First, these are observations, not theorems: nobody has proved why this works, it has simply been observed. Second, the gains are diminishing: each improvement costs more than the last. Going from mediocre to decent is far cheaper than going from excellent to slightly better.
The point that changed the game lies elsewhere: emergent capabilities. Some abilities do not appear gradually but all at once, beyond a certain scale. A model cannot do something, then it can. That is what made the race irresistible: nobody wanted to miss the next threshold.
The three walls
Data. We have largely exhausted the quality text available. That is why labs buy printed books and why expert annotation is worth twenty billion. The resource is no longer abundant.
Cost. Diminishing returns eventually meet a budget. Scaling a training run that already costs hundreds of millions tenfold to gain a few points becomes hard to justify, especially in front of shareholders.
Hardware. We have documented the memory shortage, electrical constraints and opposition to data centers. The constraint is no longer merely financial, it is physical.
Everything we have covered for months makes sense in light of this slowdown. Mixture of experts allows more parameters without more compute. Reasoning at inference time shifts the effort from training to inference. Distillation transfers capabilities into smaller models. And vertical models bet on data nobody else has. None of these paths is about going bigger.
What this implies
The slowdown in scaling laws has observable consequences, and they explain the current state of the market.
Convergence. If everyone hits the same limits, the gaps narrow. That is what our comparison of American and Chinese models shows: three points separate the top from fifth place.
The shift to efficiency. When you can no longer win on power, you win on cost. Hence the price cuts and specialised chips.
The return of engineering. What differentiates is no longer model size but what you build around it: tools, memory, orchestration, clean data.
What to take away
A nuance is needed to close, because you often read that these laws are "dead". They are not. They still work, it is simply that the cost of the next step exceeds what players are willing to pay.
That is an important difference. A physical wall forbids continuing. An economic wall waits for someone to find a cheaper way. If compute costs fall enough, the race could resume.
In the meantime, this pause tends to benefit the user. A period when you can no longer win by force is a period when you win by ingenuity, and ingenuity produces models that fit on a graphics card rather than in a data center.