AI scaling laws: why bigger wins, and whether it's ending

If you want to understand why the last six years of AI looked the way they did, why companies spent hundreds of billions of dollars building ever-larger models, and why the field is now arguing about whether that era is over, you need one idea: scaling laws. They are the empirical discovery that a language model's performance improves in a smooth, predictable way as you increase three things: the size of the model, the amount of data it trains on, and the compute spent training it. That predictability, more than any single architecture, is what turned AI from a research curiosity into a capital-intensive industry, because it meant that spending more reliably bought more capability.

This piece explains what scaling laws actually say, why their predictability was such a consequential finding, the important correction that changed how models are trained, the subtle gap between what they predict and what users experience, and the live 2026 debate about whether pure scaling is hitting walls and giving way to a different way of buying capability. It is the organizing principle behind the modern AI era, and understanding it is what lets you make sense of the industry's bets, its spending, and its current pivot.

The discovery: performance follows a power law

The foundational result, established in 2020, is that the loss of a language model, a measure of how well it predicts text, falls in a remarkably regular way as you scale it up. Plot the loss against model size, or against dataset size, or against compute, on the right axes, and you get a straight line: a power law. As each input grows, loss decreases by a predictable amount, and this relationship holds not over a narrow range but across many orders of magnitude, from tiny models to the largest ever built.

The surprising part is not that bigger is better, which anyone might guess, but that bigger is better predictably. Because the relationship is a smooth power law, you can measure how a series of small models perform, fit the curve, and forecast how a much larger model will perform before spending the money to build it. This is what made the enormous investments rational rather than reckless. A lab could estimate, with reasonable confidence, that a model ten times larger trained on ten times the data would reach a specific level of performance. Scaling stopped being a gamble and became something closer to engineering, where you could plan capability as a function of budget. Few findings in the history of the field have had a larger practical consequence.

Why it reshaped everything

Scaling laws did something uncomfortable to the research culture that produced them: they suggested that the largest gains came not from clever new architectures but from making the same basic thing bigger. Once the transformer existed, the dominant lever on performance turned out to be scale, not ingenuity. This is a version of what has been called the bitter lesson, the repeated historical observation that general methods riding on more computation tend to beat carefully hand-engineered ones.

The consequence was a decade-defining strategy: to get a better model, scale up. This is the logic behind the succession of ever-larger models, the arms race for compute, and the construction of enormous datacentres. It reframed progress as substantially a matter of resources, which is why AI became a domain where a handful of well-funded labs could pull ahead, and why compute access became a strategic asset. The whole shape of the industry followed from the shape of a curve.

The Chinchilla correction: data was undervalued

The first version of the scaling laws contained an error that mattered a great deal in practice. It implied that when you had more compute, you should spend most of it on making the model bigger, and relatively little on training it for longer on more data. Models built on this advice were large but trained on comparatively little text.

In 2022, the Chinchilla research corrected this. By running the experiments more carefully, it found that model size and training data should scale roughly together, in near-equal proportion, as compute grows. The rule of thumb that emerged was about twenty tokens of training data for every parameter in the model. Judged against this, many of the famous large models had been badly undertrained: they had far more parameters than their training data could properly exploit. The most cited example is that a 175-billion-parameter model trained on 300 billion tokens would have been beaten, at the same compute cost, by a much smaller model trained on trillions of tokens. The lesson, that it is not just scale but the right balance of each ingredient, became the standard for training, and later model families were explicitly designed around it.

It is worth knowing, as a lesson in how empirical science actually proceeds, that the disagreement between the original laws and the Chinchilla correction was partly an artifact of measurement. The earlier work used relatively small models and measured model size in a way that excluded certain parameters, which inflated its estimate; when later researchers corrected for this, the two frameworks moved closer together. The clean story of one law overturning another is messier underneath, which is a useful reminder that scaling laws are fitted empirical curves, sensitive to how you set up the measurement, not laws of nature.

The gap between loss and capability

There is a subtlety that separates people who understand scaling laws from people who quote them, and it concerns what exactly is being predicted. The power laws describe loss, the model's smooth, continuous skill at predicting the next token. But what users and businesses care about is capability: can the model do arithmetic, write correct code, pass an exam. And the mapping from smoothly falling loss to specific capabilities is not itself smooth or easy to predict.

Some abilities appear to arrive suddenly. A model shows no skill at a task across a range of sizes, and then, past some scale, the ability seems to switch on. These have been called emergent abilities, and they complicate the tidy picture of predictable scaling, because they suggest you cannot always foresee what a bigger model will be able to do, even if you can foresee its loss. There is a genuine debate here: some researchers argue these sudden jumps are partly a mirage created by how the ability is measured, since a harsh all-or-nothing metric can turn smooth underlying improvement into an apparent sudden leap. Whether emergence is real or an artifact of measurement remains contested. Either way, the practical point stands: scaling laws predict the loss curve well, and predict which specific downstream capabilities will appear far less well, which is why labs still have to build the big model to find out exactly what it can do.

The walls: where pure scaling runs into limits

By the mid-2020s, the strategy of simply scaling up began to meet real constraints, and this is the heart of the current debate. The most concrete is the data wall. Scaling laws demand more data in proportion to model size, but the supply of high-quality human-written text is finite. Estimates place the exhaustion of the usable stock of quality public text somewhere between the mid-2020s and the early 2030s, and frontier labs already report bumping against limits on how much unique, high-quality text they can gather. You cannot keep scaling data indefinitely when you are running out of data, and training repeatedly on the same text or on lower-quality material yields diminishing returns.

Alongside the data wall sits a broader worry about diminishing returns on pretraining itself. The power law never promised that scaling would keep delivering dramatic jumps; a power law flattens, so each further order of magnitude of compute buys a smaller absolute improvement in loss. Combined with the spiralling cost of each new scale-up and the data constraint, this raised the question that now dominates the field: has the era of buying capability mainly by making pretraining bigger reached the point of diminishing returns? The honest answer is that the easy, dramatic gains from naive scaling have become harder to get, which has pushed the field to find other axes to scale.

The shift: scaling compute at inference, not just training

The most important development of the current era is that the field found a new place to spend compute, and it changed the story. Instead of putting all the extra compute into making the model bigger during training, you can spend more compute at inference time, letting the model think longer about each problem. This is test-time compute, and it powers the reasoning models that generate long chains of thought before answering, or sample many attempts and select the best.

The striking finding is that this new axis has its own scaling law, and that spending compute at inference can, for some tasks, buy more capability than spending the same compute on a bigger model. A smaller model allowed to think longer can beat a larger model answering instantly. This reframes the whole enterprise: capability is no longer a function of training scale alone but of training scale and how much the model is allowed to deliberate at run time. The new paradigm is often summarised as train smarter, then reason smarter, combining efficient training, sparse architectures like mixture of experts, and quality data, with heavy use of test-time compute. It is why reasoning models became the frontier just as raw pretraining scaling was showing strain: the field did not stop scaling, it changed what it scaled.

Inference-aware scaling and the shape of things now

One more refinement shows how much the thinking has matured. The original compute-optimal rule minimised training cost for a target performance, ignoring what happens afterward. But a model that will be queried billions of times spends far more compute over its life on inference than on training. Later work extended the framework to account for this, and reached a practical conclusion: when a model will be used heavily, it can be worth overtraining a smaller model, pushing it on far more data than the compute-optimal rule suggests, because the smaller model is cheaper on every one of those billions of queries. This is why modern model families ship small members trained on enormous amounts of data: they are optimised for cheap inference over a long service life, not for the old training-only notion of optimality. The same instinct to find regularities has also been extended to the reinforcement-learning stage of post-training, which appears to follow its own scaling patterns, with the familiar diminishing efficiency at larger scale.

So is scaling over? The accurate answer is that scaling as a principle is very much alive, but the naive single-axis version, just make the pretrained model bigger, has matured into something more plural. The field now scales data quality, architectural efficiency, inference-time compute, and post-training, rather than only parameter count, precisely because the returns on raw size alone have become harder to extract. Scaling laws did not stop being true. They stopped being the whole story.

The short version

Scaling laws are the empirical finding that a model's loss improves in a smooth, predictable power-law relationship as you increase model size, training data, and compute, holding across many orders of magnitude. Their importance is the predictability: you can forecast a larger model's performance before building it, which made enormous investments in scale rational and drove the modern AI era. The Chinchilla correction showed that data and model size should scale together, about twenty tokens per parameter, revealing that many big models were undertrained. But scaling laws predict loss more reliably than they predict specific capabilities, they face a looming data wall as high-quality text runs short, and pure pretraining scaling shows diminishing returns. The field's response has been to scale new axes, especially test-time compute, letting models think longer at inference, which can beat scaling parameters, plus data quality, sparse architectures, and post-training.

The idea to hold onto is that scaling laws made capability a predictable function of size, data, and compute, which is why the industry bet everything on scale, and the current moment is not the end of scaling but its diversification, from a single axis of making pretrained models bigger to many axes including how long a model thinks at inference. The curve that built the industry is flattening on its original axis, and the field's answer has been to find new axes to climb.

Common questions

What are scaling laws in AI? Scaling laws are empirical relationships showing that a language model's loss, its skill at predicting text, improves in a smooth, predictable power-law fashion as you increase three things: the model's size in parameters, the amount of training data, and the compute spent training. Discovered in 2020, they hold across many orders of magnitude. Their significance is predictability: because the relationship is regular, you can measure small models, fit the curve, and forecast how a much larger model will perform before building it, which is what made large-scale investment in AI rational rather than a gamble.

What is the Chinchilla scaling law? The Chinchilla scaling law, from 2022, corrected the original scaling laws by showing that model size and training data should be scaled roughly in equal proportion as compute grows, rather than putting most extra compute into model size. Its rule of thumb is about twenty tokens of training data per model parameter. This revealed that many earlier large models were undertrained, having far more parameters than their limited training data could exploit, so a smaller model trained on more data would beat them at the same compute cost. Chinchilla-optimal training became the standard, and later model families were explicitly designed around it.

Why are bigger AI models better? Because of scaling laws: as a model grows in parameters and is trained on proportionally more data and compute, its ability to predict text improves in a predictable, power-law way, which translates into broadly greater capability. Larger models have more capacity to represent patterns and knowledge. The important caveats are that size must be balanced with enough training data to be effective, that the improvement per additional order of magnitude of compute shrinks as the curve flattens, and that raw pretraining size is now only one of several axes the field scales, alongside data quality, efficient architectures, and inference-time compute.

Are AI scaling laws hitting a wall? Pure pretraining scaling faces real constraints. The clearest is the data wall: high-quality human-written text is finite, and estimates place its exhaustion between the mid-2020s and early 2030s, with frontier labs already constrained on unique high-quality data. On top of that, power laws flatten, so each further scale-up buys a smaller absolute gain at rising cost. The dramatic returns from naive scaling have become harder to get. But scaling is not dead; the field has shifted to scaling other axes, especially test-time compute, data quality, and efficient architectures, so the principle continues even as the original single-axis approach matures.

What is test-time compute scaling? Test-time compute scaling means spending more compute when the model answers a question, rather than only when training it, by letting the model think longer, generate long chains of reasoning, or attempt a problem many times and select the best result. It has its own scaling law, and remarkably, for some tasks, spending compute this way can improve capability more than spending the same compute on a larger model. A smaller model allowed to deliberate can beat a bigger model answering instantly. This is the mechanism behind reasoning models and a major reason the field pivoted here as raw pretraining scaling showed strain.

Do scaling laws predict what a model can do? Only partly. Scaling laws predict loss, the smooth, continuous measure of next-token prediction skill, quite well. But the mapping from loss to specific capabilities, like doing arithmetic or writing correct code, is much less predictable. Some abilities appear to switch on suddenly past a certain scale, called emergent abilities, though there is debate over whether these jumps are real or an artifact of harsh all-or-nothing metrics turning smooth improvement into an apparent leap. Either way, labs generally still have to build a larger model to discover exactly which downstream capabilities it will have.

Is the era of scaling over? Not over, but changed. Scaling as a principle remains central, but the naive version of simply making pretrained models bigger has run into diminishing returns and the data wall, so it has diversified. The field now scales multiple axes: data quality, sparse and efficient architectures like mixture of experts, reinforcement-learning post-training, and above all test-time compute, letting models think longer at inference. Scaling laws did not stop being true; they stopped being the whole story. Capability is still bought with compute, but increasingly through how a model is trained and how long it reasons, not through parameter count alone.