AI scaling laws: Why more compute has meant better models, and where the doubts start
Scaling laws are measured curves showing that AI models improve predictably as compute, data and size grow. They are the logic behind the data centre build-out, and the limits of that logic matter.

AI scaling laws are measured patterns showing that a language model’s error rate falls in a smooth, predictable way as three things grow: the size of the model, the amount of training data and the computing power used. They let labs forecast roughly how good a bigger model will be before they build it. That forecast is a large part of why companies spend billions on data centers.
A recipe that improves in predictable steps
Think of baking bread for more and more guests. If you know how much flour, yeast and oven time one loaf needs, you can estimate the ingredients for a hundred loaves without trial and error. Scaling laws do something similar for AI. Train several small models, plot how their quality changes as you add ingredients, and extend the line to a much larger model.
The three ingredients are the number of parameters (the adjustable numbers inside the model), the number of training tokens (small chunks of text the model reads) and the compute (the total arithmetic performed during training). The quality measure is usually the model’s loss, a number that falls as the model gets better at predicting the next token.
What the 2020 OpenAI paper found

The idea became widely known through a 2020 paper by Jared Kaplan, Sam McCandlish, Dario Amodei and colleagues at OpenAI, titled “Scaling Laws for Neural Language Models.” The authors reported that loss follows a power law in each of the three factors, with trends spanning more than seven orders of magnitude.
A power law means steady, diminishing returns. Each tenfold increase buys a similar fractional improvement, so progress continues but every step costs far more than the last. The paper also reported that details such as network width or depth had minimal effect within a wide range. Raw scale mattered more than clever shapes.
Its practical advice was that larger models learn more from each example. For a fixed compute budget, the authors concluded, the best choice was a very large model trained on a relatively modest dataset and stopped before it fully converged.
The 2022 correction: Bigger was not the whole story
In 2022, Jordan Hoffmann and colleagues at DeepMind published “Training Compute-Optimal Large Language Models.” They trained more than 400 models, from 70 million to over 16 billion parameters, on 5 to 500 billion tokens. Their conclusion was that the large models of the time were significantly undertrained. Model size and training tokens should grow in equal proportion: double one, double the other.
To test it, they built Chinchilla, a 70 billion parameter model trained on about four times more data than the 280 billion parameter Gopher, using the same compute budget. According to the paper, Chinchilla outperformed Gopher, GPT-3, Jurassic-1 and Megatron-Turing NLG on a broad range of tasks, and reached 67.5% average accuracy on the MMLU benchmark, more than 7 points above Gopher. This result is often called the Chinchilla scaling law.
Since then, many labs have trained smaller models on far more data than that rule suggests. A smaller model is cheaper to run for every user query, so extra training can pay off later.
How the curves connect to the building boom
If quality rises predictably with compute, then buying more chips, power and buildings looks like a reasonable bet. That reasoning sits behind the build-out of the data centre construction companies now in the news, which the Financial Times reported have lost $4.2tn in value. The FT’s public summary does not discuss scaling laws, so the link is ours: the investment case rests on the assumption that the curves keep paying off.
Reuters, as cited in our report, listed doubts about returns on AI spending among the pressures on infrastructure stocks. That is where scaling laws become a financial question. A power law promises progress, but not that each dollar will earn what the last one did.
Limits, caveats and newer scaling axes
Scaling laws describe averages measured on a specific quantity, the loss. They do not guarantee that a new skill, such as reliable reasoning, will appear at a given size. Predicting loss is easier than predicting what a model can do for users.
Other limits are practical. Training data is finite, and high quality text may run short. Larger runs need more power, cooling and money, and the financing behind that capacity depends on investor confidence. The fitted curves also come from particular model families and datasets, and they may shift when the recipe changes.
Labs have also begun to scale a different thing: the effort spent when a model answers. Our explainer on test-time compute covers that approach, which adds a second curve alongside training scale.
What to watch
- Whether new frontier models land on the old curves. Results that fall short would weaken the case for ever larger training runs.
- Data supply. Whether synthetic or newly licensed data can stand in for scarce text.
- Cost per query. Smaller, longer trained models shift spending from training to serving users.
- Spending signals. Capacity orders, cancellations or delays would show whether builders still believe the curves. The FT’s public text, as we noted, does not claim any projects have been cancelled.
Scaling laws are best read as an empirical observation, not a law of nature. They have held across a wide range so far, and the money now flowing into data centers is a bet that they continue to hold.
Sign Up for the Daily AI Pulse
One email a day. All the stories that matter.
By signing up, you agree to our Terms of Use and acknowledge the data practices in our Privacy Policy. You may unsubscribe at any time.
Daniel ReedAI Research, Safety & Ethics AnalystDaniel Reed currently works as an AI Research, Safety & Ethics Analyst at Aiholics, writing about how changes in artificial intelligence are affecting and will affect scholarship,…
You may also like

Test-time compute: How letting a model think longer improves its answers

Reinforcement learning with verifiable rewards: How AI learns from checkable answers

Flow matching: How robots and image models turn noise into smooth actions
