Every few months, someone in AI declares that scaling is "hitting a wall." Every few months, a lab ships a model that quietly makes the claim look premature. To understand why this argument keeps recurring — and why it matters for anyone building on top of AI models rather than inside a research lab — you need to understand scaling laws: the empirical curves that describe how model performance changes as you add more compute, more data, and more parameters.
Scaling laws are not a law of nature in the way gravity is. They are observed regularities, discovered by training hundreds of models at different sizes and plotting the results. But they have proven durable enough, across enough orders of magnitude, that they now function as the closest thing the field has to a predictive science. If you want to understand why GPT, Claude, Gemini, and Llama models keep getting bigger, why training runs cost what they cost, and why "just add more data" replaced years of architecture tinkering as the dominant research strategy, scaling laws are the place to start.
What Scaling Laws Actually Say
At the core of scaling law research is a simple observation: a language model's loss — roughly, how surprised it is by the next word in a sequence — decreases in a smooth, predictable way as you increase three things:
- N — the number of parameters in the model
- D — the amount of training data (tokens)
- C — the total compute used to train it, which is roughly proportional to N × D
Plot loss against any one of these on a log-log scale, holding the others from becoming a bottleneck, and you get something close to a straight line. That straight line is a power law: loss falls off as a function of N or D raised to some negative exponent. Power laws are unusual in machine learning because they hold across many orders of magnitude — the same relationship that predicts the jump from a 10-million-parameter model to a 100-million-parameter model also predicts, with reasonable accuracy, the jump from 10 billion to 100 billion.
This is the part that makes scaling laws useful rather than merely descriptive. If a relationship is smooth and log-linear, you can fit it on small, cheap models and extrapolate to the large, expensive model you haven't trained yet. A lab can run a grid of models at manageable sizes, fit the curve, and use it to decide how big the next model should be and how much data it will need — before spending tens of millions of dollars finding out the hard way.
The Original Finding: Bigger Is Predictably Better
The idea that model performance scales predictably with size dates back to work at OpenAI in 2020, which showed that transformer language model loss follows power laws in N, D, and C independently, with remarkably little dependence on other architectural choices — depth versus width, attention head count, and similar details mattered far less than the raw resource totals. That was a significant claim: it suggested that architecture search, which had consumed enormous research effort, was less important than simply allocating more compute and data correctly.
The practical upshot of that early work was a strategy: given a fixed compute budget, make the model as large as possible and don't worry too much about running out of data, because in the regimes studied at the time, parameter count mattered more than token count. This shaped a generation of very large, comparatively undertrained models.
The Correction: Compute-Optimal Training
A few years later, further scaling research revisited this and found that the original guidance had gotten the balance wrong. Many large models of that era were significantly undertrained relative to their size — they had far more parameters than the training data could justify. The revised relationship, often called the "compute-optimal" or Chinchilla-style scaling law, showed that for a fixed compute budget, model size and training tokens should scale at roughly the same rate. Doubling your compute budget means you should roughly double both the parameters and the data, not just make the model bigger.
This single correction changed how frontier models get built. A smaller model trained on proportionally more data can match or beat a larger, undertrained model at the same compute cost — and it's cheaper to run at inference time afterward, which matters enormously once a model is serving millions of queries a day. Inference cost, not just training cost, is now a first-class variable in these decisions, which is part of why "smaller but more thoroughly trained" models have become common in production fleets.
Why This Matters Right Now
Scaling laws matter today for a reason that has nothing to do with any single announcement: they are the load-bearing assumption behind how the entire industry allocates capital. Multi-year compute commitments, chip purchases, data center buildouts, and data licensing deals are all sized against extrapolated scaling curves. When a lab commits billions of dollars to a training cluster, it is implicitly betting that the power-law relationship between compute and capability will continue to hold at the next order of magnitude.
That bet has three visible sources of uncertainty right now, all directly tied to the scaling framework itself:
- Data is finite. The compute-optimal relationship assumes you can keep feeding a growing model proportionally more high-quality tokens. Public text data is not infinite, and several labs have already turned to synthetic data, multi-epoch training, and multimodal data (images, video, audio) to keep the data side of the curve growing. Whether these substitutes scale the same way natural text does is an open empirical question, not a settled one.
- The exponents are small. Scaling laws are power laws with fairly modest exponents. That means each additional unit of capability costs disproportionately more compute — the curve keeps improving, but the returns diminish, and the dollar cost of the next meaningful jump in benchmark performance keeps rising.
- Pretraining scaling and reasoning scaling are now separate curves. The recent wave of "reasoning" models trained with reinforcement learning and extended inference-time computation (thinking longer per query rather than just being bigger) represents a second scaling axis — test-time compute — layered on top of the original pretraining scaling laws. Labs are now investing in both simultaneously, and the interaction between the two curves is one of the more active research questions in the field.
None of this means scaling has stopped working. It means the industry has moved from a single, simple scaling story (bigger model, more data, better results) to a more layered one, where pretraining scale, data quality, and inference-time compute are three separate levers, each with its own curve and its own diminishing returns.
How the Curves Actually Move Investment
It's worth being concrete about what a scaling law actually predicts and doesn't. It predicts loss — a statistical measure of next-token prediction quality — not task performance on any particular benchmark, and not qualitative capability jumps like "can now write working code" or "can now pass the bar exam." The relationship between loss and downstream capability is itself an area of active study, and it's where a lot of the popular confusion about scaling comes from.
| Concept | What it predicts | What it does NOT predict |
|---|---|---|
| Pretraining scaling laws | Cross-entropy loss on held-out text as a function of N, D, C | Which specific skills or benchmarks improve, or when |
| Compute-optimal ratio | The N:D split that minimizes loss for a given compute budget | Real-world usefulness or safety of the resulting model |
| Emergent capabilities | Not directly — these show up as instances where task performance moves discontinuously, unlike the smooth loss curve | Their own future occurrence — that's a debated, disputed phenomenon |
| Test-time / inference scaling | Task accuracy as a function of compute spent per query at inference | How this trades off against pretraining scale long-term |
That gap — smooth loss curve versus lumpy, sometimes-discontinuous capability jumps on specific tasks — is one of the more contested areas in the field. Some researchers argue that so-called "emergent abilities" are partly an artifact of how researchers choose to measure benchmarks (a metric with a hard pass/fail threshold looks discontinuous even when the underlying probability is improving smoothly), while others argue there are genuine phase transitions in capability. Either way, it means you cannot read a scaling curve and know in advance exactly which new abilities the next model will unlock — only that its general competence, in aggregate, should be higher.
Practical Implications for Builders and Businesses
You do not need to train a frontier model to care about scaling laws. They shape decisions for anyone building products on top of models that already exist.
Model selection is a scaling-law question in disguise
When a team chooses between a small, fast, cheap model and a larger, slower, more expensive one for a given task, it is implicitly making a scaling trade-off. The compute-optimal insight — that a smaller, well-trained model can match a larger undertrained one — is why so many vendors now ship families of models at different sizes trained on similar data mixtures, rather than one enormous model for everything. Fine-tuning a smaller model on domain-specific data, or using retrieval to supply context, is often a way of getting scaling-law-like gains without paying pretraining-scale costs.
Pricing follows the curve
Inference cost scales roughly with parameter count and the amount of compute spent per query. As reasoning models spend variable amounts of test-time compute depending on question difficulty, cost per query has become less predictable than it was for a single fixed forward pass. Budgeting for AI features now requires thinking about compute elasticity, not just a flat per-token price.
Capability roadmaps are compute roadmaps
If your product plan assumes "the underlying model will get meaningfully better next year," that assumption rests on someone else's scaling bet playing out. It is reasonable to expect continued improvement given the historical trend, but the pace of improvement is not guaranteed to be linear or even monotonic in every dimension — some capabilities (raw factual recall, for instance) scale differently than others (multi-step reasoning, tool use, long-context coherence).
A short checklist for teams making build decisions against this backdrop:
- Don't assume the next model generation will be uniformly better at your specific task — test against your actual use case, since scaling gains are measured in aggregate loss, not your benchmark.
- Prefer architectures (retrieval, tool use, fine-tuning) that let you improve results without waiting on the next pretraining cycle.
- Track total cost of inference, not just sticker price per token, especially for reasoning-heavy workloads with variable test-time compute.
- Treat vendor claims about "emergent" new capabilities with mild skepticism until you've verified them against your own tasks.
Limitations and Open Questions
Scaling laws are an empirical fit, not a derived physical law, and several open questions sit underneath the confident-sounding curves.
The laws are measured on loss, not on what people actually want from a model. Loss reduction correlates with better performance on many tasks, but the correlation is not perfect, and it says nothing about alignment, factual reliability, or safety — a model can have excellent scaling-predicted loss and still confidently state falsehoods, because loss measures prediction of training-distribution text, not truth.
Data quality is entangled with data quantity in ways the simple N-D-C framework doesn't fully capture. Two datasets of equal token count can produce very different models depending on deduplication, filtering, and diversity. As labs exhaust easily available high-quality text, the marginal token added to training sets is, on average, of lower quality than the ones before it — which means real-world scaling curves may bend earlier than a clean power law would predict.
Nobody has a rigorous theory for why power laws hold in the first place. There are proposed explanations rooted in the statistics of natural language and in properties of the underlying data manifold, but there is no first-principles derivation comparable to, say, thermodynamics. This is why the "physics" framing for scaling laws is useful as an analogy but should be taken with caution — it's closer to Kepler's empirical planetary laws than to Newton's underlying mechanics. The field is still looking for its Newton.
Test-time compute scaling is new enough that its own limits aren't established. Early results suggest that letting a model "think longer" at inference improves accuracy on certain reasoning tasks, following its own rough power-law-like curve, but how far that scales, and whether it will run into diminishing returns as steep as pretraining has, is unresolved.
What to Watch Next
A few signals will tell you which way the scaling story is heading over the next few years:
- Whether frontier labs keep disclosing scaling details. Early scaling papers were unusually transparent about compute, data, and parameter counts. Competitive pressure has made labs increasingly cagey about these specifics, which makes it harder for outside researchers to verify whether the curves are still holding as cleanly as before.
- Synthetic and multimodal data results. If models trained substantially on synthetic or multimodal data continue to follow the same scaling relationships as models trained on natural text, the "running out of data" concern becomes much less pressing. If they don't, expect renewed emphasis on data efficiency and architectural changes rather than raw scale.
- The pretraining-versus-test-time-compute trade-off. Watch whether labs continue investing heavily in ever-larger pretraining runs, shift more investment toward inference-time reasoning compute, or try to find an optimal blend — this allocation decision is effectively a bet on which curve has more room left to run.
- Independent replication. Because so much scaling research now comes from the same handful of labs building the models being studied, independent academic replication — where it happens — is worth watching closely as a check on whether the curves are being reported accurately or optimistically.
FAQ
What is an AI scaling law?
An AI scaling law is an empirically observed power-law relationship between a model's performance (typically measured as loss on held-out data) and the resources used to train it — parameter count, training data volume, and total compute. It lets researchers predict how a larger model will perform before training it.
What is the Chinchilla scaling law?
The Chinchilla scaling law refers to research showing that many large language models were undertrained relative to their size, and that for a fixed compute budget, model parameters and training tokens should scale at roughly equal rates rather than favoring parameter count alone. It reset how labs allocate compute between model size and data volume.
Do scaling laws mean AI progress will continue forever?
Not necessarily. Scaling laws describe smooth, predictable improvement in loss as resources increase, but the exponents involved are small, meaning each additional gain costs disproportionately more compute. Data availability, data quality, and energy and hardware constraints could all cause real-world progress to bend away from the idealized curve well before any theoretical limit is reached.
Why do bigger AI models cost so much more to train?
Training compute scales roughly with the product of parameter count and training tokens, both of which increase together under compute-optimal scaling. Because performance gains follow a power law with a modest exponent, each meaningful jump in capability requires a disproportionately larger compute investment than the jump before it.
What is test-time compute scaling?
Test-time (or inference-time) compute scaling refers to improving a model's accuracy on a given query by letting it use more computation at the moment of answering — for example, generating and evaluating longer chains of reasoning — rather than by making the underlying model larger. It is a separate lever from pretraining scale and has become a major focus for reasoning-oriented models.
Are emergent abilities in AI models real or a measurement artifact?
Both explanations have supporting evidence and the debate is unresolved. Some researchers argue that apparent sudden jumps in capability are genuine phase transitions in what a model can do; others argue they largely reflect how discontinuous benchmark metrics interact with an underlying, smoothly improving probability of correctness. In practice, this means capability improvements should be verified on your own tasks rather than assumed from a scaling curve alone.
How do scaling laws affect which AI model I should use for my product?
Scaling laws explain why smaller, well-trained models can match larger, undertrained ones at lower inference cost, which is why model providers now offer families of differently sized models rather than a single large one. For most applications, the practical choice comes down to testing model size against your actual task and cost constraints rather than assuming bigger always means better for your specific use case.
Teams weighing these model-selection and cost trade-offs in a real product roadmap can get hands-on help from Woyce Technologies.