Scaling laws in 2026: what still scales and what doesn't

Scaling laws are the closest thing AI has to physics. Two empirical curves — Kaplan's 2020 power laws and DeepMind's 2022 Chinchilla correction — have dictated how labs spend hundreds of billions of dollars on compute. The idea was seductive: double the parameters, double the data, and watch loss fall along a predictable line.

In 2026 that story is half true. The curves still predict training loss with impressive accuracy. But training loss is no longer what labs optimize for, and the ratio Chinchilla made famous — roughly 20 training tokens per parameter — is one that almost nobody follows anymore. Here's a tour of which scaling laws still hold, which broke, and which new ones took their place.

The old curves still predict training loss#

Start with what hasn't changed. The foundational relationships from OpenAI's "Scaling Laws for Neural Language Models" (Kaplan et al., 2020) and DeepMind's "Training Compute-Optimal Large Language Models" (Hoffmann et al., 2022) continue to describe pretraining well: loss falls as a smooth power law in parameters, data, and compute, with training FLOPs approximated by the formula C ≈ 6 · N · D (six FLOPs per parameter per token). Labs still use these fits to predict a large model's final loss from small-scale runs before committing tens of millions of dollars to a training run — and the predictions still land within a few percentage points.

So the machinery of scaling-law measurement is intact. What broke is the prescription: the idea that a single compute-optimal ratio tells you how to build the best model.

Chinchilla's 20:1 rule broke on the economics#

Chinchilla's headline result was that compute-optimal training splits resources equally between parameters and data, at about 20 tokens per parameter. DeepMind proved it by training Chinchilla (70B parameters on 1.4T tokens), which outperformed the 4×-larger Gopher (280B) despite using the same compute budget.

The catch, formalized in work on inference-aware scaling (including the "Beyond Chinchilla-Optimal" analysis and MosaicML's "Chinchilla Trap" argument), is that Chinchilla optimized the wrong objective for most real deployments. Training is a one-time cost; inference — serving the model to millions of users — is the recurring cost that dominates total cost of ownership. If you expect billions of API calls over a model's lifetime, you should deliberately overtrain: train a smaller model on far more data than 20 tokens per parameter, so the model is cheap to serve.

The industry moved. According to data published by Epoch AI, the average token-per-parameter ratio in open-weight models climbed from 10 in 2022 to 300 in 2025, growing at 3.1× per year. The extremes are even more dramatic: Alibaba's Qwen3-0.6B (April 2025) trained 600M parameters on 36 trillion tokens — a 60,000:1 ratio — and Liquid AI's LFM2.5-350M (April 2026) pushed the figure to 80,000:1 through a blend of pretraining and large-scale reinforcement learning. Llama 3 8B sat at 1,875:1. The 20:1 line isn't wrong as physics; it's just not the optimum anyone pursues.

What still scales: the new frontiers#

With pretraining ratios settled, the frontier of scaling research moved in several directions:

1. Inference-optimal scaling#

Overtraining smaller models on more data than Chinchilla recommends is now standard practice — for production models, the inference-optimal ratio runs 100–2,000 tokens per parameter. The lesson: scaling laws aren't about making the best model; they're about making the best model you can afford to serve.

2. Test-time (inference-time) compute#

Reasoning models — OpenAI's o-series, DeepSeek-R1, Gemini 2.5 — spend extra compute after training, generating long chains of thought before answering. Research from 2026 ("Test-Time Scaling Makes Overtraining Compute-Optimal," arXiv) showed this feeds back into the training calculus: when models will reason at inference time, overtraining at training time becomes even more attractive. Scaling laws are becoming multi-dimensional — parameters × data × reasoning compute — not just bigger models.

3. Architecture-conditional scaling laws#

The original laws treated architecture as a rounding error. New work is fixing that. A paper accepted to ICLR 2026 ("Scaling Laws Meet Model Architecture," from UW-Madison and AWS researchers) fit conditional scaling laws that incorporate architectural choices — hidden size, the MLP-to-attention parameter ratio, and grouped-query attention — over 200+ models from 80M to 3B parameters. Their optimized architectures achieved up to 42% greater inference throughput than LLaMA-3.2 baselines at the same training budget. This is scaling law research with inference efficiency in the objective function, not bolted on afterward.

4. Post-training scaling#

RLHF/RLVR and synthetic data pipelines now contribute gains comparable to more pretraining, especially on reasoning benchmarks — where math and reasoning scores are far more sensitive to post-training and inference-time compute than to raw pretraining scale. Repeating data up to ~4 epochs remains ~85% as effective as unique data (Muennighoff et al., 2023), but beyond that, returns collapse.

What doesn't scale (or is hitting walls)#

• The data wall. Epoch AI estimates the effective stock of quality- and repetition-adjusted public human text at roughly 300 trillion tokens — and projects frontier developers could exhaust it between 2026 and 2032 if trends continue. After that, the field leans on private datasets, other modalities, and AI-generated synthetic data, which carries its own risks (diversity loss, model collapse on self-training loops).
• Diminishing returns per capability. Studies across 400+ models found that as datasets grow, improvements decelerate faster than classic laws predict — the marginal uniqueness of each new sample shrinks. Different capabilities also plateau at different sizes: knowledge tasks show diminishing returns past ~30B parameters, while only creative tasks keep benefiting strongly at large scales.
• Capability ≠ loss. The laws predict loss, and loss is increasingly decoupled from the capabilities people actually buy: reasoning, tool use, agentic behavior. A 2% loss improvement can mean nothing on benchmarks that matter, or everything.

The practical takeaway#

If you're building with AI rather than training it, the 2026 scaling story reduces to three actionable facts:

1. Small, overtrained models keep closing the gap. Falcon 180B (2023) was beaten by Llama 3 8B (2024) a year later. When choosing a model, prefer the smallest one trained far beyond Chinchilla ratios — it's cheaper per token and often just as good.
2. Budget inference-time compute, not just API calls. Reasoning models trade latency and cost for accuracy on hard problems; the lever of "let it think longer" is now a first-class scaling dimension.
3. Data strategy beats parameter strategy. With public human text potentially exhausting within this decade, labs are differentiating on data pipelines — curation, synthetic generation, private corpora — not on having more GPUs than the next lab. The same logic applies to fine-tuning your own models.

Scaling laws didn't die. They multiplied. The single curve became a portfolio of curves — training, inference, reasoning, architecture, data — and the labs winning in 2026 are the ones optimizing the whole portfolio instead of chasing the one line Chinchilla drew in 2022.