The Scaling Laws Plateau

Since 2020, the AI industry has operated on a seductive premise: just make it bigger. More parameters, more data, more compute, better results. The scaling laws published by Kaplan et al. at OpenAI seemed to confirm this with mathematical precision — model performance improves as a smooth, predictable function of scale (Kaplan et al., 2020). Two years later, DeepMind's Chinchilla paper refined the ratios, arguing that training data should scale proportionally with parameters (Hoffmann et al., 2022). The message to the industry was clear: spend more, get more.

This was not just a research finding. It was a business thesis. And it justified the largest capital deployment in the history of computing. Understanding where this thesis holds and where it bends is essential for anyone trying to make sense of AI's trajectory — and the investment decisions that shape it.

The Scaling Hypothesis

The core claim is elegant. Plot model loss against compute on a log-log scale, and you get a straight line. Double your compute budget, and performance improves by a predictable amount. No architectural breakthroughs needed. No clever tricks. Just money.

This predictability was the point. It turned AI research from an uncertain scientific endeavour into something closer to engineering — or, more precisely, into a capital expenditure problem. If you know the curve, you can forecast returns. If you can forecast returns, you can raise funding. If you can raise funding, you can build bigger.

Between 2020 and 2024, this logic drove tens of billions of dollars into GPU clusters, data centre construction, and training runs. Nvidia's market capitalisation surged past $3 trillion. Microsoft invested $13 billion in OpenAI. New data centres consumed enough electricity to power small countries.

The scaling hypothesis was not a neutral observation about neural networks. It was the financial engine of the AI boom — the logic behind Nvidia's $3 Trillion Question and the broader AI valuation race.

Where It Breaks Down

The trouble with straight lines on log-log plots is that they eventually bend. By late 2024, evidence was accumulating that the era of reliable scaling returns was ending — or at least entering a new, less comfortable phase.

Diminishing returns

Each doubling of compute produces smaller absolute improvements. The leap from GPT-3 to GPT-4 was dramatic and unmistakable — the model crossed qualitative thresholds in reasoning, coding, and general knowledge that stunned even sceptics. The reported leap from GPT-4 to GPT-5 was considerably less dramatic. The Information reported in late 2024 that OpenAI's internal "Orion" project, an early effort toward GPT-5, showed disappointing improvements relative to the compute invested (The Information, 2024). Similar murmurs emerged from Google DeepMind and Anthropic.

This is not surprising in mathematical terms. A power law on a log-log scale is still a power law. You are running ever faster to stay in place. But it surprised an industry that had internalised "bigger equals better" as a near-certainty.

The data wall

Large language models learned from the internet. The problem is that the internet, while vast, is finite — and the high-quality fraction of it is much smaller. Books, academic papers, well-edited journalism, and carefully written technical documentation represent a limited corpus. Wikipedia, often cited as the gold standard of training data, contains roughly four billion words in English. That sounds like a lot until you realise GPT-4 was reportedly trained on trillions of tokens.

The industry's response was synthetic data — using AI models to generate training data for other AI models. This works, up to a point. But Shumailov et al. (2024) demonstrated that models trained recursively on their own outputs suffer from "model collapse," a progressive degradation where the tails of the distribution — the unusual, creative, and minority outputs — get shaved away with each generation. The model converges on an increasingly bland and homogeneous mean.

You cannot bootstrap your way past a data wall by talking to yourself. (For more on this dynamic, see Synthetic Data and the Ouroboros Problem.)

Energy and cost constraints

The training run for GPT-4 reportedly cost over $100 million. Estimates for frontier models in 2025 range from $500 million to over $1 billion. These figures include not just compute rental but the electricity to power it, the cooling to prevent the chips from melting, and the engineering talent to keep it all running.

At these price points, the number of organisations that can even attempt a frontier training run shrinks to a handful — the Frontier Lab Oligopoly in practice. And the returns on each marginal dollar of compute are declining, which means the economic case for ever-larger runs weakens even for those who can afford them. The Economics of Training a Frontier Model unpacks these costs in detail.

The DeepSeek counter-example

In January 2025, DeepSeek, a Chinese AI lab, released DeepSeek-R1 — a reasoning model that matched or exceeded frontier performance on many benchmarks while reportedly using a fraction of the compute budget of its Western competitors. DeepSeek achieved this through architectural innovation, aggressive mixture-of-experts designs, and efficient training techniques rather than brute-force scaling.

This was the most direct challenge to the scaling hypothesis since it was articulated. If comparable performance is achievable with dramatically less compute, then compute is not the binding constraint. Architecture, data quality, and training methodology matter more than the sheer number of floating-point operations. DeepSeek's Technical Report provides a window into how they achieved this — and DeepSeek and the Price of Open AI explores the market implications.

The market noticed. Nvidia lost nearly $600 billion in market capitalisation in a single day after DeepSeek's release — the largest single-day loss in stock market history at the time.

The Industry Pivots

The major labs did not, of course, simply concede that scaling had plateaued. Instead, they shifted the argument. The new claim: even if pre-training scaling is hitting limits, inference-time scaling — spending more compute when the model is actually answering questions — still works.

This is the thesis behind reasoning models like OpenAI's o1, o3, and their successors, as well as Anthropic's extended thinking capabilities and Google's Gemini reasoning modes. Rather than making the base model bigger, you let the model "think longer" at inference time, generating chains of reasoning that improve answer quality.

Simultaneously, the industry pivoted toward post-training improvements — reinforcement learning from human feedback (RLHF), constitutional AI methods, fine-tuning on curated datasets — and toward agentic systems that chain multiple model calls together to accomplish complex tasks.

These are genuine technical advances. But notice what happened: the goalposts moved. The original scaling hypothesis promised that you could simply spend your way to better intelligence. The revised story is more complicated, more uncertain, and less amenable to the simple capital-expenditure logic that powered the boom. How Reasoning Models Actually Work explains the technical shift in more detail.

Why the Misconception Matters for Redistribution

The "bigger is better" narrative was never just about technology. It was about power.

If the primary driver of AI capability is raw compute, then only organisations with billions of dollars can compete. The scaling hypothesis created a natural oligopoly — a small number of frontier labs backed by the world's largest technology companies, with everyone else relegated to fine-tuning the models those labs produce. As Vipra and Korinek (2024) argue in their analysis for the Brookings Institution, this concentration of AI capability maps directly to concentration of economic and political power.

This suited the incumbents perfectly. The narrative justified their capital advantages and made competition seem structurally impossible. Why fund a startup when only Microsoft, Google, and a few others can afford the training runs?

If scaling plateaus, this logic weakens. If architecture and data quality matter more than raw compute, then a well-funded lab with clever researchers — like DeepSeek — can challenge the incumbents. The moat of "more GPUs than anyone else" becomes shallower.

But do not mistake a shifting landscape for a level playing field. The industry is already pivoting toward new forms of concentration. Proprietary data — from user interactions, enterprise deployments, and synthetic pipelines — is becoming the next scarce resource. Specialised hardware (custom ASICs, inference chips) creates new dependencies. And agentic systems, which require massive infrastructure to deploy reliably, may reconcentrate power even as the base models commoditise.

The scaling plateau does not automatically redistribute power. It creates an opportunity for redistribution — one that will be captured by incumbents unless deliberate choices are made.

The Better Frame

AI improvement is not a smooth curve upward. It never was. The scaling laws described a particular regime — a period when compute was the binding constraint and architecture was held roughly constant. That regime produced real and impressive gains. But it was a phase, not a permanent law of nature.

The better frame is path-dependent technological development. Multiple approaches exist. Scaling is one. Architectural innovation is another. Better data curation, more sophisticated training objectives, inference-time computation, retrieval-augmented generation, neuro-symbolic methods — the design space is vast, and the optimal path through it is not predetermined.

The belief in inevitable, compute-driven scaling served the interests of those with the most capital. It made a particular business model — raise billions, buy GPUs, train the biggest model — seem like the only rational strategy. It discouraged investment in alternative approaches and alternative organisations.

Now that the scaling curve is bending, the question is whether the industry will genuinely diversify or simply find new ways to concentrate power under a different technical rationale.

Who benefits from the scaling narrative? Frontier labs with massive compute budgets. Nvidia and its GPU monopoly. Cloud providers selling compute by the hour. The investors who bet early on the scaling thesis.

What does the plateau mean? Potentially, a more competitive and more diverse AI ecosystem — if policymakers, investors, and researchers adjust their assumptions. Projects like EleutherAI and the Rise of Small Language Models movement suggest what a more distributed AI landscape might look like. The window is open. Whether anyone walks through it is not a technical question. It is a political one.

Related

• Economics of Training a Frontier Model — The cost structure behind the scaling bet
• Compute Is the New Oil — Why compute concentration matters for power
• Frontier Lab Oligopoly — The market structure scaling created
• DeepSeek and the Price of Open AI — The efficiency challenge to brute-force scaling
• How Reasoning Models Actually Work — The technical pivot from pre-training to inference-time compute
• Synthetic Data and the Ouroboros Problem — Why the data wall cannot be bypassed with AI-generated data
• AI Market Concentration — Research on how scaling dynamics shape market structure
• Rise of Small Language Models — The alternative path: capable models that don't require billions

Sources

• Kaplan, Jared, et al. "Scaling Laws for Neural Language Models." arXiv, 2020. arXiv
• Hoffmann, Jordan, et al. "Training Compute-Optimal Large Language Models." arXiv, 2022. arXiv
• Shumailov, Ilia, et al. "AI Models Collapse When Trained on Recursively Generated Data." Nature 631, 2024. Nature
• Vipra, Jai, and Anton Korinek. "Market Concentration Implications of Foundation Models." Brookings Institution, 2024. Brookings
• "OpenAI's Next Frontier Model Shows Slower Improvement." The Information, November 2024. The Information