For most of the past decade, AI scaling has worked because the open web kept up. Each new generation of language model needed roughly an order of magnitude more text than the last, and the internet was big enough to make the math work. That assumption is now ending. Independent research suggests AI labs will catch up to the available stock of high-quality public text data within the next two to four years. After that, “train on more data” stops being a strategy.
What comes next is less a single solution than a scramble across several fronts.
How Big Is the Remaining Pool
The reference point most researchers cite is the 2024 Epoch AI study, peer-reviewed at ICML. Its estimate places the effective stock of quality-adjusted human-generated public text at roughly 300 trillion tokens, with a 90% confidence interval from 100T to 1,000T. The largest known training datasets sit at about 15 trillion tokens — meaning the absolute ceiling is roughly 20 to 60 times current usage. At a 2.5x annual growth rate, that headroom evaporates fast.
Epoch’s projection is full utilization between 2026 and 2032, with a median around 2028. OpenAI researchers reportedly noted during GPT-4.5’s development that data shortage was a tighter constraint than compute, unimaginable two years earlier. Anthropic CEO Dario Amodei places the probability of AI scaling stagnating due to insufficient data at around 10%.
Why the Useful Pool Is Shrinking Faster Than the Raw Pool
The headline number understates the problem because the high-quality slice is being walled off in real time. A Data Provenance Initiative study led by MIT researchers tracked content access between April 2023 and April 2024 and found that 5% of all training sources and 25% of the highest-quality sources had been restricted. OpenAI’s crawlers were blocked from nearly 26% of high-quality sources, Google’s from 10%, and Meta’s from 4%. Reddit sued Anthropic for allegedly scraping over a hundred thousand times after agreeing to stop.
When publishers realized the data was worth real money, the economic logic shifted in parallel.
The Licensing Era
By late 2025, content licensing deals had become a serious line item across the major AI labs. A partial map of who has signed what:
| AI lab | Counterparty | Reported value |
| OpenAI | News Corp | Up to $250M over 5 years |
| OpenAI | ~$70M/year | |
| ~$60M/year | ||
| OpenAI | Axel Springer | ~$13M/year over 3 years |
| OpenAI | Vox Media, Condé Nast, Dotdash Meredith, FT, The Atlantic, Le Monde, Guardian, AP | Multi-year, various |
| Amazon | New York Times, Condé Nast, Hearst | Multi-year, undisclosed |
| Meta | CNN, Fox News, USA Today, People Inc. (March 2026) | Multi-year |
The total publisher-AI licensing market reached an estimated $3.4 billion in 2025. Anthropic’s $1.5 billion settlement in an author’s copyright case the same year underlined the alternative: unlicensed training is now an existential financial risk, not a footnote. In September 2025, Reddit, Yahoo, Medium, People Inc. and others backed Really Simple Licensing — a framework modelled on music industry royalties, with pricing models like pay-per-crawl and pay-per-inference. No major AI lab has signed on.
The Synthetic Workaround and Its Real Limits
If you can’t scrape it and can’t always afford to license it, you generate it. Synthetic data — text produced by existing models to train new ones — has become the most discussed alternative. The market for synthetic data was valued at roughly $0.5–0.67 billion in 2025 and is projected to top $2.3 billion by 2030.
The catch is what researchers call “model autophagy disorder,” or MAD — the phenomenon where models trained on their own outputs gradually lose diversity, accuracy, and connection to real human language. A 2024 Nature paper showed this empirically across multiple generations. The fix, where one exists, is to keep grounding synthetic data in fresh human content. This is where private datasets matter. Regulated industries — finance, healthcare, online betting platforms like casino nv — generate massive proprietary logs of human behavior under tightly controlled conditions that AI labs would value but cannot scrape. Licensing arrangements with these sectors face thicker regulatory walls than news-publisher deals, but the data quality is often higher. Whether that data ever enters the training pipeline is a legal architecture question more than a technical one.
Strategies Replacing “More Data”
If raw scale stops paying off, the field has to win efficiency from other variables. Four directions are visible in current research:
- Multimodal training, converting image, video, and audio data into token-equivalents — Epoch estimates this could add 450T to 23 quadrillion tokens of usable scale.
- Algorithmic efficiency, with techniques lowering the compute needed for a given performance level by roughly 0.4 orders of magnitude per year.
- Specialized smaller models trained for narrow domains, often outperforming larger generalists on the target task.
- Test-time compute approaches, where models reason longer at inference — OpenAI’s o-series and DeepSeek’s R1 family are early evidence this can partially substitute for training-data scale.
None of these is a clean replacement. Together they extend the runway. Whether they sustain the trajectory of capability gains that defined 2020–2024 is a separate question.
What Actually Changes for the User
The end of effortless scaling doesn’t mean AI stops improving. It means the curve gets noisier. Expect more visible plateaus between releases, more domain-specific models replacing single general-purpose ones, and a larger share of model differentiation coming from training methods and proprietary datasets rather than raw size. The dramatic pre-2024 progress was partly about chips and partly about a free, abundant resource nobody priced. The chips are still scaling. The resource is not.
