- AI got smart by training on basically the whole internet, and it is running low. Estimates put the exhaustion of high-quality human text between 2026 and 2032.
- The workaround, training on AI-generated synthetic data, has a trap called model collapse, where each generation gets blander unless real data stays in the mix.
- Data is becoming a scarce, paid resource, which is why licensing deals are exploding and why owning unique data is the new moat.
The Setup
Two forces drive how good AI gets: compute and data. Everyone talks about chips and gigawatts. Far fewer talk about the other limit, which may bite first. The models learned by reading the internet, and there is only one internet. Researchers now think the supply of high-quality human text good enough to train on runs out somewhere between 2026 and 2032.What the Data Wall Actually Is
Modern models are trained on enormous piles of text, measured in trillions of tokens, a token being roughly a word-piece. Epoch AI estimated the high-quality public internet holds around 300 trillion tokens, and frontier models already train on a meaningful slice of that. The problem is that good data does not grow as fast as the models' appetite for it. Push the scaling a few more years and the labs run into a wall: they have read everything worth reading. That is the data wall.Why You Cannot Just Make More
The obvious fix is synthetic data, having AI generate its own training material. It works up to a point, and it carries a famous risk called model collapse. Train a model on the outputs of earlier models, generation after generation, and it slowly loses the rare, weird, true edges of human writing and drifts toward bland, average text, a copy of a copy of a copy. The research has a fix: mix synthetic data with real data instead of replacing it. That keeps the model anchored to reality, and it means real human data stays essential.Why Data Became Expensive
This is why a data-licensing market appeared almost overnight. Public AI content deals went from basically zero in 2022 to around 90 by the end of 2025, and the prices keep climbing. Reddit, now the single most-cited source inside AI models, signed deals with Google and OpenAI that could reprice toward a combined 550 million dollars a year. News publishers are signing too. When the raw material runs short, whoever owns a unique, high-quality pile of it gets paid.What It Means For Investors
Follow the scarce input again. If compute was the bottleneck story of the last two years, data is the quieter one underneath it. Companies sitting on large, unique, human datasets, forums, publishers, niche archives, suddenly hold a real asset. Proprietary data is becoming a moat that money alone cannot quickly rebuild, which is why OpenAI has signed more licensing deals than almost anyone. For users of AI, it also hints that future gains may come more from better, cleaner, exclusive data than from raw size.So Is AI About to Stall
Probably not stall, and the easy gains from just feeding it more text are fading. The next edge comes from squeezing more out of the same data: smarter training, reasoning at inference time, multimodal data like video and audio, and exclusive sources others cannot touch. The era of free, infinite training data is closing. What replaces it is data as a managed, paid, strategic resource.FAQ
Didn't they already train on everything? How is there a wall?They trained on most of the easy, high-quality public text, yes. The wall is that this supply is roughly fixed while model demand keeps growing, and the genuinely good data, not spam or duplicates, is a smaller slice than it looks.
If synthetic data causes collapse, why use it at all?
Because used carefully it helps, especially for narrow skills like math and code where you can check the answer. The danger is replacing real data with synthetic across generations. Mixed in alongside human data, it is a tool, not a trap.
What should I watch as an investor?
Watch who owns unique data and who is paying for it. Licensing deals, content partnerships, and exclusive archives are turning into balance-sheet assets. The companies that control scarce, high-quality data may hold more leverage than the ones just buying chips.