The public web is running out of training data. Your company's records aren't.
The internet looks infinite, but for AI training it isn't: researchers at Epoch AI estimate the stock of quality public human text could be used up between 2026 and 2032.
What the research says
Epoch AI estimated the effective stock of public, human-written text at roughly 300 trillion tokens. At the pace models were scaling, they projected it would be fully used somewhere between 2026 and 2032.
When public text runs short (Epoch AI projection)Epoch AI, 'Will we run out of data?' (2024)
Why that matters to a normal company
Once public text runs short, developers pay for new, private, high-quality sources.
Business records show how real work gets done: decisions, fixes and handovers.
That's exactly what AI agents for the workplace need to learn from.
Key takeaways
Public training data is finite and largely used.
Private business records are the next frontier.
Value depends on quality, history and privacy handling.
