Training data supply

Enough good data to keep improving models.

capabilityBuilt today Built but scarce · 2 of 5 Open on the map →

Facts

Lane: Data
Wave: 1 of 12 (build order, not a date)
Needs: 0 boxes all the way back through hard links
Unlocks: 101 boxes downstream through hard links
Leverage: None: already built
Cluster: Science and medicine

Evidence as of October 2026

Epoch estimates the stock of public human-written text runs out sometime between 2026 and 2032. The fallbacks are synthetic data, private data and training on the same data more times.

01AI could run out of human-written training text (PBS, on Epoch AI)

Detail boxes 2

Expert human data market capability · Built today · helps, not required

Paying domain experts to write and grade training data.

Mercor passed $2bn of gross annualised revenue in June 2026, paying more than $2m a day to over 30,000 contractors a week.

01Dealroom: Mercor doubles to a $2B run rate

RL environment market capability · Built today · helps, not required

Companies building simulated workplaces and tasks for reinforcement learning.

Epoch AI finds environment contracts often run to six or seven figures a quarter; a high-fidelity Slack replica costs about $300,000.

01Epoch AI: the state of RL environments

Around it

Needs · 2

Data rights and licensingcapability · Emerging · helps, not required
Synthetic datacapability · Emerging · helps, not required

Unlocks · 1

Pretraining at scalecapability · Built today

Loops and off-ramps

Synthetic content and the commonsharmful loop
Domains make the datareinforcing loop
Training data ruled outoff-ramp · slowed