AI Industry Approaching Data Scarcity Wall, With $100B+/Year Spend Projected by 2030
Summary
- • AI scaling is shifting from compute-limited to data-limited: only ~300 trillion tokens of useful public text exist, and pretraining demand is approaching that ceiling
- • Reinforcement learning extended the training data runway by tapping verifiable coding and math tasks, but that reservoir is also running low on sufficiently hard problems
- • Labs are projected to spend over $100 billion per year on data acquisition by 2030 as proprietary, synthetic, and institutional datasets become critical
- • The author calls for a 'Stargate for Data' — a civilizational-scale data collection effort to match the ambition of the trillion-dollar compute buildout
Details
~300T Token Public Ceiling
Approximately 300 trillion tokens of useful public human text exist on the internet — a finite ceiling that AI pretraining demand is rapidly approaching
Compute-to-Data-Limited Regime Shift
AI is transitioning from a compute-limited regime (GPUs were the bottleneck) to a data-limited regime where high-quality training data becomes the primary constraint on scaling
RL Data Runway Running Low
Reinforcement learning extended the usable data frontier by exploiting verifiable math and coding tasks, but the supply of sufficiently hard RL tasks is also being depleted
$100B+ Data Spend by 2030
Labs are on a trajectory toward more than $100 billion per year in data spend by 2030 as they compete for private, institutional, and synthetic data sources
Stargate for Data Proposal
Author proposes a 'Stargate for Data' — a coordinated, civilizational-scale initiative to collect and digitize data with the same ambition applied to the trillion-dollar compute buildout
AI Gaps Map to Data Gaps
AI systems consistently underperform in domains where web data coverage is limited — suggesting future capability ceilings will directly mirror data collection gaps
Dark Matter of the Internet
Vast tacit knowledge, undocumented processes, and institutional information has never been digitized — representing a significant but hard-to-access untapped data resource
Analysis of AI's coming data scarcity crisis: the internet's usable training data is nearly exhausted, RL's runway is depleting, and labs are heading toward $100B+/year data spend by 2030.
What This Means
The AI industry has operated under the assumption that data is effectively unlimited, but that era is ending. With ~300 trillion tokens of useful public text approaching exhaustion and verifiable RL training tasks also running dry, the bottleneck to continued AI progress is shifting decisively from compute to data. The author argues this demands a coordinated, trillion-dollar-scale effort to collect, digitize, and structure data from domains currently dark on the internet — and that without it, AI's ability to accelerate economic progress will be bottlenecked by domain-specific data gaps. The projected $100B+/year data spend by 2030 suggests this transition is already well underway.