Grayscale Sees AI Token Demand Climbing 24-Fold by 2030

Grayscale Head of Research Zach Pandl expects AI token consumption to climb 24-fold by 2030, with cheaper models and autonomous agents driving usage higher. Cheaper Models Widen AI's Reach Anthropic's newest Sonnet model costs up to 30% less per task, the company says, and competition from low-priced open-weight models is pressing on token prices. Tokens measure the text AI models process and generate, and developers pay for the volume they consume. As prices fall, businesses can justify AI applications that once cost too much. Grayscale's report, "Investing for the Compute Bottleneck," looks at how cheaper intelligence could support wider adoption. Agents Multiply Compute Demand OpenAI's recently launched "dots" are always-on agents designed to work toward a user's goals around the clock. Unlike chatbots that answer a single question, agents complete multi-step workflows: preparing for a client meeting, for instance, may mean reviewing notes, analyzing portfolio changes and drafting an agenda. Grayscale estimates such work consumes five to 50 times more tokens than a chatbot interaction, depending on complexity. Agents could drive substantial growth in inference, the computation required to run AI models. Goldman Sachs projects token consumption could rise 24-fold by 2030, largely on agent adoption. Infrastructure Remains the Bottleneck More efficient models do not guarantee lower total computing demand when usage grows faster. Supplying the necessary infrastructure takes time, and power availability, data centers and GPUs remain essential to running expanding workloads. Limited supply could cap how quickly capacity grows alongside AI adoption. Grayscale cites CoreWeave, which recently reported signing new contracts at higher compute prices — a sign of continued pricing pressure even as the cost per AI task declines. Pandl's analysis argues that cheaper models and more capable agents could lift usage enough to outweigh efficiency gains. In that scenario, owners of scarce physical computing infrastructure stand to benefit.