AI token consumption could increase 24-fold by 2030 as cheaper models and autonomous agents expand usage, according to Zach Pandl, head of research at asset manager Grayscale. The outlook, laid out in the firm’s report “Investing for the Compute Bottleneck,” also flags that the physical infrastructure needed to run these systems remains constrained. Grayscale argues that rising demand for computing power could outpace efficiency gains, intensifying pressure on available data centers, electricity supplies, and GPUs.

Cheaper AI Models Could Expand Usage

Declining costs are making artificial intelligence more accessible for applications and automated workflows. Anthropic’s latest Sonnet model now costs up to 30% less per task, according to the company, while competition from lower-priced open-weight models—whose parameters are publicly available for anyone to download and run—is putting continued downward pressure on token prices.

Tokens measure the text that AI models process and generate, and developers pay for their consumption. As prices fall, businesses may find new uses for AI that previously cost too much to run. The report examines how cheaper intelligence could support broader adoption.

AI Agents Could Multiply Computing Demand

OpenAI’s recently introduced “dots” are always-on agents designed to work toward users’ goals around the clock. Unlike chatbots, which respond to individual questions, agents can complete multiple steps within a single workflow. Preparing for a client meeting, for instance, could require reviewing notes, analyzing portfolio changes, and drafting an agenda.

According to Grayscale, tasks of this kind can consume five to 50 times more tokens than a standard chatbot interaction, depending on complexity. Agents could consequently drive substantial growth in inference, the computation required to run AI models. Goldman Sachs projects that token consumption could rise 24-fold by 2030, with the bank attributing the increase largely to agent adoption.

Physical Infrastructure Remains a Compute Bottleneck

Although AI models are becoming more efficient, total computing demand can still rise when usage grows faster than those gains can offset, a dynamic that recalls the Jevons paradox, the 19th-century economic observation that greater efficiency can increase rather than reduce total consumption of a resource. Supplying the necessary infrastructure takes time. Power availability, data centers, and GPUs remain essential to supporting expanding workloads, and their limited supply could constrain how quickly computing capacity grows alongside AI adoption.

Grayscale also pointed to CoreWeave, which recently reported signing new contracts at higher compute prices, as an example of continued pricing pressure on infrastructure even as the cost of individual AI tasks declines.

Pandl’s analysis argues that cheaper models and more capable agents could raise usage enough to outweigh efficiency gains. Under this scenario, demand could benefit owners of scarce physical computing infrastructure. How the balance between falling token costs and rising infrastructure pricing evolves may become visible in further model-price cuts from providers such as Anthropic, in contract disclosures from operators like CoreWeave, and in revisions to demand estimates such as Goldman Sachs’ 24-fold projection.