Artificial intelligence (AI) adoption continues to accelerate, with projections indicating that overall token consumption could expand roughly 24-fold by 2030. This growth trajectory, highlighted in recent analysis from Grayscale Research, reflects not only broader usage but also structural shifts in how AI systems operate.
Falling model costs and the rise of more autonomous tools are expected to amplify demand for computing resources well beyond simple query-response interactions.
A conventional chatbot typically handles a single question or request in one pass.
In contrast, an agent pursues a multi-step objective by sequencing tasks, gathering information, evaluating options, and refining outputs.
These longer workflows can require five to fifty times as many tokens as a basic chatbot exchange, depending on complexity.
Preparatory work for something like a client meeting—reviewing prior notes, assessing portfolio shifts, and composing an agenda—illustrates how token usage multiplies quickly.
Always-on systems that operate continuously toward user goals can extend this pattern further, generating sustained compute loads rather than isolated sessions.
OpenAI’s recently introduced “dots,” described as always-on agents capable of working around the clock, along with new lower-cost models from various providers, are seen as additional drivers.
Lower prices per token, partly from competition by open-weight alternatives, make more automated workflows economically practical.
When intelligence becomes cheaper, organizations and individuals tend to apply it to a wider range of tasks, expanding total consumption even as efficiency improves at the model level.
Goldman Sachs projections cited in the discussion attribute much of the expected 24x rise largely to agent-driven activity, with inference—the ongoing computation needed to run models for users—already growing rapidly.
Yet the physical systems that deliver this computation face real constraints.
Power generation and transmission, data center construction, and graphics processing unit (GPU) supply all require extended timelines for planning, permitting, building, and deployment.
Digital demand can scale almost immediately once software is available, while infrastructure lags.
As a result, cheaper AI and more capable agents can coexist with tight compute markets.
Some specialized providers have recently signed new contracts at higher prices for compute capacity, consistent with scarcity amid rising usage.
Grayscale Research concludes that expanding AI activity is positioned to favor owners of scarce compute resources.
Efficiency gains at the software layer may not fully offset the volume increase from broader adoption and more intensive agent workflows.
Owners of powered data centers, grid-connected facilities, and related infrastructure that can convert electricity into reliable computing capacity stand to benefit as demand outpaces the slower pace of physical build-out. This dynamic underscores a shift in which compute itself functions as a constrained input supporting the next phase of artificial intelligence deployment.