Rapid expansion of AI compute capacity could enable tens to hundreds of millions of continuously running AI agents using today’s largest models, or billions of agents with more efficient models, within the next two years. The main bottleneck is availability of high-bandwidth memory (HBM); using projected HBM shipments for 2025–2027, researchers estimated how many GB300-equivalent GPUs could be produced and how many agents those GPUs could run concurrently. While this scale implies potential annual revenues in the trillions of dollars, whether market demand will absorb such a supply of agents remains the central uncertainty.
What the study examined
The analysis asked how much continuous agent labor the compute capacity coming online over the next two years could support. An "agent" in this context is a continuously running session of models similar to Claude Code, OpenAI Codex, or Meta’s Muse.
Methodology
- The study treated high-bandwidth memory (HBM) availability as the key constraint and used projected HBM shipments from 2025 through 2027.
- From those projections it derived how many GB300-equivalent GPUs could be manufactured.
- The GPU count was multiplied by estimates of how many concurrent agents each GB300-equivalent GPU could host, under different efficiency assumptions.
- Calculations assume agents run nonstop (including nights and weekends), allowing weekly agent-hours to be compared with full-time human workers.
Key quantitative findings
- Under the central assumptions, compute coming online over the next two years could support tens to hundreds of millions of concurrent agents when using today’s most capable models. In weekly-hours terms, the top-tier agents running nonstop would supply the equivalent of roughly 140 million to 700 million full-time employees.
- With more efficient models, the same hardware could support billions of agents. Applying DeepSeek V4 Pro serving benchmarks to the projected hardware supply yields about 1.9 billion concurrent agents under the central hardware assumption. Running nonstop, those agents would provide roughly the weekly working hours of 8 billion full-time workers.
Context for those numbers
- For scale: the United States population is about 342 million and it has an estimated 100 million knowledge workers. The comparisons above count working hours only; agents can also act faster than humans, while task quality may differ.
Revenue implications
- Leading model developers’ revenues have been growing very quickly. If recent fivefold annual growth continued, annualized revenues could reach roughly $1 trillion by the end of 2027.
- Under a conservative assumption that only 20% of total compute is used for revenue-generating inference at current API prices, the study’s main estimate implies between $2.6 trillion and $5.3 trillion in annual revenues.
Main uncertainty: demand vs. supply
Despite the large potential scale, the report emphasizes there is no guarantee these revenues will materialize. The crucial question is whether demand for AI-driven services grows rapidly enough to absorb the expanding supply of continuously running agents, or whether a capacity glut will depress utilization and revenue.
Conclusion
Projected hardware build-outs, limited primarily by HBM supply, could support an unprecedented digital workforce measured in hundreds of millions to billions of concurrent agents. That capacity could translate into very large revenues if demand follows, but the market’s ability to absorb such a volume of agent labor is the decisive unknown.
