Epoch AI estimates billions of AI agents could be supported by 2027

by@EpochAIResearch

Summary

A recent analysis has estimated that hardware shipments of high-bandwidth memory (HBM) from 2025 to 2027 could support tens to hundreds of millions of concurrent frontier-model AI agents. This capacity amounts to potentially supplying as many working hours as 140 to 720 million full-time employees, depending on the efficiency of the AI models used. Current spending for these agents varies, with workloads using the Codex model averaging approximately $16 to $18 per agent-hour, while Claude Code sessions range between $24 and $50. However, to fully utilize even 20% of this capacity would require a drastic increase in demand for AI services, projecting annual API-equivalent spending of between $2.6 trillion and $5.3 trillion, considerably exceeding anticipated revenues of about $1 trillion from leading model developers by the end of 2027.

Analysis

Codex: Codex refers to an agentic workload harness associated with OpenAI models, used for tasks involving model calls and tool use in continuous sessions. The news analyzes Codex workloads alongside others to estimate hourly costs and concurrency on AI hardware, providing a basis for projecting agent capacity from HBM shipments. Josh You: Josh You is a researcher acknowledged for providing feedback on the AI agent capacity analysis. The individual contributed to refining the estimates of hardware-supported agents and demand implications in the report. Anthropic: Anthropic is an AI company developing frontier models and agent harnesses such as Claude Code. The news highlights its role through cost benchmarks for Claude workloads and references Dario Amodei’s vision of scaled AI agents in data centers. JS Denain: JS Denain is a researcher acknowledged for providing feedback on the AI agent capacity analysis. The individual contributed to refining the estimates of hardware-supported agents and demand implications in the report. Claude Code: Claude Code is an agentic workload harness developed by Anthropic for running sessions with its Claude models, incorporating model calls, tool use, and subagents. It serves as a key example in the analysis for comparing agent-hour costs and serving requirements against closed frontier models. Dario Amodei: Dario Amodei is the CEO of Anthropic. The news cites his description of a future “country of geniuses in a datacenter” to frame the discussion on the potential scale of concurrent AI agents supported by upcoming hardware. Jaime Sevilla: Jaime Sevilla is a researcher acknowledged for providing feedback on the AI agent capacity analysis. The individual contributed to refining the estimates of hardware-supported agents and demand implications in the report. Venkat Somala: Venkat Somala is a researcher acknowledged for providing feedback on the AI agent capacity analysis. The individual contributed to refining the estimates of hardware-supported agents and demand implications in the report. Demand Outlook: Utilizing projected agent capacity at scale would require sustained rapid growth in demand for AI services across the economy. Hardware Capacity: High-bandwidth memory shipments from 2025 through 2027 provide the basis for estimating how many frontier-model agents future data centers could run concurrently. Workload Comparison: Different agent harnesses and models produce varying hourly costs that affect how many concurrent agents the same hardware pool can support.

Categories

aitechai_agentsmachine_learning
View Original Tweet