ReadingResearchRadarInvestment framework
Sign in / Sign up
← Back to reading
News3 min read·2026-10-12

Epoch AI Estimate: Chips Shipped Through 2027 Could Run Hundreds of Millions of Agents, But Demand Gap Looms

Epoch AI researcher Jason Li uses high-bandwidth memory (HBM) shipment volumes as a basis to estimate that AI chips shipped between 2025 and 2027 could theoretically support approximately 30 to 170 million concurrent frontier-model agents, equivalent to the weekly working hours of 140 to 720 million full-time employees. The report further shows that even using only 20% of potential capacity, the required API spending would reach $2.6 to $5.3 trillion annually, far exceeding projected developer revenues, highlighting a risk of over-supply in AI infrastructure investment.

ImportanceStructuralEvidenceE3 inspectableWrite-upStandard

2026-10-02

AI companies are spending hundreds of billions of dollars annually on chips and data centers, premised on the idea that this hardware will run AI agents to perform work that people do today. A critical but often overlooked question is: how many agents can this infrastructure actually support? And if capacity exceeds demand, what happens to returns?

In a study published on October 2, 2026, Epoch AI researcher Jason Li quantified this supply-demand relationship. Li previously worked at NVIDIA on LLM serving and inference optimization, and now focuses on AI inference topics at Epoch. He chose high-bandwidth memory (HBM) as the measurement basis because HBM is simultaneously an important component for inference and a main bottleneck for chip production, and its supply data is relatively transparent. Globally, HBM is produced primarily by three companies: Micron, Samsung, and SK hynix.

Li's estimation method proceeds in two steps: first, he converts HBM shipped since 2025 into equivalent GPU units by dividing by 288 GB (matching one GB300 GPU), then multiplies by the number of concurrent agent sessions each equivalent GPU can support. For closed frontier models, he analyzed actual session traces from the TraceLab dataset covering Codex and Claude Code workloads, deriving API-equivalent spending per agent-hour—roughly $16–18 for Codex and $24–50 for Claude Code—using $30 as a central reference point. Combined with GPU rental costs ($5 per GB300-hour) and an assumed API revenue-to-serving-cost ratio of 5–10×, he calculates concurrent sessions per equivalent GPU.

Sign in / Sign up
ReadingResearchRadarInvestment framework

The core finding: HBM shipped from 2025 through 2027, once fully deployed and allocated to agent workloads, could theoretically support approximately 30 to 170 million concurrent frontier-model agents. Because AI agents can work 168 hours per week rather than the human 40-hour standard, these agents could supply weekly working hours equivalent to 140 to 720 million full-time employees. For reference, the United States has a population of 342 million and an estimated 100 million knowledge workers.

However, there is a stark gap between this theoretical capacity and actual market absorption. Li assumes only 20% of potential capacity is used for revenue-generating inference (40% allocation, 50% utilization), yielding annual API-equivalent spending of $2.6 to $5.3 trillion. By contrast, if leading model developers maintain fivefold annual revenue growth, their combined annualized revenue would reach roughly $1 trillion by end-2027. The potential supply corresponds to a demand level 2.6 to 5.3 times higher than current revenue projections.

Li explicitly identifies the key uncertainty as demand-side. Short-term shortages may occur due to power or land constraints for data centers—Microsoft CEO Satya Nadella said in late 2025 that the company had chips sitting in inventory that could not be plugged in because suitable powered space was unavailable. As these bottlenecks ease, accumulated hardware will come online rapidly. Epoch's separate analysis estimates AI inference prices have been falling approximately 47% per quarter since 2023. If prices fall faster than demand grows, maintaining utilization requires agent activity volume to grow enough to offset efficiency gains.

This analysis is not a stock prediction or investment recommendation. It offers a structural perspective: the bet on AI infrastructure is not just about technical feasibility, but whether economic demand can keep pace with physical compute expansion. If AI applications fail to move beyond simple Q&A to broad economic value creation, or if users are unwilling to pay near-human-wage rates for agents, the current capital expenditure boom could face severe return-on-investment pressure.

Sources

  • 原文与论证epoch.ai↗
Continue researching →