AI chips shipped through 2027 could eventually support tens to hundreds of millions of concurrent agents running frontier models, according to a new Epoch AI analysis. Its bigger warning is about demand: the hardware build-out could outpace the market for AI services.
Epoch AI Watch analysis
What happened
Epoch AI estimates that memory shipped from 2025 to 2027 could support around 30 million to 170 million concurrent frontier-model agent sessions, if the hardware is fully deployed and allocated to those workloads. The estimate rises to about 1.9 billion sessions when applying DeepSeek V4 Pro serving benchmarks, a more efficient-model scenario.
The analysis translates those sessions into working hours, spending and assumptions about hardware supply, model serving costs and utilisation. Read Epoch AI’s analysis.
Key findings
- Frontier models: tens to hundreds of millions of agents
Epoch’s estimate for chips shipped through 2027 is 30 million to 170 million concurrent sessions, assuming full deployment and allocation. - A more efficient-model scenario reaches 1.9 billion
Applying DeepSeek V4 Pro serving benchmarks produces an estimate of about 1.9 billion concurrent agents on the same projected hardware supply. - Capacity could exceed demand
Epoch estimates that using 20% of its central capacity estimate could imply $2.6 trillion to $5.3 trillion in annual API-equivalent spending. It contrasts that with roughly $1 trillion in annualised model-developer revenue by the end of 2027, assuming fivefold annual growth.
Why it matters
These are estimates of potential capacity, not a forecast that billions of useful agents will be deployed or paid for. But they put a sharper figure on the gap between building AI infrastructure and finding enough work, customers and money to keep it busy. That is a rather expensive question to leave until after the data centres are built.
The report also shows how much the answer depends on model efficiency, serving costs, hardware assumptions and demand. Epoch says its estimates hold current model and workload requirements fixed, and notes that agent output quality varies. A theoretical capacity figure is not a headcount of digital employees ready to clock in.
Our read
This is a useful way to test the grand infrastructure story against a practical question: how much agent use could the hardware actually support, and who will pay for it? Treat the totals as scenarios, not a prediction. The most consequential finding may be that even modest use of the projected capacity implies an enormous market.
What to watch
- Whether real-world demand grows quickly enough to use the capacity being built.
- How model efficiency and serving costs change the estimated number of sessions per chip.
- Whether future estimates account for workload requirements and agent quality changing over time.
Discussion spark: If projected AI hardware capacity outruns likely demand, should companies keep building ahead of the market or slow investment until paying use cases catch up?
Sources and evidence
- How many AI agents could we run? | Epoch AI (2 October 2026, 00:00 UTC)
Independent WittyWires tracker for public updates about Epoch AI. Not affiliated with or endorsed by Epoch AI; this is not an official account.