Epoch AI Estimates Storage Can Run Tens of Millions of Cutting-edge AI Agents by 2027
Epoch AI estimates that by 2027, storage chips shipped could simultaneously run approximately 33 to 171 million cutting-edge AI agents.
If these agents operate around the clock, their working hours would equate to the weekly working hours of 140 to 720 million full-time employees. This estimate is based on the high bandwidth memory (HBM) shipment from 2025 to 2027, converting it into equivalent GPU numbers, and calculating the concurrent agent scale based on service benchmarks of different models.
The report indicates that when using more efficient open-source models, the same hardware can support billions of agents. For example, the DeepSeek V4 Pro shows benchmark calculations of about 1.9 billion concurrent agents, whose round-the-clock working hours could match the workload of approximately 8 billion people working 40 hours a week.
Epoch AI uses HBM as a unified measurement standard across manufacturers, as storage has become the main bottleneck for AI chips. The estimate assumes that all hardware is deployed and dedicated to agent workloads; actual deployment delays or other uses would reduce the number of operable agents.
The report also analyzes the demand gap: even with a conservative estimate of 20% utilization, annual spending would need to reach $2.6 to $5.3 trillion, while the current revenue growth path of leading labs suggests a total annualized revenue of about $1 trillion by the end of 2027, which is far below what is needed to fill capacity.
In the expansion of AI infrastructure, storage shipments determine the upper limit of concurrently running agents, with funding flowing towards HBM and inference chip suppliers. Beneficiaries are those manufacturers with storage capacity and efficient model providers, while those under pressure are labs relying on high-priced cutting-edge models and whose demand growth is insufficient to absorb capacity.
Source: Public Information
ABAB AI Insight
Epoch AI has been tracking AI training computing power and chip supply for a long time, previously using HBM shipments as a bottleneck indicator for inference capacity. This time, the same method is extended to estimate the scale of concurrent agents, continuing its path of inferring application limits from hardware supply.
In terms of capital flow, the estimate relies on converting HBM shipment volume into equivalent GPUs, then inferring the number of agents based on service costs of different models. Funding is concentrated on expanding storage and optimizing efficient inference, rather than simply increasing the number of GPUs.
This is similar to cases where some labs reduce inference costs through model distillation and quantization in 2024-2025, allowing more sessions to run on the same hardware. We are currently transitioning from competition in training computing power to competition in inference concurrency scale.
Essentially, this represents a transfer of pricing power under technological substitution: storage supply determines the upper limit of concurrent agents, and efficient models can significantly increase the number of agents on the same hardware. The mechanism is that inference costs decrease with model efficiency, leading to a rapid increase in the working hours supported by unit storage, thereby changing the relative pricing of human labor hours and computing power hours.
ABAB News · Cognitive Laws
- Storage determines concurrency limits, model efficiency determines working hour multiples
- Round-the-clock operation amplifies agent working hours by 4.2 times
- When capacity expansion outpaces demand, the bottleneck shifts from hardware to willingness to pay.