Epoch AI: HBM shipments through 2027 could support 30–170 million frontier agents
A new estimate from Epoch AI puts a number on the big question behind hundreds of billions of dollars in chip and data-center investment: how many concurrent AI agents can the hardware being delivered through 2027 actually run? The answer is between 30 and 170 million agents based on frontier models — up to 1.9 billion with more efficient serving. But what could limit deployment, according to the analysts themselves, is not chip supply but demand: even 20 percent utilization of that capacity would require $2.6–5.3 trillion in annual API spending, against projected developer revenue of roughly $1 trillion. At the same time, Iterate.ai's launch of its Lifeboat inference engine shows that the fight over how many agents each GPU can carry is already happening in products.
How the estimate is built
Epoch AI's analysis starts from high-bandwidth memory (HBM), the type of memory modern AI accelerators depend on. For agent workloads, the analysis argues, it is memory that sets the constraints: memory capacity limits how many agents can run concurrently, while memory bandwidth limits how fast the model can stream results (Epoch AI).
The analysts count all HBM shipped from 2025 onward, including HBM3E and newer generations, in units of 288 GB — equivalent to the memory on one Nvidia GB300 GPU. HBM shipped in 2025–26 alone could, once in operation, support 16–56 million concurrent agents. Including deliveries through 2027, the estimate rises to roughly 30–170 million concurrent agents based on frontier models [7259a40f].
It is worth stressing what these figures are — and are not. They are projections under stated assumptions, not measured deployments. The calculation assumes, among other things, that the projected hardware supply is actually built out and put into agent work. In reality, the hardware will be used for many different workloads, and some of it may never be installed. The precise details of the assumptions sit in the publication's methodology, and the results should be read as conditional estimates, not facts about current agent usage.
Billions of work hours — and an efficiency edge
To put the numbers in perspective, Epoch AI uses work hours: 30–170 million concurrent frontier agents corresponds to the weekly work time of 140–720 million full-time employees [7259a40f]. That is enormous capacity measured against human labor.
The figure could be far higher if serving becomes more efficient. Applying the serving performance of DeepSeek V4 Pro to the same projected hardware, the analysis arrives at roughly 1.9 billion concurrent agents — enough to deliver as many weekly work hours as 8 billion people working 40 hours each [7259a40f]. This assumes the performance figures from the DeepSeek model transfer to entirely different systems, and there is no independent verification of those measurements.
The bottleneck is demand, not silicon
This is where the analysis turns the chip-shortage debate on its head. If the market used just 20 percent of the central capacity estimate, Epoch AI calculates it would correspond to between $2.6 and $5.3 trillion in annual API spending. For comparison, the analysts estimate developer revenue at roughly $1,000 billion by the end of 2027 — but even that figure rests on an assumption of fivefold revenue growth per year until then (Epoch AI).
Epoch AI itself points to the possibility of a surplus: "Demand could fall behind this potential supply and create an overabundance of capacity. The key uncertainty is whether sustained, rapid growth in demand for AI services will justify the investment," the analysis writes — reproduced here in English translation.
The cost per agent-hour also varies widely by model. In Epoch AI's analysis of agent traces, Codex workloads averaged roughly $16–18 per hour of continuous agent activity, versus $24–50 for Claude Code workloads [7259a40f]. The figures show that an agent-hour is a commodity with a wide price margin — and that the total bill for agent deployment depends heavily on which models and work tasks you choose.
Iterate.ai: more agents per GPU as a business model
Analysts are not the only ones doing the math on agent density. On October 5, 2026, software company Iterate.ai launched Lifeboat, an inference engine for large language models with built-in confidential computing. The company claims the software fits two to six times as many concurrent agent sessions on each GPU (SiliconANGLE).
Technically, Lifeboat attacks the same problem Epoch AI points to: memory. According to SiliconANGLE, the optimizations address the pressure from the key-value cache — the model's working memory for each conversation — which grows as agents run over long sessions and crowds GPU memory. The company claims that managing this pressure makes room for more concurrent sessions on the same card [b8b8c8c4].
The company's own tests, as reported by SiliconANGLE, were run on a single Nvidia RTX PRO 6000 Blackwell GPU with the open model Qwen 30B-A3B. With Lifeboat's optimizations in place, the engine held 2,048 concurrent sessions with all requests completed, versus half that without. Throughput was 8,714 tokens per second, versus 4,965 without the optimizations [b8b8c8c4].
"Before any enterprise buys more GPUs for its agents, it should find out what the ones it already owns can do," said Jon Nordmark, co-founder and CEO of Iterate.ai, according to SiliconANGLE. His argument is that doubling concurrent sessions per card yields more capacity without new racks or more power.
The numbers carry caveats: they come from the company's own tests on one GPU model and one open model, and are not independently verified. They still show where the market is heading — per-GPU agent density has become an active product focus, not just a research question.
Three open questions
The biggest uncertainty in Epoch AI's numbers is implementation. The estimate assumes the hardware shipping in 2025–27 is actually built out and put into service for agent work. How much is realized — and when — will in practice determine whether the world ends up with capacity for 30 million, 170 million, or 1.9 billion agents.
The second question is the pace of efficiency. If serving innovations of the kind Iterate.ai claims to represent become widespread, the supply curve could shift as fast as demand grows — further reinforcing Epoch AI's point that the scarce thing is not compute, but willingness to pay for what it produces.
The third is the whole economic equation. The gap between $2.6–5.3 trillion in implied API spending at 20 percent utilization and roughly $1 trillion in developer revenue, built on an assumed fivefold annual increase, is enormous. Either prices must fall sharply or efficiency must rise — or a large share of the planned capacity will sit unused. Epoch AI does not hide the possibility: the world could end up with a surplus of agent capacity. For now, it remains an open question, not a verdict.