Agent Cloud — The Infrastructure AI Agents Run On
The agent cloud is the computing infrastructure built for AI agents instead of for people: an isolated machine for every task, memory that outlives the task, and a meter on every call. Jamin Ball mapped it on 6 November 2026 and expects it to grow larger than the classic cloud.
What it is
An AI agent is software that carries out a task on its own: it writes code, runs it, calls other programs and checks the result. The cloud of the past fifteen years was built for something else, one application serving millions of people around the clock. Agents need the reverse: millions of small programs, each running for seconds or days on its own isolated machine, in bursts nobody can plan.
The investor Jamin Ball called the infrastructure for this the agent cloud in his Clouded Judgement letter of 6 November 2026. In his words it has “the same core primitives, but rebuilt for agents”.
The layers
Ball names six. Sandboxes are disposable computers, one per task, the successor of the virtual server. Persistent state is storage that outlives a task, so an agent can pause, copy itself and resume. Inference clouds run the AI models the agents call. A router or gateway spreads the work across different models, as the load balancer once spread traffic. Training loops turn freely available models into customised ones. Underneath sit ordinary processors, storage and network, needed at a much larger scale.
Around them he lists a second ring: tools that keep long tasks running, browsers for agents, machines for building and testing code, code storage, identity and payment for agents, monitoring and memory.
What changes
The classic cloud was tuned for traffic per machine. The agent cloud is tuned for three other things: how well tasks are kept apart, how much they remember and how fast a machine starts. The number of units is far higher, because every task gets its own machine.
Who builds it and who bills it
Somebody has to own the machines: the large cloud companies and the specialised renters of AI capacity. Above them sit companies that charge by use, per request, per stored unit, per monitored system or per checked identity, without owning data centres.
In our indices the two groups have parted since the middle of 2026. From 30 June to 11 November the Compute Operators in the Agentic Ecosystem Index lost 23.7% and the hardware in Rubin 100 29.1%, while Data & Memory gained 22.9% and Identity, Trust & Governance 38.3%. Ball makes no claim about who keeps the profit; this reading is ours, and the report sets out the case against it too.
Where it sits in our indices
The Agentic Ecosystem Index holds 34 listed companies paid each time an agent runs, in nine sections that match Ball’s layers closely: Compute Operators, Edge & Distribution, Runtime & API Gateways, Data & Memory, Operations & Observability, Identity, Trust & Governance, Agentic Security and Foundation Models. Most of the companies Ball names himself are private.
What to watch
The outlooks of the companies that charge by use, because the market has paid only those whose meter visibly runs faster. The debt of the capacity renters. And purchases such as Dropbox buying the storage company Tigris and OpenAI buying the code host Pierre Computer Company: application companies taking layers of the agent cloud in-house.
This entry describes how we read the term; none of it is a recommendation.