Cloudflare $NET believes CPU demand could 20x with AI agents

$AMD $INTC $ARM
The company argues that AI agents could create a massive CPU/server infrastructure problem if every worker runs multiple agents using today’s cloud model
The math:
US:
⏩ 100M knowledge workers × 1 agent each ÷ ~10 agents per CPU = 10M CPUs
Global:
⏩ 1B knowledge workers × 10 agents each ÷ ~10 agents per CPU = 1B CPUs
Cloudflare compares this to current global server CPU production of only ~35M–45M per year, implying the current approach could require ~20x current annual CPU production
$NET is pitching Agent Cloud as a solution
This is how it would work:
1. Dynamic Workers instead of full containers
Agents can be spun up on demand and put back into cold storage when idle, avoiding one heavy always-on container per agent
2. Durable Objects for state
Agents need memory/state for tasks, files, sessions, progress, and tool outputs. Durable Objects give each agent/app a stateful object with local SQLite-backed storage
3. Workflows for long-running agents
Agents may run for minutes, hours, or days while waiting for approvals, retries, external events, or scheduled steps. Dynamic Workflows allow these processes to hibernate between steps instead of constantly consuming compute
4. Project Think / Agents SDK
Project Think adds agent-specific primitives like durable execution, crash recovery, checkpointing, sub-agents, persistent sessions, and sandboxed code execution
5. Sandboxes only when full computers are needed
For coding agents or tasks requiring bash, git, file systems, browsers, or arbitrary binaries, Cloudflare still offers Sandbox containers. But the idea is to use them only when necessary, not as the default runtime for every agent