The 1,000x Compute Leap Behind Agent Swarms — Box CEO Aaron Levie
Key points
Key takeaways from the MTS interview with Box co-founder and CEO Aaron Levie (October 2026):
The 1,000x compute jump comes from agents, not chatbots. Levie argues the step-change in compute will arrive when agents are deployed in swarms and background operations that execute tasks fully decoupled from a human kicking them off — on the order of 100 to 1,000 times current AI consumption. He was responding to a post from the CEO of Daytona, who calculated that serving 100 million users of the Muse personal agent would need around 65,000 CPUs and 775PB of DRAM.
Personal agents plateau; enterprise workloads scale. Every user of a personal agent is effectively running a computer in the cloud, but consumer usage caps out because any one person has only so many simultaneous tasks. Levie expects the real 100x token and compute growth to come from enterprise workflows, where knowledge work is currently rate-limited by the energy a person can contribute.
Nonlinear agent deployment. He foresees deploying 10,000 agents to upgrade a major piece of software, fanning out agents across life sciences experiments, kicking off agents on every code change to review security risk, hedge funds applying near-limitless compute to trade research, and agent swarms working inside AI research itself.
Still astonishingly early. Citing Andrej Karpathy's funnel of AI use cases, Levie notes that from non-consumption and chatbot conversations down, only thousands to tens of thousands of people have actually deployed productive agent swarms. He expects usage to grow at least 100x from here, with compute and infrastructure the binding constraint.
Box's play is unstructured data at scale. Box deploys agents on contracts, invoices and research data to extract key points, automate workflows and surface insights, with demand across healthcare, financial services, life sciences, the public sector and federal government. Customer appetite is unsatiated; the binding constraints are the cost of compute and wiring up data and workflows for agent deployment.
Security is the next rate limiter. As swarms touch enterprise data, organisations need to know what agents accessed, what work they did and be able to halt execution mid-stream. Levie cites Satya Nadella framing agents as insider risk, and calls the guardrails, auditability and fine-grained controls around them the trillion-dollar question for the economy.
Box extends Shield to agents. Box is building the same fine-grained controls it already applies to humans — read versus read-write permissions, trusted and untrusted external agents versus internal ones, and anomaly alerts — for agents, and points to a wave of agentic security startups including Enclave, Depth First and Kevin Mandia's Method.
One to two orders of magnitude more infrastructure. Levie expects 10 to 100x more compute and infrastructure ahead, with agents running 24/7 across network-traffic aggregation, IoT and critical infrastructure, security camera footage, life-science research and continuous review of medical records. He sees the sector sitting downstream of a handful of builders, among them Intel's US fab push, AMD, Nvidia and TSMC.
Read more: MTS — Monitoring the Situation
