Loading page…
The estimate came during a conference conversation about cloud infrastructure and letting one AI agent check another—not a verified account of who paid.
Listen to this story
At Cloudflare Connect 2026, Peter Steinberger described an agent setup built around cloud execution and a second agent that can approve or block the first agent’s actions. He says the two-agent negotiation can change how a task is carried out and makes him more comfortable granting agents autonomy. Steinberger also estimated that he used $10 million worth of tokens in one month at OpenAI; that figure is his estimate, not a verified bill. The discussion points to infrastructure and permission design as practical constraints on more autonomous agents.
Steinberger argued that agents should run in the cloud rather than depend on a user’s computer.
In his approach, a separate agent can withhold permission, prompting the working agent to reconsider or simplify an action.
The $10 million token-use figure is Steinberger’s estimate and has not been independently verified.
OpenClaw creator Peter Steinberger says he used an estimated $10 million worth of AI tokens in a single month at OpenAI. Speaking with Cloudflare’s Dane Knecht at Cloudflare Connect 2026 in San Francisco, he also described a future in which agents run in the cloud and check one another’s actions.
The figure is Steinberger’s own estimate, not an independently verified bill. His work provides relevant context: TechRadar says he has been working with OpenAI on next-generation personal AI agents since February 2026.
a virtually unlimited token budget, which not all of us in the room have
Dane Knecht, describing Steinberger’s access during the conversation
Steinberger argued that the infrastructure supporting AI has not caught up with what he wants agents to do. His complaint was concrete: why should the work still depend on his computer rather than happen in the cloud? He framed better cloud infrastructure as part of making agents more useful, rather than simply asking for a more capable assistant.
His longer-term vision goes beyond answering a request. He expects people to have intelligence that runs throughout the day, thinks about their needs and may anticipate something helpful. In that vision, an answer could arrive before the person asks the question. That is Steinberger’s expectation for how agents might develop, not a capability announced at the conference.
Steinberger also described how he becomes comfortable giving an agent more freedom. Rather than relying on one agent to both perform a task and decide whether its actions are acceptable, he favors a second agent that can grant or withhold permission. The agent doing the work must negotiate with the other system.
He said those exchanges sometimes lead the working agent to reconsider an action or find a simpler approach that stays within its guardrails. The second agent is therefore not just checking a finished answer in his description; it can influence whether and how the work proceeds. Steinberger said combining the two systems substantially increases his trust in letting the agent act freely.
Loading discussion...
Join the conversation
Explain where that reassurance would help—or fall short.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.