Peter Steinberger, creator of OpenClaw and member of OpenAI's technical staff, estimates he spent around $10 million on tokens running AI workloads in the last month alone. According to the event, the bill is being covered by investors.
The figure came out Wednesday at Cloudflare Connect 2026 in San Francisco, during a conversation with Cloudflare CTO Dane Knecht. Knecht recalled that Steinberger has acknowledged having a "virtually unlimited token budget", something not everyone in the room has. Facing the audience's surprise, Steinberger added: "What people don't always understand is not all of that is useful." The remark did more to fuel the murmuring than to reassure.
Image: spaces — https://huggingface.co/spaces/Echo-AI-official/openclaw-ai/tree/main
Later, Steinberger pointed to infrastructure as the core problem: "Why is this still using my computer? Shouldn't this all be in the cloud?" Cost remains a central concern for anyone without venture funding, since cloud-hosted agents can be expensive.
Knecht linked the discussion to open source contributions: many maintainers have stopped accepting pull requests out of frustration with so-called "AI slop", low-quality generated code. Steinberger argued the way pull requests are handled must change, because people do not explain how they arrived at their proposed changes. "If I see your thoughts, how you got from A to B, my trust in the output is way higher than if you just throw the finished work across the fence," he said.
On how many agents people will run, Steinberger does not expect hundreds per person: "I think everyone will have some sort of intelligence that runs and thinks and helps them throughout the day, or even predicts what could help you in the future."
The final issue was trust and what data to give an agent. Steinberger admits part of the answer depends on how much you trust the underlying model, since much of its behavior is not governed by the harness, the code wrapping it. Even with risks such as prompt injection reduced through engineering, "you might not want to give your agent all the permissions". His proposal is to use one agent to authorize another: "You want another agent that can give it permissions or not, and then this agent has to negotiate with another agent." Combining the two systems, he says, greatly increases his trust that he can let the agent rip.
No comments yet. Be the first!