Peter Steinberger, the creator of the open-source AI assistant OpenClaw and now a member of OpenAI’s technical staff, has revealed that he used approximately $10 million worth of AI tokens in a single month. The disclosure came during a conversation with Cloudflare CTO Dane Knecht at the Cloudflare Connect 2026 conference in San Francisco.
The staggering figure highlights the potentially enormous computing costs of running advanced AI agents at scale. Steinberger also acknowledged that not all the work generated by such extensive token usage is useful.
Who Is Peter Steinberger and What Is OpenClaw?
Peter Steinberger created OpenClaw, an open-source AI agent designed to perform tasks beyond answering questions. AI agents can interact with software, execute workflows and automate activities with varying degrees of human supervision.
Steinberger joined OpenAI earlier in 2026, while OpenClaw’s development was set to continue through an open-source foundation.
His work involves experimenting with AI-powered development workflows, where agents can help write code, review contributions, manage tasks and perform other software engineering activities.
Why Did He Spend $10 Million on AI Tokens?
AI companies generally charge for model usage based on tokens, the units used to process text and other information. Costs can rise significantly when an AI system repeatedly analyses large amounts of context, generates code, reviews outputs or runs multiple agents.
During the conference discussion, Knecht noted that Steinberger had access to a virtually unlimited token budget. Steinberger responded that people often fail to understand that not all the resulting work is useful.
The reported $10 million figure is his estimate of the cost of AI token usage for the previous month. It is not evidence that an ordinary OpenClaw user would face a similar bill, nor does the report provide a detailed breakdown of the models, workloads or token volumes involved.
AI Agents Could Move More Work to the Cloud
Steinberger also questioned why AI agents still rely so heavily on users’ computers instead of operating entirely in the cloud.
Cloud-hosted agents could run continuously, coordinate tasks and work without requiring a user’s machine to remain active. However, moving workloads to the cloud introduces other considerations, including infrastructure costs, data privacy and security.
For businesses, the challenge is to determine whether the productivity gains from autonomous AI systems justify their running costs.
Not Every AI-Generated Result Is Useful
The spending disclosure also highlights an important limitation of AI automation: more usage does not automatically produce more value.
Agents may repeat tasks, generate unnecessary output or require additional model calls to check their own work. This can increase costs without delivering proportional improvements in productivity.
Organisations adopting AI therefore need to monitor token consumption, establish budgets and evaluate whether automated workflows are producing measurable results.
AI Agents Need Better Security Controls
Steinberger also discussed the risks of giving AI agents extensive permissions. An agent with access to files, applications or sensitive information can create security problems if it makes an incorrect decision or follows malicious instructions.
He suggested that one agent could act as a permission checker for another, helping determine whether an action is necessary and whether it stays within defined safeguards.
Such an approach could provide an additional layer of oversight, although it would not eliminate the need for conventional security controls and human accountability.
What This Means for the Future of AI
The reported $10 million monthly token spend illustrates both the capabilities and costs of pushing AI agents towards more autonomous work. While businesses are exploring ways to automate software development and other workflows, the economics remain an important constraint.
The next stage of AI adoption will depend not only on more capable models, but also on cheaper inference, efficient infrastructure and better methods for ensuring that automated work delivers genuine value.
Summary
OpenClaw creator Peter Steinberger estimated that he used $10 million worth of AI tokens in one month, according to remarks at Cloudflare Connect 2026. Now working at OpenAI, he discussed the costs of running AI agents, the need for cloud-based infrastructure and the importance of security controls. He also acknowledged that not all AI-generated work is useful, highlighting the challenge of balancing capability, cost and productivity.
