Peter Steinberger, a technical staff member at OpenAI and the developer behind the OpenClaw project, disclosed that he expended approximately $10 million worth of computational tokens during the preceding month. The admission emerged during a Cloudflare Connect 2026 session in San Francisco, where he conversed with Cloudflare chief technology officer Dane Knecht.

Knecht had previously noted Steinberger's acknowledgment of possessing "a virtually unlimited token budget," a condition not shared by most conference attendees. When audience members reacted audibly to the scale of expenditure, Steinberger remarked that much of the consumed capacity yielded limited practical value.

The discussion turned to infrastructure readiness, with Steinberger questioning why AI workloads continued to rely on personal computing resources rather than fully migrating to cloud environments. He highlighted cost considerations for organizations lacking venture capital backing, contrasting cloud-hosted AI agents' expense with the prospect of local execution.

Knecht referenced the OpenClaw initiative and the growing reluctance of open-source maintainers to accept pull requests amid concerns over AI-generated code quality. Steinberger argued that the pull-request workflow requires revision, emphasizing that transparent documentation of reasoning behind code changes would increase trust in contributions. He suggested that future AI assistants might operate autonomously, generating solutions without explicit user prompting.

When addressing trust in AI agents, Steinberger acknowledged the difficulty of determining appropriate data access levels. He pointed to architectural choices made during OpenClaw's development, noting that enhanced model capabilities can alter previously hardcoded behaviors. While recognizing improvements in areas such as prompt-injection mitigation, he cautioned against granting unrestricted permissions to autonomous agents. Instead, he proposed a layered authorization model where specialized agents evaluate and allocate capabilities to other agents, potentially introducing self-regulatory checks within AI workflows.