Peter Steinberger, creator of the open-source AI agent OpenClaw and current OpenAI employee, disclosed a $10 million token bill for AI model usage last month during Cloudflare’s Connect conference. He highlighted token costs as a major barrier to broader AI agent automation.
- $10M spent on AI tokens in one month by OpenClaw’s creator
- Token costs remain a key limit to always-on AI agent deployment
- Calls for improved transparency in AI-generated code contributions
What happened
Peter Steinberger, who developed the open-source AI agent OpenClaw and now works at OpenAI, revealed at Cloudflare’s Connect conference that he spent approximately $10 million on AI tokens in the last month. Tokens measure the quantity of text processed by AI models and determine pricing from providers. He explained that while his budget is effectively unlimited, many developers face much tighter constraints.
During a panel discussion with Cloudflare’s CTO Dane Knecht, Steinberger discussed the current state of AI agent infrastructure and token consumption. He pointed out that despite advances, much AI workload still runs on local devices rather than fully cloud-based systems. Steinberger also addressed permissioning challenges for AI agents, advocating for layered control mechanisms where one agent governs another's permissions.
Why it matters
Steinberger’s $10 million token spend illustrates the significant cost barriers limiting the practical deployment of AI agents that operate autonomously around the clock. While some companies impose strict limits on usage—as Accenture caps per employee spending at $29 per month—those limits vastly restrict what AI agents can do in practice. This cost factor remains a critical bottleneck despite advances in model capabilities.
In addition to cost, Steinberger and Knecht addressed trust and safety concerns, highlighting the need for AI permission frameworks and better transparency in how AI-generated code is created and vetted. Open-source maintainers have grown cautious about accepting AI contributions due to a lack of clarity around the reasoning processes behind code generation. Steinberger emphasized that exposing the decision path increases human trust in AI outputs.
What to watch next
Key trends to monitor include the industry’s efforts to reduce token costs and improve AI infrastructure, potentially shifting more AI agent workloads into the cloud and onto specialized hardware like Nvidia’s RTX Spark laptops designed for local AI processing. How companies balance autonomy with control through multi-agent permissioning systems will also be crucial for secure agent deployment.
Another important area is the evolving ecosystem for AI-assisted software development, particularly the standards for transparency and explainability when AI agents generate code. Steinberger’s view suggests future workflows may require AI contributors to document their reasoning explicitly, fostering greater acceptance among human maintainers and smoother collaboration. Meanwhile, funding moves by major AI companies like OpenAI and Anthropic will impact how aggressively these technologies scale.