IT Home reported on October 23 that Peter Steinberger, the founder of OpenClaw and a member of OpenAI’s technical team, revealed that he spent up to $10 million on Token costs for AI work last month (IT Home note: at current exchange rate, this is approximately 66.882 million RMB).
Steinberger revealed this figure during an interview with Cloudflare’s Chief Technology Officer Dane Knecht at the Cloudflare Connect 2026 conference in San Francisco. The related costs are borne by investors. In response to the audience’s surprised reaction, Steinberger added that not all token consumption generates actual value.
Knecht mentioned in the interview that Steinberger had previously admitted to having a “almost unlimited Token budget,” and not everyone at the event had such a resource.
When discussing AI Agent infrastructure, Steinberger believes that the existing infrastructure has not yet kept up with the development of AI technology. He asks why some workloads still need to run on one’s own computer, rather than being migrated entirely to the cloud. However, cloud-based AI Agents may incur high operating costs, in addition to privacy and compliance issues. He revealed that he spent up to $10 million in tokens on AI work last month.
The conversation also covered the impact of AI-generated code on the open-source community. Knecht mentioned that some open-source project maintainers are troubled by a large amount of low-quality AI-generated code, and even stop accepting external pull requests (PRs). He asked Steinberger whether the current trend means a decline in open-source or a new stage of development.
According to Steinberger, the way code is submitted needs to change. Contributors should not just submit the final code, but also explain how they derived the solution from the problem. He said that if one can see the contributor’s thought process from the starting point to the final result, his trust in the output will be higher, rather than receiving a finished code without any explanation.
Regarding the extent of the future adoption of AI agents, Knecht asked whether people will be running hundreds of AI agents at the same time in a few years. Steinberger was cautious about this. He predicted that in the future, everyone might have some kind of intelligent assistant that can operate around the clock, think continuously, and provide help. It could even predict what users may need in advance, so users don’t have to ask questions themselves and receive the answers directly.
At the end of the dialogue, both parties discussed issues related to trust and permission management for AI Agents. They talked about which data users should provide to the Agent, as well as how to limit its access rights. Steinberger said that this largely depends on the level of trust users have in the underlying model, since the behavior of the model is not completely constrained by external control code (Harness). He also pointed out that as the model’s capabilities improve, some mechanisms he designed for OpenClaw in the early stages may also change.
Regarding security risks such as Prompt Injection, Steinberger believes that even if engineering measures can reduce the risk, users should not grant all permissions to the Agent. He proposes that one Agent could be responsible for approving the permissions of another Agent, and through negotiation among multiple Agents, it would be determined whether a particular operation is necessary or if a safer alternative can be found.
Related Reads: