OpenClaw creator says he spent $10 million in tokens last month

OpenAI technical staff member Peter Steinberger said he spent about $10 million in tokens running AI workloads last month, with the cost funded by investors. At Cloudflare Connect, he discussed cloud infrastructure, the limits of AI-generated pull requests, and trust and permissions for AI agents, including a proposal for agents to authorize one another. His comments highlight both anticipated productivity benefits and unresolved cost, privacy, and safety concerns.

Analysis

The economically relevant signal is not one developer’s bill; it is the tension between agent usage growth and the cost, security, and governance required to make that usage repeatable. Heavy experimentation can inflate token demand without proving durable, paid enterprise workloads. In the near term, treat usage anecdotes as weak evidence for cloud-infrastructure revenue: verify paid inference volumes, customer retention, and gross-margin trends before extrapolating. Over 1–3 months, funding pressure and any IPO timetable could sharpen scrutiny of AI spend efficiency; a shift toward smaller models, caching, or local execution would temper cloud demand even as agent adoption rises.

The less obvious beneficiary may be the control layer—identity, permissioning, audit logs, and evaluation—rather than raw model capacity. But “agent authorizes agent” is not a security moat if both agents share model vulnerabilities, context, or incentives. Enterprise deployment still hinges on independently enforceable permissions and auditable outcomes. Open-source maintainers facing low-quality submissions could also make trusted, curated code and review workflows more valuable, while raising friction for projects dependent on volunteer throughput.

Contrarian view: the anecdote may overstate near-term inference economics; much of the spend may be wasteful exploration, not recurring demand. Conversely, investors may underprice the eventual value of agents that act continuously, if reliability and cost per completed task improve. No clean single-name trade follows from this conference discussion alone. Falsify the cautious view with sustained paid usage and improving inference economics; falsify the bullish agent thesis with weak renewal, rising cost per task, or material security incidents.

Key Decisions for Investors

• Do not trade cloud or AI infrastructure on the reported spend alone. Track paid inference consumption, gross margins, and cost per completed task at providers before treating usage as a durable revenue signal.
• Build a watchlist of identity, observability, and AI-security control-plane vendors; seek evidence of enforceable least-privilege controls, auditability, and independent authorization rather than relying on agent-to-agent assurances.
• For the next 1–3 months, monitor funding and IPO developments for signs that AI labs are prioritizing inference efficiency. A credible shift toward local execution, smaller models, or aggressive cost controls would be a downside catalyst for cloud-demand assumptions.
• Over 6–18 months, favor companies able to demonstrate secure enterprise deployment and recurring paid workloads over those relying on agent-count or token-usage narratives. Reassess if renewals and unit economics fail to improve.