Peter Steinberger, the creator of OpenClaw and now a member of technical staff at OpenAI, told a Cloudflare Connect 2026 audience in San Francisco that he burned roughly $10 million in tokens last month. Then he said the part most people would have kept to themselves: not all of it was useful.
That sentence is the most honest disclosure about the agent economy anyone has made this year, and it is worth taking seriously precisely because the person who said it never sees the bill. The argument I want to make is simple. Token spend is not a capability metric, an unlimited budget is not a productivity strategy, and the $10 million month is a diagnosis of the industry's cost problem rather than a celebration of its progress.
The $10 Million Number Is an Admission, Not a Flex
Cloudflare CTO Dane Knecht framed the budget on stage as virtually unlimited, which is the detail that matters most. A virtually unlimited budget does not make Steinberger a case study for anyone else. It makes him an outlier whose experiments answer questions that only well funded labs can afford to ask.
What does $10 million of agent traffic actually buy? Mostly exploration. Retries, abandoned branches, conversations dragged far past the point of usefulness, scheduled agents that wake up, look around and accomplish nothing, and tool calls that terminate in an error nobody reads. The quote was not false modesty. It was a description of the shape of agent workloads today, and the audience murmuring in response had the right instinct.
That is the uncomfortable part. If a frontier practitioner openly reports that most of his token budget is waste, then the headline cost of every agent deployment in production is being carried in the gap between what the model is asked to do and what it actually needs to do. That gap is not closed by buying a smarter model. It is closed by better scoping, tighter harnesses and honest measurement.
- Retry loops: an agent exploring a dead end for an hour costs the same as one solving a hard problem for an hour
- Context bloat: history carried long past usefulness, billed per token on every single turn
- Idle autonomy: always-on agents that wake on schedule and find nothing worth doing
- Unreviewed output: work generated at scale that no human ever accepts or rejects
There is also a funding footnote that should not be ignored. OpenAI is reportedly moving to accelerate funding ahead of Anthropic's expected IPO. When a company is raising at that scale, a nine figure token habit is a rounding error and the experiment is rational. When a startup with 40 employees tries the same loop, the same tokens are the runway. Same technique, completely different physics.
Cloud, Local, and the Open Source Tab Nobody Wants to Pay
Steinberger also asked why the work still runs on his computer, and why it should not all be sitting in the cloud. I understand the instinct and I still think the framing is backwards. Developers run inference locally for cost, latency and privacy, not nostalgia. Asking why everything is not in the cloud is the kind of question you ask when somebody else covers the meter.
Knecht raised the related problem of open source maintainers who have stopped accepting pull requests under the weight of low quality AI generated contributions. Steinberger's answer was the most useful engineering point of the session. If a contributor shows how they got from A to B, trust in the output goes up dramatically compared with a finished diff thrown over the fence. That is a reviewability requirement, and it points at something bigger than pull requests. An agent that cannot produce its reasoning path is not a contributor, it is a tax on whoever has to review it.
- Cost per completed task is the only agent metric worth quoting, and nobody publishes it
- Reasoning traces are a merge requirement now, not a nice to have
- Agent to agent permission handshakes shift the attack surface rather than remove it
- Local inference is not a phase to grow out of, it is the price of predictable spend
His other claim is that we will not all be juggling hundreds of agents, but each of us will have some always-on intelligence that thinks, helps and occasionally answers a question before we ask it. That is a reasonable prediction with an unreasonable bill attached. An always-on agent is a meter that never stops, and the people who find that prospect exciting are, almost without exception, the people whose tokens are paid for by somebody else.
So the real lesson from the $10 million confession is about visibility. The number was invisible to the person spending it and startling to everyone in the room who pays their own bills. If you want to know whether agentic AI is working for your team, do not ask how many tokens you can spend. Ask how many of the tokens you already spent came back as work you actually shipped. Steinberger has the honesty to admit the answer at $10 million a month. Most of us cannot afford to find out the hard way.
Comments