The $10 Million Vanity Project: OpenClaw and the Delusion of 'Unlimited' Compute

In the current gold rush of artificial intelligence, there is a seductive myth that compute is a resource to be conquered through sheer volume rather than efficiency. This delusion reached a fever pitch last week, as first reported by The Register, at the Cloudflare Connect 2026 conference in San Francisco, where Peter Steinberger, the creator of OpenClaw and a member of technical staff at OpenAI, admitted to spending approximately $10 million in tokens on AI workloads in a single month.

To be clear: this is an opinion piece. As a columnist tracking the money behind the platforms, I view Steinberger’s admission not as a badge of productivity, but as a textbook case of token-incentivized vanity spending. When a creator is operating with what Cloudflare CTO Dane Knecht described as a "virtually unlimited token budget," the incentive to build sustainable, lean, and efficient systems evaporates. We are witnessing the emergence of a 'burn-first' mentality that exposes the fragility of creator-led crypto and AI economies.

According to reporting from The Register, the sheer scale of this expenditure caused audible surprise among the conference audience. The most damning part of the exchange wasn't the number itself, but Steinberger’s reaction to the room's shock. As The Register reports, Steinberger attempted to qualify the spend by adding, "What people don't always understand is not all of that is useful."

Let that sink in. A single developer spent $10 million of investor capital in 30 days, and his primary defense is that much of it was useless. In any other sector of the creator economy, spending millions of dollars on "not useful" output would be labeled a catastrophic failure of management. In the venture-backed AI bubble, however, it is framed as the cost of exploration.

This disconnect is further highlighted by Steinberger’s perspective on infrastructure. During his conversation with Knecht, Steinberger mused that the main challenge facing users is that infrastructure "hasn't really caught up yet," questioning why AI tasks are still using local computers and arguing that "this all [should] be in the cloud."

This rhetoric ignores the fundamental economic reality that most creators and developers are not setting venture funds alight. As The Register notes, cost remains a primary consideration for anyone not backed by massive capital injections. By pushing for a cloud-only future while operating on an unlimited budget, Steinberger is advocating for a model that is inaccessible to the very open-source community that OpenClaw ostensibly serves.

Speaking of open source, the conversation between Knecht and Steinberger touched on a critical tension: the "AI slop" currently overwhelming open-source maintainers. Knecht pointed out that many maintainers have stopped accepting pull requests because of the low-quality, AI-generated noise. While Steinberger argues that the solution is a change in how pull requests are handled—specifically by requiring developers to explain their reasoning from "A to B"—he fails to address how the culture of "unlimited tokens" contributes to this slop. When the cost of generation is zero for the creator, the incentive to curate and refine disappears, leaving the maintainer to sift through the wreckage.

Furthermore, the financial backdrop of this spending is precarious. The Register reports that OpenAI is currently attempting to "accelerate funding" as it faces competition from Anthropic, which is expected to launch an IPO. The irony is palpable: while the company seeks more capital to stay competitive, its staff are operating with budgets that allow for millions of dollars in "not useful" token spend.

Steinberger’s vision for the future of AI agents is equally detached from the constraints of the average user. While he doubts that everyone will have hundreds of agents, he envisions a world where intelligence predicts a user's needs before they even ask a question. To achieve this level of predictive autonomy, the token burn would likely dwarf even the $10 million monthly spend.

Even Steinberger’s proposed solution for security—using an agent to authorize another agent to ensure they stay within guardrails—adds another layer of computational overhead. While he claims this "negotiation" between agents increases his trust to "let the agent rip," it is essentially a proposal to throw more tokens at a problem that could be solved with better engineering and tighter constraints.

OpenClaw is being positioned as a leap forward in AI agency, but the way it is being funded and utilized is a cautionary tale. When the cost of failure is borne by investors rather than the creator, the result is not innovation; it is excess. If the AI economy is to survive the inevitable cooling of venture capital, it must move away from the "unlimited budget" mindset. We cannot build a sustainable creator economy on the back of $10 million months where the primary output is "not useful."