What prompted the $10 million token expenditure claim?
Peter Steinberger, the creator of the OpenClaw AI agent platform, admitted to utilizing a virtually unlimited token budget to run intensive AI workloads. This claim was made during a conversation with Cloudflare CTO Dane Knecht at the Cloudflare Connect 2026 conference held in San Francisco. Steinberger has been collaborating with OpenAI on the development of next-generation personal AI agents since February 2026.
The scale of this spending has drawn scrutiny because the costs are largely absorbed by OpenAI’s investors rather than the developer himself. According to the details shared at the conference, the $10 million monthly figure represents the sheer volume of data processed to train and refine complex agentic behaviors. While Steinberger appears focused on the technical frontier of AI capabilities, observers have noted that such profligate use of capital may seem disconnected from traditional business realities.
The context of this spending is further complicated by the current competitive landscape of the AI industry. While OpenAI continues to support high-cost research and development through massive investment, rivals such as Anthropic are reportedly preparing for an initial public offering (IPO). This suggests a diverging path between companies focusing on rapid, high-cost iteration and those moving toward traditional public market structures.
The perspective of industry leaders
Dane Knecht, the CTO of Cloudflare, provided context to Steinberger's comments, noting that not everyone in the professional AI community has access to such a massive budget. Knecht observed that while high token usage is necessary for certain breakthroughs, "not all of that is useful," suggesting a distinction between productive computation and inefficient resource consumption.
How does Steinberger view the future of AI infrastructure?
Steinberger believes that current infrastructure is lagging behind the needs of advanced AI agents, arguing that the industry must move toward seamless, cloud-based compute. He expressed frustration with the current reliance on local hardware, questioning why significant AI tasks still require significant local processing power instead of being fully handled in the cloud.
His vision for the future involves a shift where the distinction between a user's device and the AI's computational engine disappears. He posits that the primary challenge facing developers today is an infrastructure structure that has not yet caught up with the demands of agentic workflows. For Steinberger, the goal is a cloud environment that provides ubiquitous, frictionless access to intelligence.
This transition is essential for the next phase of AI evolution, which Steinberger predicts will move from reactive tools to proactive assistants. He envisions a world where intelligence is integrated into every facet of daily life, functioning as a continuous presence that thinks and anticipates needs. This evolution would shift the user experience from actively querying a system to receiving answers before a question is even formulated.
What is the role of multi-agent systems in AI reliability?
To manage the complexity and risks associated with autonomous agents, Steinberger proposes a dual-agent architecture designed to ensure safety and accuracy. He suggests that a single agent is insufficient for high-stakes tasks; instead, a system should consist of at least two distinct entities: one to execute the work and another to act as a supervisor.
In this proposed model, the second agent serves a critical governance role. It is responsible for reviewing the actions and outputs of the primary agent, providing or withholding permissions based on established guardrails. This creates a system of checks and balances that mimics human organizational structures, where oversight is a prerequisite for autonomy.
Negotiation and guardrails
The interaction between these two agents is not merely a one-way command but a process of negotiation. Steinberger noted that agents often realize a task might be better handled differently or that a specific action might violate safety protocols. By having one agent question the other, the system can identify simpler, safer paths that remain within predefined constraints.
This collaborative approach is intended to build user trust. Steinberger argued that when these two systems work in tandem, it provides the confidence necessary to "let the agent rip," allowing for much higher levels of autonomy without the fear of unmanaged errors or unpredictable behavior.
Why is token consumption a central debate in AI development?
The massive scale of token consumption highlighted by Steinberger underscores the economic tension at the heart of the AI boom. Tokens, which represent the basic units of processing in large language models, are the primary currency of AI development. As agents become more complex and capable of long-running, autonomous tasks, their hunger for these tokens grows exponentially.
This creates a divide between two different development philosophies:
- The Research-Driven Approach: Characterized by high-capital expenditure and the pursuit of capability at any cost, often supported by massive venture capital or corporate investment.
- The Efficiency-Driven Approach: Focused on optimizing model performance and reducing the cost-per-token to ensure commercial viability and scalability.
The $10 million monthly figure serves as a benchmark for the extreme end of the research-driven spectrum. As the industry matures, the ability to achieve similar agentic intelligence with significantly lower token requirements will likely become a key competitive advantage for developers seeking to move beyond the experimental phase into widespread commercial adoption.
Frequently asked questions
What is an AI token?
An AI token is a fundamental unit of text or data used by large language models to process information. Instead of reading words, models break text down into these smaller segments to perform computations. The more complex the task, the more tokens are consumed during the interaction.
Why did Steinberger spend $10 million in tokens?
Steinberger used the tokens to power the intensive workloads required to develop and test next-generation AI agents. This massive expenditure supports the high-level reasoning and continuous processing capabilities necessary to move AI from simple chat interfaces to autonomous, proactive agents.
How does the dual-agent system work?
The system utilizes two separate AI agents to ensure safety. One agent focuses on performing the requested tasks, while the second agent monitors the first, checking its work against guardrails and negotiating permissions to ensure the actions are appropriate and safe.
What is the difference between cloud and local AI?
Local AI runs on the user's own hardware, such as a computer or smartphone, which can limit its power. Cloud AI runs on massive remote servers, providing much higher computational capacity but requiring a constant connection and incurring costs like token usage.
What is the significance of the Cloudflare Connect 2026 conference?
The conference served as the venue where these industry insights were shared. It provided a platform for leaders like Steinberger and Knecht to discuss the intersection of AI agent development, cloud infrastructure requirements, and the economic realities of the current AI landscape.
Key takeaways
- OpenClaw creator Peter Steinberger reported a $10 million monthly token spend for AI workloads.
- Steinberger advocates for a transition from local computing to seamless, cloud-based AI infrastructure.
- The proposed AI agent architecture involves a dual-system approach for improved safety and trust.
- The massive token consumption is currently being funded by OpenAI's investors rather than the developer.
The future of autonomous intelligence
The revelations from the Cloudflare Connect 2026 conference highlight a pivotal moment in the AI industry. While the sheer scale of Steinberger's token spending may appear unsustainable to some, it reflects the intense computational demand required to bridge the gap between current models and truly autonomous agents. As the industry moves toward the dual-agent architectures and cloud-centric infrastructures described, the focus will likely shift from simply increasing model size to improving the efficiency and reliability of agentic workflows. The ultimate success of these technologies will depend on whether they can move from high-cost experimental phases to scalable, trustworthy tools for everyday use.
