AI development costs revealed

Peter Steinberger, the creator of OpenClaw, disclosed at the Cloudflare Connect 2026 event in San Francisco that he spent roughly 10 million dollars in AI tokens in September 2026. These expenses were funded by investor money, though Steinberger indicated that much of this spending did not result in tangible outcomes.

Unlimited budgets not common

Steinberger mentioned that he had an effectively unlimited token budget, a situation not shared by most attendees at the event. This disparity highlights the uneven access to resources within the AI industry and raises questions about the sustainability of high AI spending for organizations without similar financial backing.

Industry constraints and questions

The discussion at the event focused on the material limitations of the AI industry, with Steinberger questioning the continued need for local machines to handle AI tasks. He pointed out that cloud-based solutions could be used instead, though they come with costs and privacy concerns, especially for those without substantial venture capital.

Token prices drop sharply

According to a September 2026 report, the price of AI tokens has dropped significantly due to increased competition and lower production costs. The rise of freely available Chinese models and strategic price cuts have created deflationary pressure, benefiting end-users but challenging the profit expectations of major industry players.

Rapid cost decline outpaces other technologies

Epoch AI, an independent research institute, reported that AI costs are decreasing faster than those of other transformative technologies like DNA sequencing or lithium batteries. The cost of achieving a specific level of cognitive performance by an AI model has dropped by about 47% per quarter, or 13 times per year, a trend the institute calls "the cost of thought."

Strategic challenges in a declining cost environment

The rapid decline in AI costs presents a major strategic challenge for AI laboratories and application developers. Analysts note that maintaining customer loyalty and generating recurring revenue becomes harder when prices fall quickly and alternatives appear at much lower costs. However, methodological concerns, such as the reliability of evaluation benchmarks, suggest caution in interpreting these trends.

Cost distinction in AI deployment

It is important to differentiate between the "cost of pure thought," which refers to the price of achieving a specific cognitive score, and the "cost of trust," which includes the full cost of deploying AI in real-world scenarios. This includes multiple requests, verification systems, and human validation, all of which add to the overall expense.

Shift toward open-weight models

More companies are adopting open-weight models to gain flexibility and reduce costs, although closed models still dominate in complex tasks. The decision to use AI is not based solely on price or computational power; factors like interface familiarity, integration into existing systems, data security, and organizational stability often lead customers to stick with their current providers.

Future of open source and AI growth

During the conference, Dane Knecht of Cloudflare asked Steinberger if the industry was moving toward the end of open source or an age of AI prosperity. Steinberger suggested that the management of pull requests must evolve, advocating for greater transparency to build trust in AI results. He also doubted that everyone would have hundreds of AI agents in the future, suggesting that individuals might instead have a form of AI that can function, think, and assist throughout the day.

Managing AI agents and trust

Steinberger discussed the challenges of managing security and cross-control between AI agents, noting the difficulty of deciding which data to entrust to an AI. He emphasized that the level of trust in the underlying model is crucial, as many decisions in OpenClaw can now be adjusted with improved models. He advised caution and proposed establishing supervision between multiple AI agents to ensure compliance with limits and increase trust in their autonomy.