SemiAnalysis estimates ~90% of Anthropic's business comes from agentic AI, while sources say nearly 25% of its revenue in 2025 came from just two clients (Financial Times)
Powerful AI tools burn through budgets, prompting a rethink of how the technology is priced ahead of frontier lab IPOs.
In early 2026, Uber had already spent its entire annual AI budget, highlighting the volatility and unpredictability of AI costs. Companies are employing leaderboards and usage targets to encourage staff adoption of AI agents, recognizing their ability to execute complex tasks with minimal human intervention as a key to unlocking AI's potential.
However, these autonomous agents can quickly exhaust computing resources, resulting in soaring costs. One Amazon project, for instance, exceeded its budget by 860% after background tasks ran for five months and cost over $1 million. Leading AI labs have partially mitigated this risk by transitioning from fixed subscriptions to uncapped usage-based pricing, though early adopters like Uber, Meta, Cisco, and Walmart have imposed spending caps to manage expenses.
The core issue lies in how AI is priced; current models track computational power used rather than the value generated, prompting both customers and providers to seek alternative pricing structures. OpenAI's president, Greg Brockman, recently advocated for pricing based on tasks completed rather than tokens consumed, emphasizing the importance of value and speed.
Research from SemiAnalysis suggests that a significant portion of Anthropic's business, nearly 90%, stems from agentic AI applications, particularly coding tools. The company's revenue surged from $9 billion at the end of 2024 to $65 billion in July 2025, with projections indicating it could surpass $120 billion by year-end. Yet, Anthropic's IPO prospectus also revealed a concerning reliance on just two major clients, accounting for almost a quarter of its 2025 revenue.
Despite the growing sophistication of AI agents, each interaction with a large language model begins with a standard process that is limited by the model's context window, or working memory. Originally set at 4,096 tokens for GPT-3.5, this threshold has since expanded, enabling more complex tasks but also increasing token usage and associated costs.
The price per token has decreased significantly over the past year, but the volume of usage has outpaced these savings, leading to substantial expenditures for businesses. Tokenization, the process of breaking down text into tokens, is crucial for model training and execution, yet the exact count can vary depending on the tokenization method employed.
The challenge lies in balancing token efficiency with the growing demand for powerful models and the need for robust, outcome-based pricing.
Written by urgent.news from Techmeme's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
