Since the start of this year, the rapid evolution of AI agents has made “raising AI assistants” a widespread trend. Many tech giants have even encouraged their employees to fully embrace AI advancements. However, this honeymoon phase proved short-lived. As these companies delved deeper into AI, they quickly realized that while AI was powerful, the costs were becoming unsustainable. Recently, news surfaced that ride-hailing giant Uber had depleted its annual AI budget within just four months. Can even Uber no longer afford the tokens?

01 Uber Exhausts Annual AI Budget in Just Four Months

According to reports from NetEase News and Bloomberg, an Uber spokesperson confirmed that the company has imposed spending limits on AI-based programming tools used by employees, including Claude Code and Cursor. The new policy caps each employee’s monthly token consumption at $1,500 per tool. Budgets are not transferable between tools, meaning overspending on one does not affect another’s allocation. Employees can track their usage via an internal dashboard and request exceptions in special cases.

TechCrunch, citing related sources, noted that Uber had previously incentivized employees to use AI extensively, even employing leaderboards to drive adoption. However, the company burned through its annual AI budget within the first few months of the year. After initially deploying AI on a large scale as an internal efficiency experiment, Uber faced a familiar dilemma: who foots the bill, and how to manage expenses? Uber has not disclosed its financial specifics, leaving the exact total AI budget for this year unknown.

Uber’s strategy is not unique in Silicon Valley. GitHub Copilot also adopted a new billing model based on token consumption around the same time, sparking backlash from developers. The key difference is that Uber has directly allocated quotas to individual employees, transforming AI costs from a company-wide budget into a tangible usage limit for each person.

What does a $1,500 monthly cap per tool mean for an employee who relies heavily on AI for daily coding? At the very least, it makes “casually initiating a long task” a decision that requires careful consideration.

02 Why Can’t Even Uber Sustain Token Costs?

Recently, Uber, a leading company in the mobility sector, was reported to have exhausted its annual AI budget in just four months, leading to restrictions on internal AI tool usage. This is not merely a financial dilemma for a tech giant but also a wake-up call for the entire tech industry. How should we interpret this situation?

First, AI agents have completely disrupted the controllable consumption model of traditional large models. During the application phase of traditional large models, corporate AI usage was mostly limited to lightweight tasks such as single conversations, simple Q&A, and basic content generation. Token consumption was characterized by simplicity, fixity, and low frequency, with a clear upper limit on overall usage. For most tech companies, the costs of such shallow AI applications were predictable, quantifiable, and manageable, allowing for precise planning and steady consumption of the annual AI budget.

However, as the industry fully enters the era of AI agents, the application landscape has been reshaped. Autonomous decision-making, multi-round iteration, scenario linkage, and continuous computation have become core features. AI agents are no longer passive tools responding to human instructions but intelligent entities capable of autonomously breaking down tasks, repeated trial and error, and multi-process linked computation. This model directly drives token consumption to grow exponentially. When I personally tested AI agents like Lobster, whether it was OpenClaw or various domestic alternatives, a single instruction often resulted in the instantaneous consumption of hundreds of thousands of tokens. If the instruction required constant monitoring for execution, the consumption was even greater.

For platform-based tech companies like Uber, which operate across a vast array of business scenarios and rely heavily on digitized operations, deploying AI agents across all areas—such as scheduling, customer service, operations and maintenance, data analysis, and business optimization—leads to massive and sustained computational power and token consumption. This consumption has no fixed upper limit or stable cycle but continuously accumulates with business operations, ultimately rendering the company’s established annual AI budget system ineffective. Exhausting the entire year’s budget in a short period becomes an inevitable outcome. This is also a structural cost challenge that all platform-based companies will face in the era of AI agents.

Second, the severe imbalance in the input-output ratio of tokens is the biggest issue. In economics, investments must yield corresponding marginal returns. However, the current development of AI agents is still in its wild growth phase, with logical reasoning and planning capabilities far from mature. This has led to an awkward situation: AI agents often get stuck in “dead loops” or “ineffective reasoning” when executing tasks. They may repeatedly circle around a single logical node or call upon massive amounts of irrelevant data to complete a trivial subtask. They may even engage in meaningless bickering in multi-agent collaborations.

At the core of this issue lies the ineffective burning of massive tokens. Companies are footing the bill for these meaningless internal consumptions, but business efficiency has not seen a substantial improvement, nor has there been corresponding growth in revenue. It’s like a factory introducing supposedly state-of-the-art automated equipment, only to find that the equipment spends most of its time idling or producing defective products, consuming huge amounts of electricity and materials, while the yield rate and output remain stagnant. This severe imbalance in the input-output ratio is the core reason why Uber truly feels the pain. AI has not become a catalyst for profits but has instead become a pure cost consumer on the financial statements, undoubtedly causing even greater concern among the company’s management.

Third, the lack of diminishing marginal costs in the token economy is the root of the problem. Looking back at the tech industry’s history, whether in hardware manufacturing or software development, the underlying logic of business models has always been “economies of scale.” As the number of users or output increases, fixed costs amortized become lower and lower, with marginal costs infinitely approaching zero. The benefits of “diminishing marginal costs” in the internet era have made tech companies naturally optimistic about new technologies. We often see products starting out expensive but becoming increasingly affordable as technology advances, with costs even approaching zero.

However, in the token economy, this classic economic principle fails. The reasoning costs of large models are real computational power consumption, with each generation requiring high-intensity GPU operations. Every time you use it, you have to pay for the computational power rental. If a thousand people use it a thousand times, the costs stack up linearly or even super-linearly. There are no economies of scale, only rigid expenditures.

This means that as the scale of AI applications expands, the computational cost curve for companies continues to climb and is difficult to amortize. When AI becomes infrastructure, companies are essentially operating under a pay-per-use model, which is extremely fragile in terms of business logic. Without the moat of diminishing marginal costs, so-called large-scale AI adoption is akin to building a skyscraper on sand dunes. The larger the scale, the higher the risk of financial collapse.

Fourth, reconstructing the cost logic of AI is the most crucial task for every company. Over the past two years, the tech industry’s attitude towards AI can be summed up in four words: “Just do it first.” Many companies adopted AI not because they had figured out how to use it but because they were afraid of falling behind. This mindset might have been acceptable in the early stages of technology, but when AI enters the deep waters and costs start to materialize, this mindset becomes fatal.

After years of rapid development, the AI industry has bid farewell to the initial stages of wild growth, blind deployment, and burning money to change tracks. The development model that solely pursues technological advancement, full scenario coverage, and maximum intelligence has completely failed. In the future, the core competition in AI commercialization will no longer be about who uses AI more or deploys it in more scenarios but about who uses AI more precisely, at lower costs, and with higher outputs.

All tech companies must abandon the inertial thinking of “blindly embracing AI” and return to the essence of industrial operations. They should plan AI investments with cost constraints as a prerequisite and evaluate AI value based on the input-output ratio. On the one hand, companies need to establish a refined AI usage control system, screen high-value, high-return AI application scenarios, eliminate ineffective, low-efficiency, and high-consumption intelligent agent applications, and put an end to meaningless computational power and token internal consumptions. On the other hand, they should continuously optimize AI application models, promote technology reuse, scenario intensification, and computational power sharing, weaken the rigid nature of AI costs, and gradually build a commercial AI model with diminishing marginal costs.

In the final analysis, AI has never been a strategic gimmick for companies, nor is it a tool for following trends. It is a core productivity that serves corporate profitability. Any AI that cannot continuously create value for companies is destined to be eliminated by the market. Ultimately, companies will return to rationality and only pay for AI and computational power that truly generate profits.

END