Uber has spent its annual AI budget in four months because AI is making us addicted to it

Uber CTO Praveen Neppalli Naga recently explained how his company decided to deploy Claude Code to its 5,000 engineers. Adoption of the tool skyrocketed from 32% to 84% in one month, and everyone started using it so much that Uber ran into a problem: the real cost went from $500 to $2,000 per month per programmer, which destroyed the company’s spending forecasts: In four months the entire annual budget was spent to implement AI in the company. Welcome to the end of AI grants.

Microsoft will also control spending. The Uber case is not an isolated event. Microsoft has virtually unlimited computing resources with Azure. However, has made the decision to withdraw Claude Code’s internal licenses from its developers in the Experiences + Devices division. The reason is twofold: first, they want to curb operating spending before the end of their fiscal year. Secondly, they want to force the use of their own tools with GitHub Copilot as the clear protagonist.

GitHub ends its flat rate. This company, owned by Microsoft, also wants to prepare for the future, and from June 1, 2026 all GitHub Copilot plans abandon their “flat rate” option to move to a usage-based billing model. The base subscription price remains the same, but is converted into “AI credits” that will be consumed as the model is used. If developers use GitHub Copilot intensively, the credits will run out quickly and the system will stop unless we pay extra to continue working. On GitHub they pointed out that “charging a flat rate for autonomous agents is no longer sustainable.”

Source: Hedgie (X).

The graphic that explains it all. An X user named Hedgie warned that this is just the beginning and added a useful image to understand what is happening.

1. The traditional SaaS (Software as a Service) software model works in a straight line. You pay a monthly fee and the server costs barely vary whether you use that app or service for ten minutes or ten hours: the profit margin is predictable, and the load is manageable. This is what happens with “flat rates” for streaming services, for example. Whether you watch more or fewer hours of Netflix doesn’t make a big difference to Netflix’s infrastructure.
2. But agentic AI operates under an exponential curve. As seen in the image, when a programming AI agent like Claude Code starts working, it can use thousands or even millions of calls to the provider’s API (in this case, Anthropic) to receive, process and redeem millions of tokens.
3. The flat rates offered by ChatGPT Plus or Claude Pro are adequate for conversational use of AI, but AI agents devour tokens and consumption skyrockets. That’s why Anthropic, OpenAI, and others put limits on their flat rates and even prohibit their use for agentic tasks (such as those provided by OpenClaw or scheduling agents). There they ask you to pay per use with the API, and that increases the costs.

Crossroads. This situation puts companies like OpenAI, Anthropic and Google in a dilemma. If their clients (like Uber) begin to cut back on the use of AI to protect their budgets, these companies’ revenues may be dampened and that will affect their valuations. The other potential solution is to artificially lower prices to keep those customers happy, but that means absorbing significant operating losses that would harm your profitability.

AI dependence and addiction. These companies are realizing that using AI can be really beneficial, but also expensive. Anthropic or OpenAI’s business model is not new, and we have seen that pattern in the past. A company launches a product or service, often free or very cheap, but after gaining a sufficient volume of users it ends up changing its conditions to charge you more and more for that product or service.

It’s already happened. We have a good example in Google Photos or streaming services, which trapped us and then squeezed us with increasingly higher monthly fees. With AI the scenario is the same: catch us now with reduced costs and then cover the free service and make us pay if we really want to take advantage of it. There will always be alternatives such as using local models or opting for cheaper platforms, of course, but for those who offer the most advanced models and features the strategy is clear.

In Xataka | Nvidia’s financial results are simply dizzying. And it still hasn’t sold a single chip in China