---
title: "AI budgets in the red zone: how Uber and Microsoft are fighting the explosive growth of token costs"
description: "Major IT giants, including Uber and Microsoft, are facing an AI spending crisis. Budgets are melting due to the explosive consumption of tokens by agentic systems, and the promised savings on salaries do not yet justify the costs. 📉🤖"
date: 2026-06-01T10:26:43.000Z
lang: en
url: https://xab.info/en/posts/ai-budgets-in-the-red-zone-uber-and-microsoft-fight-token-costs
tags: []
publisher: "XAB.info"
---

# AI budgets in the red zone: how Uber and Microsoft are fighting the explosive growth of token costs

![Massive data center at sunset, symbolizing the explosive growth of AI token costs for Uber and Microsoft](https://xab.info/media/2026/06/01/uber-i-microsoft-krizis-raskhodov-na-ii/uber-i-microsoft-krizis-raskhodov-na-ii-1.webp)

The era of unchecked artificial intelligence growth in the corporate sector has collided with the harsh reality of economics. Major tech giants, previously actively implementing neural networks, are forced to rethink their strategies. The cause is not the efficiency of algorithms, but astronomical bills for tokens, which have become one of the main expense items for the industry.

### Uber and Microsoft: from enthusiasm to strict economy

The situation has reached a critical point, as evidenced by internal processes among market leaders. Uber experienced a real shock: the company's CTO, Pravin Nepal Nagi, reported that the entire AI budget until 2026 was spent in just a few months. Operating Director Andrew McDonald confirmed in an interview with Business Insider that no direct link was found between the volume of tokens consumed and the emergence of useful features for users.

The company acknowledges the paradox: the volume of code written has increased, but drawing a clear line between the number of lines and real software improvement is impossible. Microsoft is taking similar measures. The corporation has begun revoking developers' access to the external assistant Claude Code, planning to switch employees to the internal Copilot CLI tool by June 30. Officially, this is explained by consolidation around their own products, however, the coincidence with the end of the financial year indicates a desire to cut costs.

### The phenomenon of "agentic" AI and the cost gap

The main driver of cost growth has been the shift to agentic AI. Unlike ordinary chatbots, agents can consume more than 1,000 times more tokens. Goldman Sachs forecasts that in the coming years, token consumption for such systems could increase by more than 24 times.

Companies are trying to justify costs with figures: the head of Airbnb stated that 60% of code is generated by AI, Chime — 84%, and Google — 50%. Uber also reports that more than 60% of code is created by neural networks. However, according to the company's own assessment, the effect does not yet cover the costs. The example of Peter Steinberger's team (OpenAI) demonstrates the scale of expenses: three people spent more than $1.3 million on tokens in one month.

### The paradox of efficiency and the race for new hardware

Today, the cost of AI is growing faster than the salaries of employees it is supposed to replace. This calls into question the justification for personnel cuts conducted under the slogan of increasing efficiency. Investors and analysts are hoping for a "race for new hardware." Goldman Sachs points out that a massive increase in efficiency from new-generation chips should make AI cheaper, allowing investments to continue.

Nvidia is actively promoting the Vera Rubin platform, which promises up to 10 times greater performance per watt. However, reality is making corrections: more than 50% of announced data centers based on Blackwell architecture have been cancelled or delayed. At the end of 2025, Google, Oracle, and Microsoft adjusted their plans, stating their intention to operate equipment for six years before replacement.

In the short term, even giants like Microsoft and Uber are forced to restructure their work with AI. The explosive growth in the number of requests from agentic systems cannot be compensated by an increase in hardware efficiency, until the mass adoption of which is still several years away.