Microsoft curbs excessive AI use after promoting it widely

Microsoft curbs excessive AI usage after promoting it everywhere

Satya Nadella, the CEO of Microsoft, acknowledged in June 2026 that his teams had a tendency to overuse the most powerful artificial intelligence models for relatively simple tasks. This phenomenon, known as tokenmaxxing, involves systematically using the most expensive AI model, regardless of the complexity of the task at hand.

During a podcast, Nadella was asked about the practice of tokenmaxxing within Microsoft. He candidly replied that it was happening "a lot." This excessive use of costly models contributes to inflating Microsoft's expenses, which is already planning an investment of $190 billion in AI infrastructure for the year 2026, approximately €174 billion.

Instead of restricting access to AI, Microsoft has opted to impose internal cost discipline. Nadella admitted to being a "tokenmaxxer" himself, describing this practice as addictive. He provided clear instructions: "Do not use advanced models for non-advanced problems." Indeed, the most sophisticated models charge significantly more per token than their lighter versions, without necessarily delivering better results for simple queries.

Adoption of Copilot's Auto mode

To avoid overconsumption, Microsoft has directed its teams towards the Auto mode of Microsoft Copilot. This mode automatically selects the most suitable AI model for each request, relieving the user of that decision. This is a significant shift for a company that, for the past two years, has integrated AI into its flagship products like Windows, Office, and Azure, encouraging unrestricted use.

Engineers in the Experiences & Devices division must abandon the Claude Code tool from Anthropic before June 30, 2026, the end of the fiscal year. This tool, introduced in December 2025, quickly replaced GitHub Copilot CLI in daily use, leading to excessive token consumption.

Financial and cultural consequences

Without usage caps, some companies have spent hundreds of millions of euros on AI tokens within a few weeks. In large tech companies, internal rankings measure productivity by the volume of tokens processed, as seen at Amazon. This type of metric encourages consumption over the quality of work.

Currently, 30% of Microsoft's code is generated by AI, according to Nadella. However, this integration has not prevented excesses. AI usage must be regulated to control costs, clearly defining which task requires which level of model.

A striking example is that of an unidentified company that spent €460 million in one month on Claude, approximately $500 million, in the absence of usage caps. Uber also exhausted its annual budget for AI tools in just four months, according to its chief technology officer.

For Nadella, the essential question for his teams is: "What am I trying to create?" Rather than focusing on the number of tokens consumed or the model used, the emphasis should be on the value of the outcome achieved.