Microsoft cancels Claude Code licenses and exposes the real problem of AI costs in business
Microsoft is at the center of a conversation nobody in tech expected to be having this soon: using AI in day-to-day corporate operations can cost more than keeping human employees on the payroll.
Sounds contradictory, right?
The promise was always that artificial intelligence would cut costs, streamline processes, and make companies leaner and more competitive. But the reality starting to surface behind the scenes at major corporations tells a very different story — and one quite a bit more complicated than those presentation slides made it seem. 🤔
The latest case came from inside Microsoft itself, which decided to cancel most of its Claude Code licenses after the tool became popular way too fast among engineers, designers, and product managers. According to The Verge, the company is steering its developers toward GitHub Copilot CLI as an internal alternative.
And the Redmond giant is not alone in this situation.
From Uber to Nvidia, along with Meta and Amazon, the pattern repeats: the more companies push heavy AI usage, the more the bill catches them off guard — in a bad way.
What happened with Claude Code inside Microsoft
Claude Code is an AI tool developed by Anthropic, designed to help technical teams with writing, reviewing, and debugging code. When it reached Microsoft’s internal teams about six months ago, adoption was almost immediate. Thousands of software engineers, product designers, and project managers started integrating the tool into their daily workflow, and usage scaled organically and at a pretty aggressive pace.
The problem is that rapid growth came with an equally hefty invoice — one that quickly caught the attention of the executives managing the tech budget.
The decision to cancel most of the licenses was not driven by quality or performance concerns. Quite the opposite — Claude Code was delivering results, and the engineers themselves had already come to rely on it day to day. The core issue was financial: the volume of usage generated by employees pushed operational costs to a level that was never part of the original plan. When you multiply the price per token or per API call by the number of interactions an entire team makes throughout the day, the final number can catch any CFO off guard — and that is exactly what happened.
It is worth noting that, according to The Verge, the cancellation of Claude Code licenses does not affect the broader agreement between Microsoft and Anthropic under the Foundry program. That deal includes an investment of up to 5 billion dollars from Microsoft in Anthropic, access for Foundry customers to Claude models, and Anthropic’s commitment to purchase 30 billion dollars in computing capacity on Azure. In other words, the strategic partnership between the two companies remains solid — what changed was the unchecked internal usage of the tool. 💸
Receive the best innovation content in your email.
All the news, tips, trends, and resources you're looking for, delivered to your inbox.
By subscribing to the newsletter, you agree to receive communications from Método Viral. We are committed to always protecting and respecting your privacy.
The episode raises an important alarm about how companies are approaching the adoption of generative AI. Many organizations open up access to tools without a clear usage governance strategy, without defined consumption limits, and most importantly, without a real understanding of how costs scale as engagement ramps up. Microsoft, even as one of the largest investors in the global AI ecosystem, fell into that trap inside its own house.
Uber blew through its AI budget in just four months
Microsoft is not the only major company dealing with this problem. The Uber case might be even more revealing. According to a report from The Information, the company’s CTO, Praveen Neppalli Naga, revealed in April that Uber had already burned through its entire budget for AI coding tools earmarked for 2026 — in just four months.
This happened after Uber itself actively encouraged adoption of these tools. The company went as far as creating internal leaderboards that ranked teams by their level of AI tool usage. The logic was simple: more usage, more productivity. But the financial logic did not keep up with the engagement logic, and the result was a budget blowout that forced the company to completely rethink its adoption strategy.
This type of incentive was not exclusive to Uber. Meta jumped on the same bandwagon. One Meta employee built an internal dashboard called Claudeonomics — a nod to Anthropic’s Claude model — to track which employees were using the most AI at work. Meanwhile, Amazon started encouraging its teams to practice what they internally called tokenmaxxing, meaning using as many AI tokens as possible — the basic units of computation that power language models.
The race toward maximum AI usage turned into almost a cultural competition inside big tech. But the problem is that in a token-based pricing model, more usage automatically means more spending. 📈
When AI costs more than an employee’s salary
This phenomenon raises a question few were willing to ask out loud until recently: what if AI is, at least for now, more expensive than the people it was supposed to replace or complement?
Bryan Catanzaro, VP of applied deep learning at Nvidia, was pretty blunt about this in a recent interview with Axios. According to him, for his team, the cost of compute already far exceeds the cost of the employees. When someone from Nvidia itself — the company profiting the most from the AI revolution — makes that statement, you simply cannot ignore the message.
The pricing model of major AI platforms — typically based on consumption per token, per request, or via premium per-user subscriptions — was designed to be accessible at the entry point but can scale aggressively as usage intensifies. When a company has hundreds or thousands of employees using these tools continuously throughout the workday, the aggregate cost can quickly blow past the tech budget. And worse: in many cases, the overrun is only noticed after the bill has already arrived.
There is yet another factor that complicates this equation: the difficulty of measuring the real return on investment. Unlike an employee whose performance can be evaluated through delivery metrics, speed, or quality, the value generated by an AI tool is much more diffuse and subjective. An engineer might use Claude Code to solve a problem in 20 minutes that would have taken two hours without the tool, but consistently quantifying that and turning it into a concrete financial number is a real challenge for any manager.
When the benefit is hazy and the cost is explicit on the invoice, the decision to cut licenses becomes much easier to justify. 📊
The token paradox: unit price drops, but the total bill keeps climbing
There is an interesting paradox in this whole story that deserves attention. The individual cost of each AI token is falling — and it should continue to fall. A recent Gartner report estimates that by 2030, inference on a language model with one trillion parameters will cost AI companies nearly 90% less than it did in 2025. In other words, the raw material of AI is getting cheaper.
But here is the detail that changes everything: more advanced models, especially AI agents, consume far more tokens per task than standard models. According to projections from Goldman Sachs, agentic AI could drive a 24x increase in token consumption by 2030, reaching a staggering 120 quadrillion tokens per month as consumers and businesses adopt AI agents at scale.
That means even with cheaper tokens, total enterprise spending could rise significantly. Gartner was very clear on this point: cheaper tokens will not translate into cheaper enterprise AI. The reasons boil down to three factors:
- Agentic models require far more tokens per task
- The increase in consumption can outpace the drop in unit cost
- AI providers will not fully pass cost reductions on to customers
Will Sommer, senior analyst at Gartner, was emphatic in warning that product leaders should not confuse basic token deflation with the democratization of frontier reasoning. In other words, just because the token got cheaper does not mean using cutting-edge AI became affordable.
The dream of 100 AI agents per employee could get expensive
This rising cost scenario seriously complicates some of the market’s most ambitious plans. Nvidia CEO Jensen Huang has publicly stated that he envisions a future where 100 AI agents work alongside every employee in a company. He is part of a larger wave of CEOs promoting the vision of a corporate environment dominated by digital workers operating across every area of the business.
The vision is exciting, no doubt. But if token consumption scales faster than the drop in unit costs — and everything suggests that will be the case in the short and medium term — that reality could come with a price tag far heavier than what executives are projecting in their financial models.
This is not to say that agentic AI will not happen. It will. But the path between the vision of 100 agents per person and operational reality involves solving an economic problem that still does not have a definitive answer. And the cases of Microsoft and Uber show that even the wealthiest and most technologically advanced companies on the planet are stumbling over this bill. 🤖
What this says about the future of AI in business
The situation Microsoft and other tech giants are facing does not mean corporate AI is a scam or that the promise of efficiency was a lie. What it reveals is that the transition to an AI-integrated workplace is more complex than it appeared in product demos, and that companies are still learning how to manage this new variable within their business models. There is a learning curve, and it comes with a cost.
One of the main challenges emerging from this scenario is the need to create internal AI usage policies that balance access and control. Opening up tools to everyone with no criteria can lead to chaotic adoption and runaway spending, just like what happened with Claude Code at Microsoft. On the other hand, restricting access too much can produce the opposite effect: teams that could gain real productivity end up being cut off from resources that could genuinely transform their work. Finding that balance is one of the most important — and least glamorous — tasks for technology departments right now. 🎯
The corporate AI market is still maturing, and platform pricing models should evolve as pressure from large customers increases. There are already moves in that direction: some AI companies are starting to offer enterprise plans with consumption caps, flat-rate subscription models per user, or volume-negotiated contracts. But until these formats become the standard, organizations need to treat AI spending with the same financial rigor they apply to any other operational cost line — and not simply as an innovation investment that justifies any number on the spreadsheet.
The role of AI governance in cost control
What is becoming increasingly clear is that the adoption of generative AI in business needs to be accompanied by a robust layer of governance. It is not enough to buy licenses, open up access, and hope for the best. It is essential to define who uses it, what they use it for, how much they can use, and how to measure whether that usage is generating real value for the business.
Companies that are navigating this transition phase more successfully generally share a few common practices. They set consumption limits per team or per project. They establish clear metrics to evaluate AI return on investment. They create internal committees that periodically review tool usage and costs. And perhaps most importantly, they treat AI like any other corporate resource that requires active management — and not like a magic wand that solves everything on its own.
The absence of these practices is precisely what led Microsoft and Uber to the situations now making headlines across the industry. And the takeaway is valuable for any organization in the process of adopting AI at scale: technology alone does not deliver results if there is no strategy behind it.
The Microsoft and Claude Code episode is, above all, a reminder that adopting new technology smartly goes far beyond simply flipping the switch for everyone and hoping productivity shows up in the quarterly results.
AI has real potential to transform the way companies operate — but that potential needs to be managed with strategy, data, and most importantly, with eyes wide open to what shows up on the invoice at the end of the month. Neither Anthropic nor Microsoft had officially commented on the situation at the time of the original reporting by Fortune. 👀
