Microsoft cancels AI licenses and exposes a problem nobody wanted to face: the technology costs more than paying employees
Microsoft is at the center of a discussion that promises to shake up the tech industry in the coming months — and possibly for years to come.
The promise that sustained billions in investments was simple: AI would reduce costs, eliminate repetitive work, and supercharge team productivity. But reality is delivering a bill quite different from what any optimistic projection had anticipated.
Instead of saving money, large companies are watching their tech budgets evaporate in a matter of months — sometimes weeks. The episode involving the mass cancellation of Claude Code licenses within Microsoft itself is just the tip of the iceberg of a problem that companies like Uber, Meta, and Amazon are already feeling firsthand. 🤯
And what makes all of this even more interesting is that the cost of using AI is, in some cases, exceeding the cost of simply hiring humans to do the same work.
This is not an exaggeration, and it is not alarmism. It is what the numbers and statements from industry executives themselves are pointing to — including a surprising statement from inside Nvidia, the company that benefits most from the expansion of artificial intelligence worldwide.
So, what is really happening with the economics of AI? Let me break it down. 👇
Microsoft canceled AI licenses it had encouraged in the first place
It sounds contradictory, but that is exactly what happened. Microsoft — one of the largest investors in artificial intelligence on the planet, strategic partner of OpenAI, and developer of Copilot — cut most direct licenses for Claude Code, the Anthropic AI tool designed for programming. Instead, the company is directing engineers to use GitHub Copilot CLI.
According to a report by The Verge, the decision came just six months after the company opened access to Claude Code, encouraging thousands of developers, project managers, designers, and other employees to experiment with the tool for coding. Adoption was fast — maybe too fast. Usage scaled to the point where the company was forced to pull back on a tool its own engineers had already built into their daily workflows.
An important detail: the cancellation of Claude Code licenses does not affect Microsoft’s commercial agreement with Anthropic through Foundry. That deal includes an investment of up to $5 billion in Anthropic, access for Foundry customers to Claude models, and a commitment from Anthropic to purchase $30 billion in Azure compute capacity, also according to the report by The Verge.
The irony here is that Microsoft already has its own AI development tools, like GitHub Copilot. Yet internal teams were seeking external alternatives, which by itself says a lot about how the market is still trying to figure out which tool actually delivers results — and which one just generates a bill at the end of the month.
Uber burned through its entire annual AI budget in four months
Microsoft is not alone in this. The Uber case might be even more revealing about the scale of the problem.
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.
According to a report by The Information, Uber’s CTO, Praveen Neppalli Naga, revealed in April that the company had already burned through all of its 2026 budget for AI coding tools in just four months. And here is the kicker: the company itself had actively encouraged adoption, creating internal leaderboards that ranked teams by how much they used AI tools.
In other words, leadership pushed for adoption, teams embraced it enthusiastically, and the financial result was a monumental budget blowout before the first half of the year was even over. This raises an inevitable question: when the AI adoption strategy is measured by volume of use rather than by concrete returns, this is exactly what happens — runaway spending with no clear payoff.
Meta, Amazon, and the race to consume more tokens
The pattern repeats across other tech giants. At Meta, an employee created a dashboard nicknamed Claudeonomics — a reference to Anthropic’s Claude model — to track which workers were using the most AI. It was basically a scoreboard of who was consuming the most computational resources.
Over at Amazon, the term tokenmaxxing emerged internally, encouraging employees to use as many AI tokens as possible — these are the basic units of computational processing for language models. Every question asked, every line of code generated, every response reviewed consumes tokens. And tokens cost money.
These initiatives reveal a corporate culture where the primary goal was to maximize adoption, almost as if the mere act of using AI was synonymous with innovation. The problem is that in a consumption-based pricing model, more usage automatically means more spending. And without clear return metrics, the investment becomes pure cost. 💸
What is making AI so expensive?
To understand the scale of the problem, it helps to look at how AI pricing actually works in practice. Most tools charge by usage — whether by token processed, by API call, by compute hour, or by subscription per active user. On paper, the per-unit costs seem reasonable. The problem is that usage volume in corporate environments grows very quickly, and costs scale right along with it, often without managers noticing until the invoice arrives.
It is a dynamic similar to consumption-based streaming services: it seems cheap until you see how much you actually used at the end of the month.
The situation gets even more complex when we look at the future of AI agents. Goldman Sachs recently published a projection indicating that agentic AI — the kind that operates autonomously, executing tasks without human intervention — could drive a 24x increase in token consumption by 2030, as businesses and consumers adopt AI agents at scale. The estimate references a staggering 120 quadrillion tokens per month.
There is also another factor that significantly complicates the picture: the infrastructure required to run cutting-edge AI models is extraordinarily expensive. Nvidia GPUs can cost tens of thousands of dollars per unit, and the data centers powering these models consume energy at volumes that would make any budget manager flinch. All of this gets baked into the prices that AI companies charge end customers, creating a cost chain that is only likely to grow as long as demand stays hot.
The token paradox: cheaper individually, more expensive overall
Here is perhaps the most counterintuitive point in this entire story. The per-unit cost of AI tokens is, in fact, dropping — and it should continue to drop significantly over the coming years.
A recent report from consulting firm Gartner estimated that by 2030, the cost of inference on a one-trillion-parameter language model — essentially an extremely sophisticated AI model — will be nearly 90% lower than it was in 2025.
But Gartner was also careful to warn: cheaper tokens do not mean cheaper enterprise AI. There are three main reasons for this:
- Agentic models consume far more tokens per task than traditional chat or text generation models
- The increase in total consumption can outpace the drop in per-unit cost, pushing the final bill higher even with cheaper tokens
- AI providers do not necessarily pass cost reductions on to customers — just as happens in many other industries
Will Sommer, senior analyst at Gartner, was blunt in a public statement: product leaders should not confuse the deflation of commoditized tokens with the democratization of frontier reasoning.
In simpler terms: just because processing text has become cheaper does not mean that using advanced AI for complex tasks has become affordable. These are completely different things. 📊
The statement nobody expected to hear from Nvidia
If it was already surprising to see Microsoft cutting AI spending, what came from Nvidia was even more unexpected. Bryan Catanzaro, vice president of applied deep learning at the company, publicly stated in an interview with Axios a line that echoed across the entire industry:
For my team, the cost of compute is well beyond the costs of the employees.
The company that manufactures the chips that literally power the artificial intelligence revolution — and that has seen its stock price soar in recent years thanks to this wave — had one of its executives publicly acknowledging that, in certain contexts, keeping humans on the payroll is still cheaper than running AI for the same work.
This admission does not mean Nvidia is betting against artificial intelligence, far from it. What it does reveal, however, is that even the biggest beneficiaries of this wave are recognizing that the current economic model still has serious gaps. The efficiency that AI promises still depends heavily on the type of task, the volume of usage, the quality of available data, and the level of automation a company can implement around the tools.
Without those factors aligned, the return on investment can take much longer to materialize than any sales pitch suggests. 🤔
The dream of 100 AI agents per employee could come with a hefty price tag
To put the scale of the bet being placed into perspective, it is worth recalling a recent statement from Jensen Huang, CEO of Nvidia. He said he believes that, in the future, 100 AI agents will work alongside every employee at a company.
Huang is part of a broader wave of CEOs promoting an agentic future, where digital workers operate autonomously across the enterprise. The vision is ambitious and, in many ways, exciting.
But if token consumption grows faster than per-unit costs fall, that future could come with a much heavier bill than executives are forecasting. The simple math is relentless: multiply the consumption by 100 agents per employee, multiply by the number of employees, multiply by the tokens required per task — and the numbers become astronomical pretty quickly, even with individually cheaper tokens.
Promised productivity versus actual productivity
Here is the heart of the debate. AI clearly improves productivity in many scenarios — that is undeniable and documented in numerous studies. Developers using tools like GitHub Copilot report writing code faster. Writers using AI assistants can produce more in less time. Customer support teams that adopt intelligent chatbots can resolve more tickets per hour. These gains are real and measurable in specific, well-structured situations.
The problem lies in the gap between the actual gain and the expected gain — especially when that gain needs to be large enough to justify the rising costs of AI infrastructure. In many cases, companies underestimated the effort required to integrate the tools into existing workflows, train teams, fine-tune models for the specific business context, and deal with the errors and hallucinations that models still produce with uncomfortable frequency.
These factors consume time, money, and energy — resources that were rarely included in the original AI project budget.
There is also a human component that tends to be underestimated in projections: resistance to change. Even when tools work well technically, actual adoption within teams is usually lower than expected. Some professionals use AI only superficially, others ignore it entirely — which means the company pays for everyone’s license but only reaps the benefit from a few. This gap between activated licenses and effective usage is one of the biggest sources of invisible cost in the corporate tech ecosystem today.
What this changes going forward
The discussion emerging now is not about abandoning AI — no serious company is considering that. What is changing is how organizations are thinking about adopting these tools. The phase of excited experimentation is giving way to a more deliberate approach, where every license needs to be justified, every use case needs to be carefully evaluated, and productivity gains need to be genuinely measurable, not just estimated based on generic benchmarks.
These reports and episodes involving Microsoft and Uber could throw cold water on the bets that the largest tech companies have placed on AI. While some still cling to the promise of a renaissance or revolution through artificial intelligence, the real cost of adoption is proving to be a persistent bottleneck. These developments also suggest that the economics of replacing or augmenting human work with AI may be far more complicated than initial projections implied.
For tech companies that develop and sell these tools, the message is also clear: pricing and business models need to evolve. A model that generates unpredictable and hard-to-control invoices will not survive the scrutiny of chief financial officers at major corporations for much longer. The trend is toward more predictable models, with usage caps, fixed packages, and greater transparency about what is being consumed — much like what happened in the cloud computing market a few years ago, when companies realized they needed more control over what they were spending on the cloud.
The Microsoft episode with Claude Code might seem like an isolated detail, but it is actually an important signal that the AI market is maturing — and that this maturity will demand honesty about what works, what does not work, and how much all of this really costs. The promise of artificial intelligence remains valid. What is being revised now is the path to get there, and this adjustment process is natural, necessary, and ultimately healthy for the entire tech ecosystem. 🚀
