Microsoft backs away from Claude Code: too popular, too expensive

When Microsoft opened access to Anthropic's Claude Code to its thousands of developers, designers, and product managers six months ago, it expected a productivity boost. Instead, it got billing shock. According to The Verge, the company is now canceling most direct licenses and redirecting engineers to GitHub Copilot CLI — its own tool that it keeps under control.

Claude Code gained popularity among Microsoft employees surprisingly fast. Too fast. Massively scaling its usage exposed a crack in AI economics: the more employees use AI, the higher the bill — and it can spiral out of control. The license cancellations do not affect Microsoft's strategic partnership with Anthropic (Foundry, $5 billion investment), but exclusively the internal deployment among its own employees.

Nvidia: "AI costs are higher than people costs"

An even sharper statement came from the company that profits most from the AI boom. Bryan Catanzaro, VP of Applied Deep Learning at Nvidia, said in an interview with Axios: "For my team, the cost of compute is far beyond the cost of employees." In other words — even at Nvidia, which manufactures AI chips, running AI costs more than paying people who would do the same work themselves.

This admission resonates with a 2024 MIT CSAIL study that analyzed the economic viability of replacing human labor with AI. Researchers found that AI automation would be economically viable for only 23% of job roles where vision is the main component of the work. In the remaining 77% of cases, it was cheaper to leave the work to humans.

Tokenmaxxing: when companies push AI too far

Across Silicon Valley in 2026, a phenomenon Fortune calls "tokenmaxxing" has spread. Companies actively incentivize employees to use as many AI tokens (the basic units of AI computation) as possible, even creating internal leaderboards:

• Meta launched a dashboard called "Claudeonomics" that tracks which employees use AI the most
• Amazon directly calls on its employees to embrace tokenmaxxing — using the maximum amount of AI tokens
• Uber motivated teams through internal leaderboards, and CTO Praveen Neppalli Naga admitted in April that the company burned through its entire 2026 AI coding tools budget in just four months

The result? Instead of savings, sticker shock. Agentic AI — autonomous systems that independently perform multi-step tasks — consumes an estimated up to 1,000× more tokens than regular chatbots. Every "smart" decision an agent makes means another and another compute cycle — and another line item on the invoice.

The cheap token paradox: per-unit price drops, the bill goes up

At first glance, it doesn't make sense. The price of individual AI tokens is dropping dramatically. According to Gartner, inference costs (query processing) for a model with a trillion parameters will drop by more than 90% by 2030 compared to 2025. Yet at the same time, Goldman Sachs predicts that agentic AI will cause a 24× increase in token consumption by 2030 — to an incredible 120 quadrillion tokens per month.

In other words: tokens get cheaper, but we consume orders of magnitude more of them. Gartner analyst Will Sommer adds: "Product leaders should not confuse commodity token deflation with the democratization of advanced reasoning." Even with declining unit prices, total corporate AI bills will rise.

Other giants have already felt this paradox. Keith Lee, AI and finance professor at the Swiss Institute of Artificial Intelligence, described the current situation to Fortune as a "short-term mismatch." Fixed subscriptions (typically $20–200 per month for developer AI tools), according to him, don't cover the actual operational costs for the most active users. The result will be an inevitable shift to a usage-based model (pay-per-use), which companies like OpenAI are already testing with GPT-5.5 at twice the price of its predecessor.

What it means for Czechia and Europe

While American companies wrestle with rising bills, European businesses are just starting to adopt AI tools. According to U.S. Federal Reserve data, only about 18% of American companies had adopted AI tools by the end of 2025 (a 68% year-over-year increase). In Europe, penetration is even lower, which may paradoxically be an advantage in this context — companies can learn from American giants' mistakes.

For Czech companies and developers, there are several takeaways:

• Don't blindly follow the trend of deploying AI on everything. It pays to measure concrete return on investment (ROI) — not just the number of tokens or hours spent with an AI tool.
• Usage-based models are fairer for smaller teams. Fixed licenses at $25–200 per month per developer look tempting, but if you use the tool intensively, the final bill can be many times higher. Czech companies should negotiate transparent pricing models.
• The EU AI Act adds additional compliance costs. European companies must account not only for token pricing but also for regulatory compliance costs — audits, documentation, AI decision-making transparency.

In Czechia, no major employer has yet publicly announced restricting AI tools for their developers due to costs. Companies like Ecomail, on the other hand, are actively integrating AI into their services. But as tools like Claude Code, Cursor, or GitHub Copilot become standard developer equipment, the question of "how much is this actually costing us" will become increasingly pressing for Czech CTOs as well.

The future: cheaper AI, or cheaper people?

Analysts' outlook offers mixed signals. Gartner predicts inference will be 90% cheaper by 2030. Yet at the same time, agentic models consume orders of magnitude more tokens, and experts say AI providers won't pass the full savings on to end customers — they'll keep part of them to cover their own infrastructure.

Jensen Huang, CEO of Nvidia, still believes that every employee will one day have 100 AI agents working for them. If this vision materializes, it will mean that the economic equation must change radically — otherwise companies would go bankrupt on token bills before they even managed to deploy the agents.

The reality of May 2026 is more sobering: AI is an amazing tool, but its operating costs do not yet justify mass replacement of human workers. Microsoft figured it out. Uber figured it out. And soon every company that succumbs to the "use AI for everything" pressure without calculating the actual costs will figure it out too.