Microsoft Cancelled Its Own Claude Code Licenses Because the Bill Got Too Big

Welcome back to SavvyMonk, your one-stop for AI and tech news that actually matters.

Today's story is one of those ones where you read the headline, assume it's exaggerated, and then realize it's actually worse than it sounds. Microsoft gave thousands of engineers access to Claude Code, watched them fall in love with it, and then cancelled the licenses because the bills got out of control. And Microsoft isn't the only one hitting this wall.

Let's get into it.

WeAreDevelopers World Congress comes to San José, CA — September 23–25, 2026. 10,000+ developers, 500+ speakers, and the full software development lifecycle under one roof, in the heart of Silicon Valley.

Kelsey Hightower. Thomas Dohmke (fmr. CEO, GitHub). Christine Yen (CEO, Honeycomb). Mathias Biilmann (CEO, Netlify). Olivier Pomel (CEO, Datadog). The people actually building the tools you use every day — all on one stage.

AI, cloud, DevOps, security, architecture, and everything real builders ship with. Workshops, masterclasses, and the official congress party.

Tickets from $599. Use code GITPUSH26 for 10% off.

Secure Your Pass

Back in December 2025, Microsoft gave its Experiences + Devices division access to Anthropic's Claude Code. This is the team that builds Windows, Microsoft 365, Outlook, Teams, and Surface, and the idea was straightforward. Engineers would try Claude Code alongside GitHub Copilot CLI, Microsoft's own AI coding assistant, and the company would see which one they preferred.

Claude Code won almost immediately, with engineers picking it over Microsoft's own tool, using it constantly, and basically ignoring Copilot CLI. By May, internal sources were describing Claude Code as very popular, perhaps a little too popular, which is a real problem when your entire business strategy depends on selling Copilot to every enterprise on the planet and your own engineers won't use it.

On May 14, Rajesh Jha, EVP of Experiences + Devices, announced that nearly all Claude Code licenses in his division would be cancelled by June 30, 2026, which is not a random date since it falls on the last day of Microsoft's fiscal year. Every engineer in the group now has to switch to Copilot CLI.

Jha framed this as a strategic move, saying Copilot CLI gave them something especially important in a product they could shape directly with GitHub, but the real story is simpler. Token-based pricing means every query, every code review, every debugging session costs money, and when thousands of engineers are running agentic workflows that read entire codebases, plan changes across dozens of files, run tests, and open pull requests on their own, the bills pile up in ways that nobody had budgeted for.

Keep in mind, this is the same Microsoft that invested up to $5 billion in Anthropic just seven months earlier. Claude's models will still be available through Copilot CLI and Microsoft Foundry, so the relationship isn't over. But the tool engineers actually loved using every day is gone.

The Nvidia Confession

Then came the quote that really put the whole thing in perspective.

Bryan Catanzaro, Nvidia's VP of applied deep learning, said in late April 2026 that the cost of compute for his team was far beyond the costs of the employees, and that admission landed differently because of who was saying it.

This is a VP at the company that sells the chips powering all of this, the company that has made more money from the AI boom than anyone else, and even he is acknowledging that running AI costs more than paying the humans who use it.

A 2024 MIT study found the same thing from a different angle. Researchers looked at what it would actually take for AI to match human performance across different roles and found that AI automation only made financial sense in 23% of vision-based jobs. In the other 77%, it was cheaper to keep paying people.

The Bigger Math Problem

For the past two years, every CEO on every earnings call has said some version of the same thing, that AI will make us more efficient, reduce headcount, and cut costs. Wall Street loved it, and every time a company announced layoffs the stock went up, and every time a company announced AI adoption the stock went up again.

But the companies actually deploying these tools at scale are finding the opposite, because the more employees use AI, the higher the bill gets.

Goldman Sachs forecasts a 24x increase in token consumption by 2030, reaching 120 quadrillion tokens per month as companies adopt AI agents for more and more tasks, and Gartner published a report in March that laid out the paradox at the heart of all of this.

Individual token prices will drop more than 90% by 2030, but total enterprise AI costs will still go up because agentic models consume 5 to 30 times more tokens per task than basic chatbots, which means usage is growing so much faster than prices are falling that the overall bill keeps climbing anyway.

And the culture inside Big Tech is making this worse. A Meta employee built an internal dashboard called Claudeonomics that tracked which of the company's 85,000 employees used the most AI tokens, ranked the top 250 users, and handed out badges like Token Legend and Cache Wizard. In one 30-day window the dashboard logged 60 trillion tokens before Meta shut it down after the data leaked externally.

Over at Amazon, employees started what Silicon Valley now calls tokenmaxxing, where after the company set targets requiring over 80% of developers to use AI tools weekly, some engineers began running AI agents on trivial tasks just to inflate their usage scores. Multiple employees said there was so much pressure to use these tools that people were automating unnecessary work to climb internal leaderboards.

Meanwhile, combined 2026 capital expenditure from Amazon, Microsoft, Alphabet, and Meta is somewhere between $650 billion and $700 billion, with Wall Street projections exceeding $1 trillion for 2027. And the first companies to actually deploy these tools across their engineering teams are already pulling back because the economics aren't working the way anyone expected.

The Bottom Line

There's a growing gap between what companies tell investors about AI and what they're discovering when they actually use it at scale. Token-based pricing turns every productivity gain into a metered cost, and the more useful the tool is, the more expensive it gets, which means the AI cost curve is not heading where the earnings calls promised.

The companies that know this best are the ones building the technology themselves.

Category: Image Generation

"Create a 4:5 ultra realistic cinematic studio portrait using uploaded reference photos, replace original subject while preserving exact facial identity, proportions, hairstyle, beard density, skin tone, smile structure, natural imperfections, maximum accuracy. Vertical portrait composition, medium-full body framing, subject standing casually with relaxed posture. Both hands naturally inside pockets, body slightly turned, head tilted upward while laughing genuinely. Candid confident expression, natural energetic vibe, preserve exact pose, body angle, framing, perspective, composition. Shot on 85mm prime lens, shallow depth, cinematic studio photography, ultra sharp face focus, soft background separation, realistic depth, DSLR quality. Dramatic cinematic lighting, deep shadow contrast, strong warm key light from upper left, subtle rim light around face and shoulders, smooth shadow falloff, vibrant blue seamless studio background, moody editorial grading, realistic skin reflections, shadow depth. Oversized matte black t-shirt, modern minimal streetwear aesthetic, natural eyeglasses realistically integrated, premium casual creator look, subtle watch visible if fitting. Do not modify or stylize face, preserve exact identity from references, maintain real facial proportions, skin texture, smile shape, laugh expression. No beautification, exaggerated jawline, hairstyle change, beard pattern changes, artificial smoothing. Preserve natural asymmetry, realistic human details. Hyper realistic, cinematic portrait photography, realistic anatomy, ultra detailed skin texture, soft realistic shadows, premium Instagram creator aesthetic, editorial fashion feel, HDR lighting, 8k render, natural realistic color science."

The companies selling the AI future and the companies living the AI future are telling two very different stories right now. One group is on earnings calls promising efficiency and cost savings, while the other is scrambling to figure out why the bill is higher than the humans the tools were supposed to replace. At some point those two stories have to meet in the middle, and when they do, a lot of assumptions are going to break.

Hit reply, I read every response.

See you in the next one.

— Vivek

P.S. If you know a founder, CTO, or finance lead trying to budget for AI tools, forward this to them. They can subscribe at https://savvymonk.beehiiv.com/