For the better part of two years, the story in tech was simple: AI makes your engineers faster, happier, and more productive. Companies handed out access to AI coding tools the way offices once handed out free snacks. Generous, enthusiastic, and completely without a plan for what the bill would look like. That bill has now arrived, and nobody is particularly thrilled about the number on it.
Featured Image: Igor Omilaev (UnSplash)
What started as a few raised eyebrows at expense reports has turned into a full-blown industry reckoning. Microsoft, Uber, and GitHub are the names making headlines right now, but they represent a much wider pattern: the gap between what companies expected AI to cost and what it actually costs at scale is proving to be wider than anticipated.
Microsoft Pulls the Plug on Its Own Experiment
In December 2025, Microsoft gave thousands of its internal employees access to Claude Code, Anthropic’s AI-powered coding tool. The rollout covered the Experiences and Devices division, the team responsible for building Windows, Microsoft 365, Outlook, Teams, and Surface. Access went beyond developers to include designers, product managers, and non-technical staff, as part of a push to let people without coding backgrounds build tools using AI.
It worked, perhaps too well. According to The Verge’s Tom Warren, Claude Code became “very popular, perhaps a little too popular” inside Microsoft over the following six months. Engineers preferred it over GitHub Copilot, Microsoft’s own product. Which is, to put it diplomatically, an awkward outcome when you are Microsoft.
By May 14, 2026, Microsoft began cancelling those Claude Code licenses, with a full transition deadline of June 30 that conveniently lines up with the end of Microsoft’s fiscal year. Engineers are now being directed to GitHub Copilot CLI instead. Rajesh Jha, Executive Vice President of Microsoft’s Experiences and Devices group, framed the decision in an internal memo: “When we began offering both Copilot CLI and Claude Code, our goal was to learn quickly, benchmark the tools in real engineering workflows, and understand what best supported our teams. Claude Code was an important part of that learning… Copilot CLI has given us something especially important: a product we can help shape directly with GitHub.”
The decision is partly about cost, partly about the optics of your own workforce preferring a rival tool, and partly about platform consolidation. On the Q3 earnings call, CEO Satya Nadella confirmed the broader shift: “Any per-user business of ours, whether it’s productivity, coding, security, will become a per-user and usage business.” The direction of travel is clear.
It’s worth emphasizing that this doesn’t mean Microsoft has walked away from Anthropic. The two companies have a substantial partnership in place, with Microsoft committing to invest up to $5 billion in Anthropic and Anthropic agreeing to purchase $30 billion in Azure compute capacity. Claude models remain available through Microsoft Foundry and within Microsoft 365 Copilot for specific tasks. The license cancellation is exactly what it looks like, a cost management move.
Uber Spent Its Entire Year’s AI Budget in Four Months
Uber’s situation is more straightforward, and more alarming. The ride-hailing company also rolled out Claude Code to its engineering team in December 2025. By March 2026, 84 percent of Uber’s roughly 5,000 engineers were classified as active agentic coding users. By April, roughly 95 percent of engineers were using AI tools every month, and around 70 percent of all committed code was coming from those tools.
The problem is that agentic AI tools don’t work like traditional software subscriptions. When an engineer uses Claude Code for a multi-step autonomous coding session, the tool reads through entire codebases, writes tests, runs them, catches failures, and iterates through multiple passes. Each of those steps consumes tokens. Token costs add up. Monthly API costs per engineer at Uber ranged from $500 to $2,000, depending on how heavily they used the tool.
Uber CTO Praveen Neppalli Naga told The Information that the company had burned through its entire planned 2026 AI budget by April. Four months into the year. “I’m back to the drawing board,” Naga said, “because the budget I thought I would need is blown away already.” The finance team’s original model had been built around fixed seat counts and low-frequency usage. It completely failed to anticipate what happens when thousands of engineers adopt an agentic tool as their primary workflow.
Uber’s COO Andrew Macdonald later described a “head-exploding moment” when Naga’s comments went viral, sparking internal conversations about whether higher token usage was actually translating into a proportional increase in useful consumer features. For now, Macdonald said it was becoming harder to justify the costs. Uber is exploring OpenAI’s Codex alongside Claude Code as it searches for a cost structure that works at its scale. The internal leaderboards that once ranked engineers by AI tool usage in a trend known as “tokenmaxxing”, are likely being viewed through a slightly different lens.
GitHub Rewrites the Pricing Contract
GitHub, which is owned by Microsoft, announced in April 2026 that all Copilot plans will move to usage-based billing starting June 1, 2026. Under the old model, a quick chat question and a multi-hour autonomous coding session cost the user the same amount. That pricing structure made sense when Copilot was primarily an inline code suggestion tool. It stopped making sense when Copilot became an agentic platform running long, multi-step coding sessions across entire repositories.
GitHub’s Chief Product Officer Mario Rodriguez didn’t dress it up: “GitHub has absorbed much of the escalating inference cost behind that usage, but the current premium request model is no longer sustainable.” The new system replaces premium request units with GitHub AI Credits, where one credit equals one cent and usage is calculated based on token consumption. Existing monthly plans migrate automatically; annual subscribers will face price increases as their plans expire.
The broader message is that the AI tool industry is moving toward a model where your bill reflects what you actually used, not what you paid for up front. For individual developers, that means more transparency. For enterprise IT and finance departments managing thousands of seats, it means unpredictable monthly invoices and a new cost category they’ll need to track, forecast, and explain to the board.
The Wider Picture: An Industry-Wide Reckoning
Microsoft and Uber are the loudest examples right now, but they’re far from alone. AI software prices across the industry have reportedly climbed between 20 and 37 percent as the real economics of running agentic tools at scale become clearer. The original pitch for AI coding tools was that they’d pay for themselves through productivity gains. That math is being stress-tested by actual usage data.
The underlying problem is structural. Traditional enterprise software budgets are built around seats and licences: you pay a fixed amount per user per month, and finance can model that out without much drama. Token-based billing introduces a variable that existing procurement and FinOps teams just aren’t set up to manage. Analysis from EPC Group found that 81 percent of enterprise leaders are worried about AI vendor dependency, yet only 6 percent say they could switch providers without significant disruption. That combination of dependence and unpreparedness is not a comfortable position to be in.
For smaller companies and startups, the implications are more immediate. If a company the size of Microsoft can have its engineering budget blindsided by AI usage costs, a startup with fifteen engineers and no dedicated FinOps function faces the same problem at a scale that could genuinely be existential. The tools are powerful. They are also, at high adoption rates, expensive in ways that flat-rate pricing never revealed.
What This Means for Employment
The employment picture here is more complicated than it first appears, and it cuts in two directions.
One argument says AI is proving too expensive to actually replace human workers at scale. If running AI agents costs $500 to $2,000 per engineer per month, on top of that engineer’s salary, the tool is an amplifier, not a substitute. At some organizations, compute costs now exceed the salaries of the human teams the tools are supposedly replacing. This is AI making skilled workers more central, not less so, because someone still has to govern, direct, and clean up after the agents.
The counterargument is that cost pressure tends to find its outlet somewhere, and labor is often where it lands. Marell Evans, founder of Exceptional Capital, predicted in late 2025 that as AI budgets grow, human labor budgets will be cut to compensate. Rajeev Dham of Sapphire Ventures agreed that 2026 would mark real shifts in how companies allocate resources between people and tools. Jason Mendel of Battery Ventures described 2026 as the year AI would move from augmenting workers to automating work itself in certain categories.
What’s happening right now looks less like replacement and more like compression. Fewer engineers may be needed to maintain or ship the same volume of work as AI tooling matures. Entry-level and mid-level roles that once involved writing significant amounts of routine code are already seeing fewer postings. The demand for engineers who can direct, review, and architect AI-generated output is growing. The demand for engineers whose main job is writing that output is contracting.
There’s also the question of what Microsoft’s internal mandate means in practice. When a company tells its engineers to stop using the tool they prefer and switch to a different one, productivity takes a short-term hit. Engineers who built workflows around Claude Code now need to rebuild them around Copilot CLI. That kind of transition is rarely frictionless, and the people closest to it tend to have opinions about it that don’t always make it into the press release.
What the Numbers Actually Show
None of this means AI is failing. The Uber story, read carefully, is a story about a tool that worked so well engineers couldn’t stop using it. The problem was that nobody modeled what widespread adoption of a genuinely useful product would actually cost.
The AI industry is going through a maturation that every major technology has gone through: the gap between early adopter enthusiasm and sustainable business deployment. Cloud computing went through it. Mobile went through it. AI is going through it now, just faster and louder because the sums of money involved are large and the claims made on the way up were ambitious.
Companies aren’t abandoning AI. They’re getting more serious about it. There’s a difference between those two things, even if the budget memos look similar from the outside. The next phase of enterprise AI will look less like unlimited access for everyone and more like deliberate, governed deployment with actual cost controls. That’s probably a more durable model, even if it’s a harder one to sell internally.
For engineers watching all of this, the clearest takeaway is that the tools aren’t going away, but neither is the expectation that you know how to use them well. The premium is shifting toward people who can evaluate AI output critically and understand where it adds genuine value, rather than just reach for it by default. That’s a different skill than writing fast code, and it’s the one most relevant right now.
Leave a Reply