The two stories that broke quietly in May 2026 tell you more about the actual state of enterprise AI than any earnings call ever will.
Microsoft the company that poured roughly $13 billion into OpenAI and generates up to 30% of its own code using generative AI told engineers in a major division to stop using an AI coding tool because the bills got out of hand. At almost the exact same time, Uber burned through its entire $3.4 billion 2026 AI budget in four months, with per-engineer API costs hitting between $500 and $2,000 per month.
So what’s actually going on? The short version: AI works. It works so well that companies can’t afford to let their engineers use it freely.
Why Microsoft Told Its Engineers to Stop Using AI And What That Actually Means
Let’s be clear about what happened and what didn’t.
The affected team was Microsoft’s Experiences and Devices division the people who build Windows, Microsoft 365, Outlook, Teams, and Surface products. Claude Code had been launched internally there in December 2025, with a cancellation deadline set for June 30, 2026. Engineers were redirected to GitHub Copilot CLI instead.
This wasn’t a company-wide ban. This wasn’t Satya Nadella walking into an all-hands meeting and saying AI is a fraud. It was one division hitting a cost ceiling and getting redirected to a different presumably cheaper tool that Microsoft already owns a stake in.
That distinction matters more than most coverage is giving it credit for.
Here’s the thing people keep missing: Microsoft isn’t anti-AI. Microsoft is AI at this point. The real problem is that Claude Code, Anthropic’s agentic coding tool, apparently cost more than the budget for that division could absorb. Switching to GitHub Copilot keeps engineers in an AI-assisted workflow while bringing the cost back under internal control. It’s a vendor swap, not a philosophical pivot.
What this actually signals is something the industry has been dancing around for a year: enterprise AI tool economics are genuinely broken right now. The token costs, the API fees, the per-seat pricing none of it was designed for engineers who use these tools 8 hours a day at the level of depth they actually use them.
I’ve talked to engineering leads at companies a fraction of Microsoft’s size who ran into the same wall. They give their teams access, usage skyrockets because the tools are genuinely good, and then the finance team sends an email nobody expected. The ROI conversation suddenly gets very uncomfortable very fast.
The Uber Story Is More Revealing
Uber deployed Claude Code to around 5,000 engineers in December 2025. By April 2026, its CTO was telling reporters the company had burned through its entire annual AI coding budget in four months. Individual engineers were spending between $500 and $2,000 per month on API costs alone. 95% of engineers were using AI tools monthly. 70% of code commits were AI-driven.
Uber CTO Praveen Neppalli Naga told The Information: “I’m back to the drawing board, because the budget I thought I would need is blown away already.”
What makes Uber’s situation fascinating is the irony buried in it. This happened because the tool worked. Engineers found the AI genuinely useful and made it part of their daily workflow. You don’t get 95% monthly usage by forcing adoption. That’s organic. That’s engineers actually preferring to work this way.
By March 2026, about 84% of Uber’s engineers had adopted Claude Code. Around 70% of code committed at Uber now originates with AI. Additionally, 11% of live back-end updates are shipped by an agent with no human in the loop.
That last number deserves a pause. 11% of live backend changes going out without a human reviewing the final push. That’s not just “AI assists coding” — that’s agents operating autonomously inside production infrastructure at one of the world’s most used consumer platforms.
And the budget still blew up. Not because the tool underdelivered. Because it over-delivered in usage volume at a price point nobody had modeled correctly.
Uber’s response was not to ban AI coding tools but to impose a $1,500 monthly cap per employee per tool, tracked through an internal dashboard, with an approval process for exceptions. That’s the mature read. They’re not walking away from AI. They’re installing governance.
The Nvidia Admission Nobody Is Talking About
Here’s the data point that puts everything in context.
Bryan Catanzaro, VP of Applied Deep Learning Research at Nvidia the company now valued at over $5 trillion and making the chips that power most of the AI industry told Axios: “for my team, the cost of compute is far beyond the costs of the employees.”
Read that again. The company building the physical infrastructure for AI is saying compute costs more than people do. That’s not a warning sign buried in a footnote that’s the VP of Applied Deep Learning saying the quiet part loud. If you want to understand why Microsoft’s engineers got cut off and why Uber’s CTO is redoing his entire budget model, that sentence is your answer.
This is the cost structure the whole industry is sitting on. Companies like OpenAI, Anthropic, and Google are pricing their APIs against a cost of compute that’s still astronomical and they’re passing that cost downstream to enterprise customers who built their AI budgets before they had real usage data to model against.
You can read more about how AI compute costs are beginning to exceed workforce costs at some organizations in this deeper breakdown.
What This Tells You About Enterprise AI Adoption Right Now
There’s a pattern forming across the industry that the headlines are only half-catching.
Companies are not backing away from AI. They’re backing away from uncontrolled AI spend. Those are completely different things, and confusing them leads to very wrong conclusions about where this is heading.
GitHub is responding to the pricing pressure by shifting all Copilot plans to usage-based billing through GitHub AI Credits starting June 1, 2026. AI software prices across the US have climbed 20-37%.
So the tools are getting more expensive at the same time companies are discovering their usage was higher than expected. That’s a squeeze from both ends. The companies absorbing it are the ones that didn’t think hard enough about governance before deployment.
The ones doing it right and I’ve seen this play out across the companies I’ve followed closely — are treating AI tools the way they treat cloud infrastructure. You don’t give every engineer uncapped AWS access. You set budgets, you track usage, you create approval flows for anything above a threshold. The fact that it took Uber and Microsoft burning through real money to arrive at that obvious conclusion is a little embarrassing for an industry that was supposed to be smart about this.
The honest truth? Most enterprise AI rollouts in 2024 and 2025 were done by executives trying to show the board they were “AI-first” before they had any real framework for what that meant operationally. The chickens are coming home to roost in 2026.
The ROI Question Nobody Wants to Answer Honestly
Here’s what every article covering these two stories conveniently sidesteps: did the AI actually make the engineers more productive in a way that justified the cost?
In Uber’s case, the usage numbers suggest yes but there’s no public data on whether velocity, quality, or output improved proportionally. Around 70% of code committed at Uber now originates with AI, and 11% of live back-end updates are shipped by an agent with no human in the loop. That’s a seismic operational shift. But code volume isn’t the same as business value. Shipping more code faster doesn’t help if the code introduces more bugs, requires more maintenance, or creates technical debt that costs more to unwind than the speed gain was worth.
The ROI gap in AI tooling is real. Most companies deploying tools like Claude Code or Cursor at scale are measuring inputs (usage rates, code commits, time saved per task) but not outputs (revenue per engineer, defect rates, customer-facing reliability). Until those output metrics improve, “we used AI” and “AI worked” are two very different statements.
This matters because it shapes what comes next. If Uber’s engineering velocity meaningfully improved — if features shipped faster, if fewer bugs reached production, if the 11% autonomous deployments performed as well as human-reviewed ones — then the $3.4 billion AI budget isn’t a failure story, it’s a pricing calibration problem. They need to negotiate better rates or find a more efficient tool stack.
If the velocity improvement didn’t justify the spend, that’s a different conversation entirely. And nobody from Uber or Microsoft has published those numbers yet.
The question of who owns AI outputs especially when 11% of production deployments are agent-driven is also getting more complicated by the month. That ownership question is one the industry is still actively untangling.
The Real Danger Isn’t Overspend It’s Mispriced Risk
What actually keeps me up about stories like this isn’t the budget blowout. It’s the operational risk hiding inside these adoption numbers.
11% of live back-end updates at Uber are shipped by an agent with no human in the loop. At a company processing millions of rides and food deliveries daily, that is not a trivial exposure. One bad autonomous commit in a billing system, a routing algorithm, or a driver payout calculation doesn’t stay contained.
The savings on developer time can disappear fast if a single agent-deployed change requires an emergency incident response that pulls 20 engineers for a week. That math almost never makes it into the ROI spreadsheet.
I’m not saying agent-driven deployments are inherently risky. The evidence is actually that these tools perform well most of the time that’s why adoption reached 84-95% organically. But “most of the time” is not the same as “always,” and in production infrastructure, the tail risk on the exceptions is enormous.
The companies getting this right are the ones building hard circuit breakers into agent workflows. They let the AI write, review, and even test — but the deployment to production still requires a human sign-off above a certain risk threshold. That’s not AI-phobic. That’s just engineering with appropriate safeguards.
What’s Actually Happening With AI Company Valuations Right Now
There’s a broader financial picture worth understanding here.
The same month Microsoft pulled back on Claude Code spending, Anthropic surpassed OpenAI as the most valuable private AI company which tells you something interesting about how investors are reading the enterprise adoption story. They’re not seeing Microsoft’s budget decision as a referendum on Anthropic’s product. They’re seeing it as a sign that AI tool usage is deep enough that cost management is now the dominant concern, not adoption.
That’s actually a more mature market signal than “AI is going to replace everyone.” It means we’ve moved past the “will companies use AI” question and landed squarely on “how do companies afford to use AI at the scale they want to.” OpenAI’s own path to a public offering reportedly targeting September 2026 will put a hard number on what the market thinks enterprise AI economics are actually worth at scale.
What Should You Actually Do With This Information
If you’re an engineering leader or a company thinking about deploying AI coding tools at scale, the Microsoft and Uber stories are your free case study. Here’s what the data actually tells you to do:
Set usage caps before you see the bill, not after. Uber’s $1,500 monthly cap per employee per tool, with a separate allowance per tool and an approval process for exceptions, is the right structure. Build that before rollout, not as a crisis response.
Track output metrics, not just usage metrics. Usage rate is vanity. Ship rate is sanity. Defect rate post-AI-assist is what actually tells you whether this is working. Most teams are measuring the wrong thing.
Treat agentic deployments differently from assisted coding. An engineer using AI to write a function is low-risk. An agent pushing to production with no human review is a different risk tier entirely. Build different governance for each.
Don’t assume your usage pattern matches the pilot. Pilots run hot because early adopters self-select. Full rollouts go wider and deeper. Model your budget on 80% adoption at 150% of pilot usage volume — that’s closer to what actually happens.
Pick tools based on cost per output, not cost per seat. Per-seat pricing looks manageable in a spreadsheet. Token-based pricing at real usage volumes is what breaks budgets. Do the math on actual usage, not theoretical usage, before you sign.
The bottom line is this: AI coding tools are real, they work, and engineers genuinely want them. The problem isn’t the tools. It’s that enterprise procurement teams are buying the 2023 pitch deck at 2026 usage volumes. Close that gap and you avoid the headline.
The question isn’t whether to deploy AI tools. It’s whether you’ve done the cost modeling to deploy them without a mid-year budget crisis. Most companies haven’t. Now they have two very public examples of what that looks like.
