May 14, 2026, Microsoft has begun revoking internal licenses for Claude Code for most employees. The deadline is June 30th—also the last day of Microsoft's fiscal year.
Just six months ago, Microsoft was doing the exact opposite. In December 2025, it opened up Claude Code to thousands of employees, including engineers, product managers, and designers, encouraging everyone to reshape their workflows using vibe coding. Employees loved the tool. Perhaps, too much.
But six months later, Microsoft pulled back on its own.
Almost in the same week, YC partner Tom Blomfield said something else during a batch talk: "If your API bill doesn't make your heart ache, you're not burning enough."
In the same spring, Silicon Valley is giving two completely opposite answers to the same question—is using AI ultimately more expensive than humans?
01 The Scene of Vibe Coding's Failure
Microsoft isn't canceling the Claude model itself. Anthropic's models will continue to be available to Microsoft employees through Copilot CLI. What it's canceling is the product entry point for Claude Code itself.
The department most affected is "Experiences + Devices"—the engineering teams behind Windows, Microsoft 365, Outlook, Teams, and Surface. EVP Rajesh Jha packaged the decision as "toolchain unification" in an internal memo, but Microsoft insiders cited by The Verge were more blunt: employees generally found Claude Code better to use than Copilot CLI. The popularity of Anthropic's tool within Microsoft even led to Microsoft's own Copilot CLI being "neglected."
In other words, Microsoft removed Claude Code not because it was ineffective, but because it was too effective.
That June 30th deadline is no coincidence—it's the last day of the fiscal year. Cutting off a tool employees widely preferred, switching back to its own product, timing it at the fiscal year's end—how much of this is product judgment versus financial consideration, everyone understands.
Microsoft is not an isolated case.
A month ago, Uber CTO Praveen Neppalli Naga revealed to The Information: the company's entire 2026 budget for AI coding tools was burned through in the first four months. Uber had previously even created internal leaderboards, using competitions to incentivize employees to use AI more—the result was a budget blowout.
More direct was what NVIDIA's VP of Applied Deep Learning, Bryan Catanzaro, said in an interview with Axios: "For my team, compute costs far exceed employee costs." This came from an executive at a hardware company—a company whose core product is selling compute power.
Fortune strung these clues together, giving its article a very Fortune-esque title: "Microsoft's report exposes AI's real cost problem—using this thing is more expensive than raising employees."
If you only read this far, the conclusion is simple: vibe coding failed, the story of AI replacing humans can be shelved.
But that conclusion is premature.
02 The Copilot Model Has "Hit a Wall"
To explain Microsoft's retreat, we first need to clarify what vibe coding is.
This term was proposed by Andrej Karpathy in early 2025—he described a new way of programming: developers no longer write code line by line, but use natural language to describe intent, letting the LLM generate code. Developers might not even read the code, just look at the results—accept it if it runs, have AI fix it if it doesn't.
This is one of the most enticing productivity promises of the AI era. It means: an engineer who doesn't know Rust can have AI write Rust for them; a product manager can have AI build a prototype; a designer can have AI write runnable code. The target audience for Microsoft's December 2025 opening of Claude Code—engineers, PMs, designers—happens to be exactly these three types. This is no coincidence; it's the classic landing posture for vibe coding.
But when vibe coding lands in a large company, it becomes something structurally awkward.
Suppose Microsoft has an engineer with an annual salary of $300,000. After equipping him with Claude Code, his output increases by 20%—this is the ideal state of vibe coding. But at the same time, the token cost he burns through each month is $200, $500, or $2,000? This number will monotonically increase as his dependence on AI deepens.
What's more troublesome is, he won't be laid off because he "uses AI"—his $300K salary remains, his benefits remain, his desk remains.
In other words, Microsoft's total cost structure becomes "original employee wages + new token bill." This formula only goes in one direction—costs skyrocket.
And does "employee output +20%" translate to "revenue +20%" financially? No. It translates to "revenue remains unchanged, but the cost structure now includes an AI bill"—because the output of most employees does not directly correspond to new revenue; him writing faster doesn't mean the company sells more.
This is the real meaning behind Catanzaro's statement "compute is more expensive than employees." It's not saying AI is dumb; it's saying that when you attach AI to existing employees, you can't make the math work.
This logic also has data support.
In a recent Gartner forecast: by 2030, the inference cost for a trillion-parameter large model will drop by nearly 90% compared to 2025. It sounds like AI is getting cheaper, but Gartner's real conclusion is: this won't make the total AI bill for enterprises cheaper. Gartner senior director analyst Will Sommer said something—"CPOs should not confuse 'deflation of commodity-level tokens' with 'the commoditization of cutting-edge inference capabilities.'"
Goldman Sachs's prediction is more direct: by 2030, agentic AI will drive a 24-fold increase in token consumption, reaching 120 quadrillion per month. A 90% drop in per-token price, a 24-fold increase in consumption—the result is the total bill still rises.
Jensen Huang has an even more radical version. He said publicly a few months ago that in the future, every NVIDIA employee will have 100 AI agents working alongside them.
It sounds beautiful. But if you're the CFO, what you hear? It's 100 token-burning furnaces, burning non-stop 24/7.
The problem isn't that AI is too expensive. The problem is the assumption itself of "giving each employee an AI copilot."
This posture has a popular name in tech circles—"copilot mode." Its core assumption is: the human remains in the driver's seat, AI sits in the passenger seat giving suggestions. It doesn't replace you, it just makes you faster.
This assumption is very gentle on the surface—"AI won't take your job, AI is just helping you." But financially, its implicit meaning is: all original wages remain unchanged, but an additional token fee is added.
And tokens aren't a fixed cost; they're billed based on consumption. The more employees use, the more the company pays—this happens to be the cost structure enterprises least want to see: variable, uncapped, inversely amplifying with productivity.
When Microsoft opened up Claude Code in December 2025, it might not have fully realized this. It likely thought: let employees try it, see how much AI can boost work efficiency. But six months later, employees really got addicted, Claude Code became too popular inside Microsoft—the result was a token bill far exceeding expectations, surpassing the extra output Microsoft could recoup from this popularity.
Microsoft pulled back. But it's not pulling back on AI—it's pulling back on the structure of "employees remain in the driver's seat, AI in the passenger seat."
This is a structural failure. It won't disappear because models get cheaper, nor will it disappear because employees get more skilled—it will become more severe as employees become more proficient with AI.
03 Burning Tokens Because It's Not Burning Headcount
Almost in the same week as Microsoft's retreat, Tom Blomfield presented a completely different perspective at a YC batch talk. He didn't discuss "how AI should be used"—he discussed "what a company in the AI era should look like."
Blomfield's judgment was direct: most companies today still have a "Roman legion" structure—information flows up the chain, commands flow down, people are the core of coordination. Attaching AI to this structure is like giving firearms to Roman infantry—they'll use them more intensely, but the tactics won't change.
A truly AI-native company should look different.
Blomfield used a very specific description: every action should produce a recordable, callable artifact, making everything legible to AI; the company should be designed as a "self-improving AI loop," a system that can sense the environment, make decisions, call tools, receive feedback, and self-correct.
Humans in such a company are left with only two roles. First, individual contributors—everyone, regardless of department, is a builder and operator, bringing prototypes to meetings, not just ideas. Second, DRI (Directly Responsible Individual)—every output has a clear responsible person, "can't hide behind AI."
Then Blomfield said that golden line: "If your API bill doesn't make your heart ache, you're not burning enough."
If this sentence appeared in Microsoft's CFO office, it would be considered a joke; but in front of a room full of startup founders at YC, no one thought it was crazy.
Why?
Another YC partner, Diana Hu, gave the answer in early May at Startup School. She said something—"Maximize not headcount, but token consumption." She had an even blunter version: "One person equipped with AI tools equals what used to be a large engineering team."
Note the key word here: "equals." Not "comparable to," not "similar to"—replaces.
In YC's P26 Spring 2026 batch, there are already several companies using 5 or 6 people to do what used to require 20 or 30 people. Their token bills are certainly high, but their personnel bills are extremely low—the overall calculation shows a profit.
A more radical case is Block. This fintech company under Jack Dorsey recently laid off 40% of its employees. This isn't "cost-cutting and efficiency improvement" in the traditional sense—Block simultaneously increased internal investment in AI tools. The new structure is exactly what Diana Hu described: IC + DRI + AI agent.
In YC's context, burning tokens isn't an expense; it's a replacement. It replaces not costs outside of AI, but headcount wages. The math works because the company simultaneously removed the positions that would have burned salary money.
This is the fundamental reason why Microsoft and YC see the same thing but give opposite answers—they aren't burning the same kind of token at all. Microsoft's token is refueling the copilot for the original crew, YC's token is replacing the original driver.
04 True Assets Are Being Redefined
In his talk, Tom Blomfield also said another more intriguing line—"People are transient; context documentation is what's important."
This is an accounting-level judgment.
How is a traditional company's balance sheet written? On the left: fixed assets, accounts receivable, goodwill, IP. On the right: liabilities and shareholders' equity. Employees aren't in the asset column—employees are a cost. But every company knows internally that employees are the real assets: customer relationships are in the salesperson's mind, business intuition in the product manager's mind, technical know-how in the engineer's mind.
The characteristic of this kind of "asset" is that it walks away. Employees leave, the asset runs away.
What Blomfield describes for an AI-native company is doing one thing: extracting all these assets that originally existed only in human minds, turning them into "context assets" that are AI-readable, callable, and iterative.
What form does this take? It's detailed requirement documentation; it's process documentation that captures every decision, every email exchange, every Slack discussion; it's open MCP interfaces and APIs; it's the artifacts produced by every internal tool—all these things constitute a company's new, inheritable, non-evaporative layer of assets that won't disappear when employees leave.
Humans in such a company instead become the "variable"—they can be quickly onboarded or quickly depart, because the company's core assets aren't in human minds, they're in the documentation.
If this structure holds, it doesn't just mean a new organizational model—it means the company's balance sheet is being rewritten. A 6-person, AI-native company burning a staggering token bill might appear financially unhealthy, but its true assets could be thicker than a 60-person traditional company—it's just that current accounting standards haven't yet learned how to calculate this type of asset.
In other words, vibe coding isn't dead. It just doesn't belong to traditional companies.
The day Microsoft withdrew Claude Code was not the day AI economics failed—it was the day a posture of attaching AI to old organizations falsified itself.
And in that room of startup companies at YC, another posture is growing—they're small, they're burning, they don't have "employee AI usage rate" on their KPI sheets, their CFOs also don't panic when token bills skyrocket—because what they're burning is not "employees' copilots," it's "replacements for employees."
In the coming years, all medium-sized companies still urging employees to "use a bit more AI" will hit the wall Microsoft hit—a structurally bound-to-rise token bill.
But the real reason for hitting the wall isn't that AI is too expensive—it's that the organization hasn't changed.
And the vast majority of companies probably won't change anytime soon.
