Why Are Programmers Paying to Keep Their Jobs?

An AI subscription costs about as much as a phone bill. That comparison is the problem.

Programmers buying AI tokens out of pocket looks less absurd through a certain lens: Tokens aren't expensive—maybe a hundred yuan at most. Companies buying them risk misuse and resale. Individuals buying their own is the sustainable approach.

The math checks out. But look at the assumption buried in it: the premise treats AI programming tools like a phone plan. Whoever uses it pays for it.

That premise is what makes this unsettling.

I

For the past twenty years, the division of programmer "means of production" was simple:

• The company provides the computer.

• The company pays for the IDE license.

• The company grants cloud access.

• The company covers databases and monitoring.

These were called "infrastructure." They weren't part of your salary. They defined the boundary of the employment relationship. Without them, you couldn't work.

Now AI Copilot arrives. It's equally necessary. Yet it's been slotted into a different category: personal skill-enhancement tool.

A real conversation:

Programmer: "Can we get an enterprise account?"
Technical Director: "That's for personal efficiency. Budget's tight. Use your own for now, and we'll see."

"We'll see" means: if commit volume doesn't rise, that's your problem. If it does, that's good management. Either way, the monthly fee isn't the company's concern.

The amount is less than 1% of a monthly salary. But the category has shifted. Companies used to absorb tooling costs. Now employees pay for the right to not be left behind.

That's not consumption upgrading. That's paying to stay at the table.

II

Why won't companies cover it? Management ran the numbers and concluded AI doesn't save headcount.

Day one: AI writes the skeleton. Code volume spikes. Dashboards look great.

Week two: AI-generated code starts failing on edge cases.

Week three: manual review time exceeds the time "saved" during generation.

Week four: the team spends two days on a concurrency bug the AI introduced. If they'd written the code by hand, that bug would never have existed.

This isn't bad luck. It's a structural property of AI programming: generation is fast; delivery is not.

Most software cost isn't in writing the first version that runs. It's in keeping the code from falling apart under continuous change. AI is good at the first part and almost useless for the second. In a legacy system, the cost of tearing down and rebuilding AI-generated code is orders of magnitude higher than patching it manually.

Management dashboards don't have a line for "complexity debt." They see: AI writes a module in ten minutes. Why do you need three days?

So programmers are trapped. They use AI to produce more half-finished work that needs verification. Then they spend weekends and late nights—and more expensive Claude Opus credits—patching the holes. The company pays for code volume. The programmer bears the maintenance cost alone.

When the subscription cost shifts to the individual, the causal chain is straightforward: management sees no headcount savings → enterprise budget gets cut → programmers renew out of pocket to meet the AI-era output standard.

III

On the vendor side, the story isn't better.

OpenAI and Anthropic started with a labor-replacement model: replace a programmer making six figures with a few hundred dollars per month in AI subscriptions. They called it a productivity revolution. The sales pitch was aimed at enterprises.

But enterprises ran ROI calculations and found the numbers didn't hold. Code generation was faster. Trouble tickets weren't fewer. Maintenance cycles weren't shorter. Technical debt kept accumulating. The hours saved fell short of the new debugging costs. Enterprise budgets tightened. B2B orders slowed.

To keep cash flow, vendors pivoted to consumers. The $9.99 monthly subscription became the primary revenue stream. They went to market selling a productivity lever and ended up selling premium phone credit.

This downgrade matters for the whole industry: when token prices truly reach phone-bill parity, the technological premium disappears. Investors stop treating large language models as productivity disruptors. AI becomes a better autocomplete.

After the bubble deflates, the remaining giants have one reliable move: keep more users subscribed. Programmers—frequent users, relatively price-insensitive compared to their hourly rate, under the pressure of "if I don't subscribe, I might fall behind"—become the steadiest paying cohort. They fund, with their fixed monthly fees, a system that was supposed to work for them but instead creates more acceptance work than it saves.

IV

There's a subtler cost that has nothing to do with money.

In the old model, code review meant one person reading another's code—logic, edge cases, design. Both parties built shared understanding of the system.

In many teams now, code review has become "AI review." The reviewer doesn't read the code. They feed it to an AI, run a check, and merge if it passes. When a bug surfaces, the author says "AI generated it," and the reviewer says "AI approved it."

AI has become a cushion in the accountability chain.

After a year or two of this, a subtle erosion sets in. Programmers find it harder to construct a global causal map of the system without AI assistance. The old skill of zeroing in on a problem—relying on experience, intuition, and intimate familiarity with the codebase—gets displaced by a workflow of "let the AI try four directions and wait a minute for each."

In daily development, this doesn't matter much. It even saves mental energy. But when an extreme failure hits—one that AI can't cover, like a legacy-specific data state or a non-standard external response—this generation of programmers, conditioned to "let the AI try first," will find their diagnostic instincts blunted. At that point, it's no longer about token costs. It's about a rupture in accumulated technical expertise.

V

Back to the "rational explanation."

A few dozen yuan isn't much. Paying it yourself doesn't really matter. But the money was never the source of the unease.

The unease comes from this: the company no longer promises to provide the tools you need to work effectively—yet it reserves the right to eliminate you by AI-era standards.

When "individual purchase of AI tools" becomes the uncontested default, the high-salary, high-tech, high-welfare veneer of the IT industry starts peeling from the inside, worn away by a tiny recurring expense. What we're witnessing isn't AI-empowered work. It's the complete externalization of productivity tool costs—practitioners paying, rationally and willingly, for a ticket to more intense competition.

The irony: the AI vendors that profit from these token sales will use that revenue to sell Wall Street a story of efficiency gains—then roll out a more expensive tier. For your renewal next month.