I Watched A 50-person Dev Shop Get Vaporized In 12 Months And The Ceo Is Still Optimistic

I rent a desk in a company that doesn't really exist anymore.

A year ago this place had 50 developers packed into an open space. Real ones. Their entire business was low-code for enterprise clients. €200k contracts to deliver in three months what full-stack would deliver in nine for €600k. Beautiful business. Defensible moat. Everyone home by 6 PM.

Last year I told the CEO he should integrate AI into the workflow. Aggressively. Now. He smiled at me the way you smile at a child explaining clouds. « Our sector is protected. Low-code is too specialized. Our clients need expertise »

Twelve months later. Zero new clients. Mass layoffs. The upper floor is empty. Chairs still there, monitors still on, post-it notes from the last sprint still stuck to the walls. The company survives on maintenance contracts for projects deployed before the asteroid hit.

What killed them is almost too poetic. Full-stack development with AI is now faster and cheaper than low-code. Read that twice. Low-code's entire value proposition for fifteen years was "faster and cheaper than full-stack." That was the floor. AI didn't lower the floor. AI deleted it.

The CEO still walks the empty floor like a captain inspecting a perfectly fine ship. Sometimes he stops at a desk that hasn't had a person at it for four months and adjusts the chair. He'll figure it out eventually. Probably during the bankruptcy hearing.

Now let's talk about the math, since we're here.

I've been pricing AI infrastructure for the last six months and the numbers are absurd in both directions.

The Forbes story about Anthropic losing $5,000 per Claude Code Max user is fiction. The number confuses retail API pricing with actual compute cost. Real compute is roughly 10% of retail. Anthropic loses maybe $300/month on the worst heavy users, the gentlemen running Claude Code in agentic loops while they "sleep," for very generous values of sleep, and prints money on everyone else.

Max 5x at $100? Profitable. Max 20x at $200? Fine on average. The "AI labs are bleeding cash on inference" narrative is mostly Cursor doing PR for itself, because Cursor actually pays retail and is, in fact, hemorrhaging. Anthropic loses money on training runs and on paying ML researchers $4M a year to threaten to defect to Meta every six months. Not on you. You're not the protagonist of this thriller. You're an extra.

But the vise is closing anyway. Memory chip shortage runs through 2027. AI data centers will eat 70% of global memory production in 2026. Half the US data centers planned for this year have been canceled or delayed. Transformer lead times went from 18 months to 48 months. The physical world, charmingly, does not care about your roadmap. Anthropic is openly compute-constrained. Dario said so on a podcast, in the calm voice of a man holding the only water in the desert.

Now let's talk about the sessions limit.

Because this is where the abstraction becomes a humiliation.

You hit the wall around 2 PM. You were in flow. The architecture was finally clicking. Six tabs of context, three half-finished functions, you were five minutes from cracking the bug that's been mocking you since Tuesday. Then the message appears:

« You've reached your session limit. Your usage will reset in 1 hour and 47 minutes. »

You sit. You stare. You consider, briefly, doing the work without AI. You open the file. You look at the function. You realize you have, somewhere in the last eighteen months, forgotten how to do this. Not all of it. Just enough of it. Theoretically possible. Practically humiliating. You'd produce something. It would take four hours. It would be slightly worse. You'd hate yourself the entire time.

So you don't. You do what everyone does. You wait.

You make coffee. You make a second coffee. You read Hacker News. You read the same Hacker News post three times because you weren't really reading it the first two times. You consider going for a walk. You don't go for a walk. You watch the timer. The timer moves at a pace that suggests time itself has been throttled along with your tokens. You check email. You answer email you would normally have ignored. You reorganize your downloads folder.

« 43 minutes remaining. »

You feel, briefly, the precise emotional texture of a 19th-century factory worker waiting for the lunch whistle, except you're alone, in your home, and the whistle is being blown by a server farm in Virginia.

The truly remarkable part is that we pay $200 a month for this. The factory worker at least got paid by the hour. We pay by the month, get rationed by the hour, and call it a productivity tool.

« I'll just self-host an t-weight model. »

Sure. Show me the invoice. I have time. I'm waiting for my reset.

Kimi K2.6 in production: 8x H100, $450k for the GPUs alone. Add electricity, cooling, an ML ops engineer who knows vLLM and won't quit the moment Anthropic offers him $400k to come maintain the very thing you're trying to escape. Three-year TCO: $700k to $1.2M. For one model. Which you'll need to replace in 18 months.

Gemma 4 is the only model that makes economic sense for actual humans, and that's because Google is using it as bait to drag you onto Vertex AI. It's a comfortable trap. Google has had thirty years to perfect comfortable traps.

The fundamental math: when you pay $200 a month for Claude Max, you're sharing Anthropic's multi-billion dollar infrastructure with millions of suckers exactly like you. When you self-host, you pay alone what millions pay together. This will not work for an individual. It will not work for 95% of SMBs. The Medium articles claiming "18x cheaper per token" are technically correct in the way "you can technically eat 50 hot dogs in one sitting" is technically correct.

The new working class.

Going back to pre-AI work isn't an option. Not philosophically. Economically.

Refuse AI today and you pay a triple penalty. Worse output. Slower delivery. Higher cost per unit of work. Your competitors deliver in one hour what takes you four. The market silently raised the floor on "good work" and didn't email you the new standard.

The 50-dev shop downstairs didn't refuse AI. They just didn't move fast enough. That's the genuinely terrifying part. They weren't Luddites. They were normal, competent people running a normal, competent business who assumed they had time. They didn't. Nobody does. The window between "AI is interesting" and "AI just incinerated your business model" was twelve months and they spent it congratulating themselves on a moat that had been quietly drained.

Three classes are forming.

The Solvent Augmenteds, paying $200 to $2000 a month, productivity 3-5x, capturing disproportionate value. Maybe 5-10 million worldwide. The lucky ones, which means the ones with the most to lose.

The Constrained Augmenteds, free tier or 20$ subscription, hitting limits constantly, told « you have 3 messages remaining today »while trying to finish a deliverable so their kid can eat next week. The dystopia is here. It's just billed in tokens.

The Non-Augmented, by choice or by cost, producing 2023 output in a 2026 market. Watching their quotes get rejected as "too expensive" while charging the same rate they did five years ago. Slowly becoming unemployable while telling themselves they're old school. It is not about being old school. It is about math. Math doesn't care that you're old school. Math cashed out and bought a yacht.

The cosmic joke.

AI was supposed to free humans from drudgery. What it actually did, short-term, is make us more dependent on infrastructure we don't own than we have ever been in human history. We weren't freed from work. We were freed from the possibility of working without it. Subtle distinction. Brutal implication.

Your means of intellectual production quietly changed hands. You used to sell your brain. Now you rent your access to one and stop owning it the moment you miss a payment. The knowledge worker's independence, always more myth than reality but a comforting one, got structurally downgraded, and almost nobody noticed because the productivity gains were too euphoric to think clearly about.

What you can actually do?

Move now. Not next quarter. Not after the holidays. Now. The 50-dev shop downstairs had twelve months and burned them. You don't know how many you have. The honest answer is fewer than you think.

Welcome to the new working class. Membership dues: $200 a month. Going up. No union. No exit. No alternative. The food is decent and the productivity is unreal and one day, probably soon, the bill will be more than you can comfortably pay and you'll pay it anyway, because the alternative is becoming the 50-dev shop downstairs, and you've seen the 50-dev shop downstairs, and you don't want to be the 50-dev shop downstairs.

I'll be in the queue at 5 AM Pacific when the weekly cap resets. Bring coffee. We have an hour to work before they take it back. After that we can stare at the wall together. The new monasticism. Pray, wait, refresh.

« Reset is in 1 hour and 47 minutes. »