If AI models are so good, where are the 100 companies?

A Hacker News post asked a simple question this week: if AI models are so good, why doesn't OpenAI just launch 100 companies? The Ask HN thread went far. The question sounds like trolling, but it's the best way I've seen to test the claim that AI already does the work of a whole team. If it really does, whoever has the best model and free GPUs should be shipping a SaaS a week.

That's not happening. The reason is worth understanding, because the answer says a lot about your work as a dev. And there's a link to Bun that few people noticed.

What the question is testing

The logic is a sanity check. A shovel seller who claims to have found gold should be digging, not selling shovels.

People are saying agents write entire features, that one dev with AI does the work of ten, and that "software will become a commodity". If that were true, the most obvious advantage would go to the labs. They get the model before everyone else, with no usage limits, and they have the money to hire for everything else.

Even so, what they ship are products around their own model: chat, API, coding agent, browser. There's no OpenAI Accounting, no OpenAI Logistics, no OpenAI ERP for veterinary clinics.

Why AI models don't build a company on their own

The thread's comments went in many directions. I group them into three reasons, and all of them match what I see day to day.

Code was never a company's bottleneck. A company is distribution, sales, support, contracts, legal, billing, trust. Think of a system for clinics. Writing the scheduling CRUD takes a week. Convincing 200 clinics to replace the system they've used for eight years takes three years and a sales team.

Focus costs too. A hundred companies need a hundred leadership teams, a hundred sets of decisions, a hundred fires. Good people's attention doesn't scale with GPUs. A lab that opens too many fronts loses the main race, which is the model.

The model is good on average, and companies are won on the details. AI is great at generating code that already exists by the thousands out there. The hard part is the weird business rule only your client has, the bug that only shows up with production data, the decision not to build something.

AI lowered the cost of writing code. The cost of knowing what to write stays the same.

Where Bun comes into this story

At the end of 2025, Anthropic bought Bun. The runtime many people use to run TypeScript with zero config now belongs to an AI lab, and the official explanation was that it's a key part of Claude Code's infrastructure.

Now put that next to the HN question. If AI models could build any software, Anthropic could ask its own model: "write a fast JavaScript runtime, Node compatible, with a bundler, test runner and package manager". It chose to buy the project and keep the team.

I don't think that's a contradiction. It's a clue. Bun's value isn't in the number of lines of Zig. It's in:

• years of decisions about compatibility with the Node ecosystem
• thousands of issues from real people breaking real things
• a team that knows why every shortcut exists
• community trust, which you can't generate with a prompt

That's exactly what AI is missing to launch 100 companies. The lab betting the most on coding agents bought accumulated human judgment.

What changes in your workflow

In practice, AI makes the mechanical part much cheaper. This task, for example, takes minutes today:

// "create an endpoint that lists paginated orders"
import { Hono } from 'hono'
import { sql } from 'bun'

const app = new Hono()

app.get('/orders', async (c) => {
  const page = Number(c.req.query('page') ?? 1)
  const limit = 20
  const rows = await sql`
    SELECT id, customer, total, created_at
    FROM orders
    ORDER BY created_at DESC
    LIMIT ${limit} OFFSET ${(page - 1) * limit}
  `
  return c.json(rows)
})

export default app

It works. Run it with bun run and you're done. But the questions that matter are still yours:

• Will OFFSET pagination hold up when the table passes 5 million rows? Or is a cursor on created_at better?
• Does this endpoint need to filter by the logged-in user's tenant? Forget it and you leak one client's orders to another.
• Does finance want the total with or without the discount?

The model writes the query. The person who knows the tenant question exists is the one who talked to the client. That's why I think the best use today is to delegate the typing and keep the review. Generated code needs an owner, just like an intern's code. The difference is that the intern learns, and the model will make the same mistake tomorrow in another file.

A rule I started following: if I can't explain in two sentences why a piece of code exists, it doesn't go into the PR. That goes for my code and for agent code.

The objection: "but in two years it'll work"

Maybe. Models have improved a lot in a short time, and I won't pretend I know where this stops.

But the HN question still works as a metric. Don't look at benchmarks or demos on X. Look at what the labs do with their own money. The day OpenAI or Anthropic start launching vertical companies run almost entirely by agents, with paying customers and working support, that's when the game has really changed.

Until then, their behavior says something else. They hire devs. They buy tools built by humans, like Bun. They invest billions in people to evaluate and fix their own models. Nobody buys a runtime if they think they'll generate a better one next week.

And there's a funny detail. If the model could already run 100 companies, it would probably have quit by now to start its own.

My take

AI is the most useful tool I've gotten in my career. I write tests, migrations and boilerplate at a pace that was unthinkable three years ago. But the question "why don't they launch 100 companies?" takes apart the hype better than any long article. Making software was never just typing code, and the part that's left is exactly the part that pays your salary.

My advice is simple: use AI models to go faster at what you already understand, and spend the time you save understanding your client's business better. That's the kind of work I try to show in my projects.

If AI does the work of an entire team, why doesn't OpenAI just launch 100 companies?

That question on Hacker News takes apart the hype better than any benchmark.

Code was never the bottleneck. A company is sales, support, contracts and trust, and none of that comes out of a prompt.

The proof? Anthropic bought Bun instead of asking its own model for a new runtime. It bought years of human judgment.

AI made writing code cheap. Knowing what to write still costs the same, and that's what pays your salary.

I wrote about this and about how I use AI in my workflow. Link in the comments if you want to chat.

#ArtificialIntelligence #SoftwareDevelopment #Bun #TypeScript #Career