Who Really Wins From Subsidized AI

A reply to Haseeb Qureshi's "The Rise of the 149-Person Company." He is right that cheap AI pushes companies to stay small. A couple of things worth adding to his picture: who else this cheap price quietly helps, and why the biggest companies may be less stuck than they look.

---

Right now, small companies get a remarkable deal on AI. Pay one flat monthly fee and use as much as you want. Pass about 150 staff and it changes: you start paying by the unit, the way a taxi charges by the mile. The investor Haseeb Qureshi, in The Rise of the 149-Person Company, says this gap is changing how companies get built, rewarding founders who stay small on purpose. He is right, and the gap is wider than most people notice. It is worth following to both ends, though, because the same cheap price is doing more than pushing companies to stay small.

How good the deal is

Start with just how good. A research firm, SemiAnalysis, bought every plan and ran the AI as hard as they could, day and night, until they hit the limits. A heavy user, they found, gets something like forty to seventy times more value than paying for the same work by the unit. The sellers can afford this because almost nobody runs the AI that hard. It works like a gym: the price stays low because the average member goes twice in January and never comes back. The heavy users are quietly subsidized by everyone who signs up and barely shows up.

For a small company, this is close to free money. Each extra use costs nothing until you reach the ceiling, and most firms never get near it.

Big companies do not get this deal. Once you pass about 150 staff, the seller moves you onto an "enterprise" plan and charges for every use, at full price. Same software, very different bill. Qureshi calls the jump a tax on AI labor. Call it what you like: a meter, an extra charge, a price that only big companies pay. The effect is simple. At the small end, using more AI is free. At the big end, it adds up fast.

His main point is that this gap changes how companies get built. If staying under 150 keeps your AI nearly free, then 150 becomes a line nobody wants to cross. He points to France, where labor rules get much heavier at 50 staff, so a suspicious number of French firms stop hiring at 49. Expect the same here, he says: tiny teams, lots of AI, few people, everyone bunched just under the line.

He may well be right. It is worth looking at both ends of that line, though, because the same price pulls differently at each one.

Another winner, if it does the work

Start at the small end, where Haseeb is most clearly right. The startup that runs AI around the clock, swarms of bots writing code overnight while the founders sleep, is getting more out of this deal than anyone. Free, unlimited use is rocket fuel for a lean team racing to out-build its rivals. If the 149-person company is the big winner here, this is why, and the point stands.

There is a second winner at the same end, though, quieter and easy to overlook: ordinary small business. The accountant, the law office, the marketing shop, the clinic, the local print shop. They will never push the AI hard enough to feel a limit, so the unlimited-use part barely touches them. What the flat price changes for them is something else. It is small, predictable, and you put it on a card and expense it. No budget meeting. No purchase order. No call with a salesperson. No year-long contract.

That sounds minor. For this group it is the whole thing. For decades, what kept a small business off serious software was everything around the price: the approval, the contract, the sense that this was a Big Decision someone had to sign off on. The monthly fee itself was rarely the problem. A flat, cheap, sign-up-yourself price clears all of that away. It turns AI from a project into a line on the card statement, and the numbers show the door swinging open: the JPMorgan Chase Institute, which sees millions of small firms through their bank payments, found AI use among small businesses jumped from under 2% in 2019 to nearly 18% by the end of 2025, most of them on the cheapest tier there is. A wave that size comes from one thing: starting got easy.

So the same subsidy does two jobs. The first is the one Haseeb names, and names well: an innovation subsidy that hands a lean startup the means to out-build everyone. The second is quieter. Call it an adoption subsidy. It opens a door that was shut before, and opens it for a very large number of firms at once. Whether that turns into anything is the open question.

Because an open door is only an opening, and this is the honest part. The same tool sits on the same cheap plan for every accountant and every agency, so simply having it raises the floor for all of them and hands none of them an edge. Most small firms still use AI for small jobs, and survey after survey finds the real barrier was always the same: knowing what to do with the tool. The flat price solved the easy problem, the cost of getting in. The hard one, the skill to turn cheap AI into an advantage, is still theirs to solve. The startups already live by that. The small businesses that win will be the ones that treat the open door as a starting line.

The giants aren't as stuck as they look

Now the other end. The fear is that big companies are the losers here, taxed so heavily on AI that they cling to their staff and miss the shift. This is partly true. Their developers really are watching the meter. Uber handed AI coding tools to thousands of engineers late last year and used up its whole AI budget for the year in about four months. Microsoft started pulling AI tools from some teams once the bills landed. They feel it.

But two things make this less grim than it sounds.

First, the meter is charging them for the wrong thing. It falls on open-ended experiments, engineers running the AI hard to see what works. Haseeb is right that this is where a big company is exposed: the small, risky, surprising uses are exactly what a fast-moving startup finds first. But that experimenting was never what kept the giant alive. Its real edge is the stuff a startup doesn't have: years of its own data, deep knowledge of how its business runs, a huge base of customers to put a finished tool in front of. None of that is billed by the use. JPMorgan put an AI assistant in front of more than 200,000 staff, built on the bank's own knowledge. What makes that valuable has nothing to do with the price of an experiment.

Second, the giants are not trapped. If the bill ever truly threatened them, they have ways out. They can run cheaper open models on their own machines. They can cut private deals. The biggest tech firms already design their own chips to drive the cost down. Most don't bother, and the reason they don't is revealing. Leaving has its own price: new systems to build, scarce engineers to pay, a year or two of the open models trailing the best. Today that price is higher than the bill they already pay, so they pay it and carry on. The sellers have simply set their charge just below the cost of walking out. For now.

What actually slows the giants down is duller than any price. It is themselves. Most big-company AI projects stall, and the reason is rarely cost. Changing how a large company works is slow and hard, and most of their AI trials so far have produced nothing you can measure. The meter is not what holds the giants back. They are.

The 150 line won't sit still

There is one more reason not to build your whole company around staying at 149. The deal might not last.

When something gets cheap, people rarely use a little more of it. They use far more. An economist, William Stanley Jevons, noticed this back in 1865: as steam engines got better and coal got cheaper to run, Britain burned more of it. We are watching the same thing with AI. The cheaper it gets, the more uses we find for it, which is why those flat prices are starting to strain. Uber's runaway bill is that same story, sped up.

The sellers can see it too, and they will not subsidize it forever. GitHub's coding assistant switched everyone to paying by use this June. Anthropic, the company behind Claude, recently made its big-company plans stricter, stripping out the bundled usage that used to soften the bill. The cliff Haseeb spotted got steeper. A flat, all-you-can-use price is a moment in time, and building a permanent company on a passing price is a shaky plan.

What to take from this

Put the two ends together and the lesson is the same at every size.

If you run a small business, this is your moment to walk in. The tool a giant pays a fortune for, you can rent for the price of a lunch, today, with nobody's permission. The catch is that the tool is the easy part. The advantage goes to whoever knows what to point it at.

If you run a big one, stop treating the AI bill as your main problem. Your data, your people and your slow inner workings matter far more. Spend less energy counting the meter and more turning the things a startup cannot match, your own data and the customers you can already reach, into something those customers feel.

And if you are a founder deciding how big to grow, stay small if you like. Just do it for the right reason: small teams move faster and waste less. Build the company on the advantages cheap tools can't copy: judgment, trust, the work only you can do. Those hold still while the price of AI keeps moving.