Building in the Token Economy

For about two years, building with AI felt free. It wasn’t. Someone was just paying for it.

Now the bill is showing up. Uber capped how much its employees can spend on AI after blowing through the budget in four months. Companies are watching their token bills climb the way they used to watch AWS bills, except faster and with less to show for it. The subsidies that made every prompt feel weightless are quietly lifting, and the cost of building with AI is becoming a real line item again.

I think this is good. Cheap building was fun, but it hid a problem.

When building is basically free, you build everything. Every idea in the backlog becomes a “why not, it’s an afternoon.” Every feature request becomes a yes. The model makes it so easy to generate working software that the hard part stops being the building and becomes the restraint. And restraint is the first thing to go when the meter isn’t running.

That’s the trap. Building with AI is fast, it’s fun, and it feels productive. But shipping code was never the same as moving the business, and a pile of features nobody asked for is still a pile. The token economy just put a price tag back on the difference.

How I decide what to build now

Before I open the editor, I align with whoever has a stake in the outcome. Then I cross-reference the idea against the goal and against the actual numbers, the analytics, what people are really doing in the product. Then I decide what ships and what waits.

The thing that does ship gets a real spec with acceptance criteria, so “done” actually means something. I build it, push it, and then test whether it did the job. Here’s the part most people skip. I don’t leave it open. If it moved the needle, I keep building on it. If it didn’t, I kill it.

That last step is the whole game. Killing what doesn’t work is how you keep the cost of building honest. Tokens you spent on something you then delete aren’t an investment. They’re just a bill.

The upside of cheap, fast building was never “ship more stuff.” It was “find out faster.” You can put a real thing in front of real users in a day and learn something true about whether it matters. That only compounds if you act on the answer. Shipping faster is worthless if you can’t kill faster too.

Bloat is still what kills products

The real killer of products was never the cost of building them. It’s bloat. Every product that got slow and bossy and stuffed with settings nobody touches got there one reasonable feature at a time. Each one made sense in the moment. Together they buried the thing people actually came for.

And every time a product bloats, it opens a gap. Some smaller, simpler tool walks right into that gap and takes the users who just wanted the one thing to work. Simple to use wins. It has always won. AI didn’t change that. It just made it easier than ever to bloat your way out of the market while feeling busy the whole time.

I see it across everything I build. The features that survive are the few that show up in the numbers. The rest felt great to ship and changed nothing, and the honest move is to cut them before they calcify into weight the next version has to carry.

The new estimate is tokens, not hours

Here’s where this goes. The big companies will do what big companies do. They’ll build a version of agile around it. Instead of estimating a feature in engineer-hours, they’ll estimate it in tokens, millions of them, and the planning meeting will argue about the budget the same way it used to argue about story points.

And honestly, that’s better than the alternative. A token budget is at least a number that can say no. It beats shipping everything with your eyes closed and finding out at the end of the quarter that you spent a fortune feeling productive.

But it’s still the wrong unit to fall in love with. Tokens spent are an input, not a result. A feature that burned a million tokens and moved nothing is more expensive than one that burned ten million and doubled retention. Lines of code generated were never a measure of a better product, and the token bill isn’t either. The question underneath all of it hasn’t changed: did this move the needle? Budget the tokens, sure. Just don’t mistake the budget for the point.

What to build when tokens cost money

Less, but the right less.

The cheap-model routing stuff is real and worth doing. Route the easy calls to the cheap model, save the expensive one for when it earns its keep. Kimi-style economics will matter more every quarter, and being deliberate about which model touches which job is going to separate the products with healthy margins from the ones quietly losing money on every user.

But the bigger lever isn’t the model you pick. It’s the work you choose not to do. The most expensive token is the one you spent building something that didn’t need to exist.

AI is incredible at enabling and augmenting the building. It’s a terrible reason to build everything. The founders who win the next couple of years won’t be the ones who shipped the most. They’ll be the ones who shipped what mattered and had the discipline to walk away from the rest.

The free trial is ending. That’s not a loss. It’s the part where building with AI grows up, and the people who already knew what they were building barely feel the change.