Three Laws of AI Microeconomics

A founder friend runs a coding tools startup. Raised in 2024 at a high mark. Good revenue. Still opens Stripe first thing every morning. Then Posthog. Then his investor update doc. Then back to Stripe.

Different worry each week. Sometimes a competitor. Sometimes pricing. Sometimes a feature foundation labs shipped that overlapped his product.

He doesn’t know which one to worry about. That’s the part that gets to him.

There’s a simpler way to think about this. Start with rent.

Every Claude token has a cost. Somewhere a GPU draws power. A datacenter cools itself. A utility bills. A chip depreciates. Intelligence appears free at the application layer because someone else pays the rent at the compute layer.

Most founders building on AI act as if this isn’t true. They price products on current token costs. They treat foundation models as a free natural resource. Like air.

They are not air.

They are office space in Montgomery St with a five-year lease and a clause for annual rent revision that nobody read.

Rent goes up. Slowly at first. Then all at once.

Anthropic raises prices. AWS adjusts margins. Nvidia raises allocation costs. The application company that thought it owned its margin discovers it was renting it.

First law of AI microeconomics. Intelligence is rented from compute. Rent is paid forever.

Watts cost money. Silicon depreciates. Heat must be removed. Physics, not economics.

The second one is harder.

Coffee cools because the room is colder than the coffee. To keep coffee hot you need a flask, a flame, or willingness to drink fast. There is no natural thermos for profit either. In any market with free entry, profit decays toward cost of capital unless something prevents it.

What’s different in AI is the speed.

In SaaS, this took years. Customers were sticky. Sales cycles were long. Replication required engineering effort. Build a product. Hit profitability. Ride the moat for a decade.

AI runs at compute speed. Model commoditizes in weeks. Wrapper commoditizes in months. Harness commoditizes in quarters.

Every AI application company is learning this right now. They thought their moat was the model. The model commoditized. They thought their moat was the UX. The UX commoditized. They thought their moat was the integration. The integration commoditized.

The hard part isn’t that founders missed it. In fact they saw it coming. Saw it in their dashboards. Saw it in their churn cohorts. Saw it in the foundation lab release notes. Still couldn’t escape.

You don’t really build a moat. You keep working on it.

Second law. Profit without defensibility is heat dissipating into a cold universe.

Anthropic ships every two weeks. Not optional.

This is also why their marketing has the tone it does. Dario makes sure of it. Every announcement reminds everyone the moat exists. Silence reads as decay. The “SaaSpocalypse” framing isn’t sales tactics. It’s the same maintenance work as the shipping. Different surface.

Friends in software companies wince. They wince at the wrong thing.

There is a floor.

Third law of AI microeconomics. At zero defensibility, profit approaches cost of capital. No company operates below it for long.

There are only three exit paths. Build the moat. Sell to someone who needs the moat. Die.

The AI wrapper graveyard. Every vertical AI startup at SaaS-era multiples without SaaS-era moats is racing the third law. Either the moat materializes and the valuation is vindicated, or the multiples collapse to cost-of-capital levels and the company gets acquired or winds down.

Acquisition is the soft landing. xAI buying Cursor at $60B is the third law expressing itself through M&A instead of bankruptcy. Price was high because the acquirer needed the asset before the third law fully priced it. Once the third law fully prices it, the asset isn’t worth $60B.

This is why fast acquisitions happen at the top of the AI cycle. Acquirers buy before the third law catches up. Sellers sell because they can read the curve. Buyers pay a premium for time.

The third law is patient. The premium for time is finite.

Every AI company has an implicit clock. Measures time until the third law fully prices the company’s defensibility. Companies with real moats reset the clock by shipping. Companies without real moats run out the clock.

Clock is invisible to founders sometimes. Visible to investors. This is why down rounds happen.

The laws explain things that otherwise look strategic.

Microsoft has distribution. Optimizing for what the first law lets them capture. Distribution rent. They don’t need to win the model layer. They need to keep their position above it.

Google has compute and a model. Trying to fold harness on top. What Anthropic has. Paying $750M to partners to do it because organic folding has been hard. Partner money is a hedge against the second law. Buying maintenance work they can’t do internally fast enough.

xAI has compute and a CEO with extraordinary will but weak harness. Bought Cursor to import a harness. This will probably not work. The second law doesn’t accept imports gracefully. Imported moats decay faster than native ones because the team that built them is no longer the team maintaining them.

Anthropic has the unusual combination. Model, harness, UX, governance in one team. Second law operates more slowly against them. They get to ship faster than dissipation. For now. This will not last forever. Nothing does. But it’s lasting longer for them than for the others. That’s what shows up in the valuations.

Open source is the most interesting one. Not fighting the laws. Accepting them. Shifting where the rent flows. Instead of paying Anthropic for inference, you pay the cloud provider. Or you pay yourself in GPU depreciation in your own datacenter.

First law still applies. Rent is paid forever. Recipient changes.

Second law still applies. Profit dissipates. Dissipation happens at the commodity layer where it was going to happen anyway.

Third law still applies. Cost of capital is the floor. Floor is lower because there’s no application margin to defend.

Open source isn’t a competitor to Anthropic. It’s a different equilibrium. Anthropic serves customers willing to pay for frontier intelligence. Open source serves customers who want commodity intelligence at commodity prices. Both will exist permanently because the demand curve has both ends.

The market is not a war with a winner. The market is a curve with two stable points. Both will be served.

When my friend says he’s worried about a competitor, I tell him the competitor isn’t the problem. They’re just where it’s showing up.

The competitor is exploiting the second law against his specific moat. The pricing pressure is the first law showing up at the application layer. The acquisition offer is the third law approaching. The buyer is paying to get ahead of it.

Naming it helps. Otherwise everything feels like a threat.

What you do about it is a separate question. Some fight harder. Some sell. Some pivot to the long tail where these things move slower. Some leave AI entirely and build something where decay runs at SaaS speed.

All reasonable. Pretending the laws don’t apply is the only unreasonable answer.

The thing is and this took me a while to see it’s not actually changing, just moving.

Same forces that closed the railroads. Same forces that hollowed Kodak. Same forces that ate SaaS gently enough that strategy felt like it mattered more than physics. Apply to AI. Apply faster.

Compute is faster than human coordination. The laws follow compute.

The three laws aren’t new. Just faster now.

Still figuring out what that really means.