If AI Coding Is Not Winner-Take-All, Gross Margin Decides Who Survives

Key takeaways

• Cognition's $48B valuation is being read as a signal that investors expect several durable players in AI coding, not one.
• Winner-take-all markets reward land grab. Fragmented markets reward margin.
• AI products carry a variable cost per user that traditional SaaS does not, so growth can make the P&L worse.
• Flat pricing on a usage-driven cost base is the most common way AI startups quietly break.
• The relevant metric is gross margin per user per cohort, not blended gross margin.

What the signal actually is

TechCrunch's framing is the interesting part: a valuation higher than what Cursor commanded before its SpaceX sale is being read not as one company pulling away, but as investors concluding that AI coding supports multiple large winners.

If that read is correct, it changes the operating advice for everyone building in and around this space. In a winner-take-all market, burning margin to capture share is rational, because share converts into a monopoly later. In a market with several durable players, that same burn is just a smaller company.

Why does market structure change your pricing?

Because it changes what your losses are buying.

In a market that consolidates to one, negative gross margin per user is an investment. You are purchasing a position that will later be defensible and repriced. In a market that stays fragmented, negative gross margin per user is simply a subsidy your competitors do not have to match, and there is no future moment where you get to raise prices without a customer who can leave.

So the strategic question is not whether you can grow. It is whether growth compounds value or compounds cost.

Why is AI gross margin harder than SaaS gross margin?

Classic SaaS had near-zero marginal cost per user. Serving your ten thousandth customer cost roughly nothing more than your hundredth, which is why the entire playbook of flat pricing, generous free tiers and land-and-expand worked.

AI products broke that. Every active user consumes tokens, and heavy users consume disproportionately. Which means the shape of your cost curve is set by your power users while your revenue curve is set by your price list, and those two curves do not track each other.

Here is the point I think is under-discussed: in AI products, your best customers can be your least profitable ones. The user who runs your agent all day is the one giving you retention, expansion and word of mouth, and they are also the one with the worst unit economics on a flat plan. Traditional SaaS instincts say delight them. AI unit economics say measure them first.

How do you tell if your pricing is holding?

Blended gross margin will hide the problem for a long time, because light users subsidize heavy ones until the mix shifts. Two better views:

Gross margin per usage decile. Split your users into ten buckets by consumption and compute margin for each. If the top decile is negative, you have a time bomb, because product-market fit pushes more users into that decile over time.

Cost per user over cohort age. Track inference cost per user by months since signup. If it rises, your customers are learning to use the product more, which is good for retention and bad for a flat price.

If both charts point the wrong way, you do not have a cost problem, you have a pricing model problem, and no amount of prompt optimization will fix it.

What are the actual options?

Three, roughly, and most products end up blending them:

1. 1Keep flat pricing and cap usage. Simple to sell, predictable margin, but caps annoy exactly the users you most want.
2. 2Move to usage-based or credits. Margin follows cost automatically, but adoption friction rises and customers dislike unpredictable bills.
3. 3Hybrid: a base fee plus metered overage. Predictable for the customer at normal usage, protected for you at extreme usage. This is where most AI products with durable margins land.

Whichever you choose, model it against your real usage distribution before you announce it, not against an average user who does not exist.

In a market that does not consolidate, nobody gets rescued by a future monopoly, so the company that knows its gross margin per user is the one still standing in three years. You can model tiers, credits and margin per user in Calcaas before you commit to a price list.