AI labs are not software companies. They are airlines.

Investors still talk about AI labs as if they were software companies: high margins, a winner that takes most of the market, a moat that lasts. That is the wrong model.

Frontier models are heading toward airline economics. The product is becoming good enough across providers. Customers can switch cheaply. They will buy on price. You can have a huge industry and still fail to keep an attractive return.

The punch

If performance is close and switching costs are low, price becomes the strategy. That is not a temporary discounting phase. It is what the industry structure produces.

Airlines already taught this. Air travel is essential. The market is enormous. Major carriers fly similar planes on similar routes. A passenger treats a seat as a seat and buys the cheaper fare. The result has been chronic weak returns, even in years when planes are full.

Token generation is starting to look like that seat.

Three facts, not a technology story

1. Performance is converging. OpenAI, Anthropic, Google, and strong open-weight models now clear the same 'good enough' bar for a wide range of work. Leads still move. They move in months. That is not a decade of uniqueness. Once several options clear the bar, the remaining comparison is cost, reliability, and how easy it is to swap.

2. Switching costs are low. A user can open a different chatbot tomorrow. An engineering team can change a model name in an API call, or put a gateway in front of three vendors and route to the cheapest one that passes evals. That is not replatforming an ERP. Procurement will run a bake-off and pick the lowest cost at acceptable quality. That is what procurement does.

3. The rational response is a price war. Training and serving are expensive. Idle GPUs do not sit on a shelf like inventory. Airlines discount empty seats. Model providers discount empty clusters. Once price is the main competitive variable, margin is what is left over, not what was planned.

Same structure, same outcome

Michael Porter's point is simple: industry structure sets the ceiling on profits. Growth does not. A large, exciting market can still be a bad business if rivalry is brutal, buyers can play vendors off one another, and firms cannot make the customer care who they buy from.

	Typical software	Airlines	Frontier models
Differentiation	High	Low	Falling fast
Switching costs	High	Low	Low
How buyers choose	The product	The fare	The price per token
Expected margins	Attractive	Structurally weak	Structurally weak
Who keeps the surplus	The vendor	Almost nobody, or the plane maker	Compute suppliers, maybe apps

Read the last two columns. They rhyme. That is the thesis.

Where this can still be wrong

A consumer app with memory, habit, and a workflow people live in can raise switching costs. An enterprise deal with compliance, data rules, and agents already in production is stickier than a hobbyist API key. Regulation can also shrink the set of approved vendors. Those are paths to better structure. They are not the default at the model layer.

Capital intensity does not save you. Airlines' planes are expensive. AI training runs are expensive. High fixed costs do not create pricing power. They create a reason to fill capacity, which usually means discounting. Capex is not a moat unless you can refuse unprofitable volume.

So who actually keeps the money?

Value flows to the parts of the system whose structure can hold it. In air travel, that has often been aircraft makers and airports, not the average carrier. In AI, the analogous places are scarce compute (Nvidia and the clouds that control distribution) and applications that own the workflow, the data, and the customer relationship.

The model, sold as a substitutable token, is set up to compete like a fare.

What to do with this

If you are building with models, assume you will multi-home. Make the model swappable. Compete on the job you are replacing: the workflow, the data, the trust, and the unit economics. Do not build a company that only works if you have exclusive access to a slightly better checkpoint.

If you are investing or partnering, do not underwrite software-monopoly margins for a business that looks, on structure, like a high-growth airline. Size of market is not attractiveness of industry. Airlines taught that at global scale. AI may teach it faster.