The subsidy ends where the margin begins

A $25 billion bond sale is a strange way to say “AI is getting cheaper”.

Yet that is the point. RTTNews reported that Amazon is seeking to raise at least $25 billion in a bond sale to fund AI investments. Put that next to The Decoder’s report that OpenAI and Anthropic are giving away millions in computing power to attract startups, and the pattern gets clearer: the cheapness is not proof that AI infrastructure has become naturally cheap. It is proof that someone else is paying the bill early.

The obvious reading is that AI providers are being generous to win developers. The better reading is that subsidised compute is customer acquisition with a GPU-shaped invoice attached.

Free compute is the SaaS free trial, rebuilt for an industry where the marginal cost can bite back. A startup that builds its product, evals, routing logic, prompts, latency expectations and pricing around one provider is not merely “testing” a model. It is making an infrastructure decision before it has the organisational scar tissue to know what that decision costs.

That is why the subsidy ends where the margin begins.

The stack is becoming the strategy

Microsoft’s reported shift is the cleanest signal. TechCrunch reported that Microsoft is relying more on its own models as part of the AI cost-cutting trend. For users, nothing needs to look different. The button, workflow or assistant surface can stay where it is. Behind the interface, though, the product team can start asking a harder question: which model is good enough for this task at this price?

That question will define AI product engineering.

The early AI era rewarded teams that could find the best model and wrap it in a decent interface. The next phase rewards teams that can route work intelligently: small model here, frontier model there, cached answer when possible, retrieval before generation, human review only where the risk justifies it. Model quality still matters, but model spend is becoming a product constraint, like latency or payment fees.

DeepSeek’s reported chip work points in the same direction from the other end of the stack. Reuters via MarketScreener reported that DeepSeek is developing its own AI chip. Training gets the glory. Inference gets the invoice, every time a user clicks, asks, drafts, searches or retries.

If you are a model company, owning more of the stack is a way to protect margin. If you are financing AI capacity, debt becomes part of the operating model. If you are a product builder, cheap access today is useful, but it is not a pricing strategy.

There is an old economics pattern here. Railways were not won by the prettiest locomotive demonstration. They were won by track, financing, routing rights and operating discipline. AI has its demos, but the durable advantage is moving towards cost per useful action.

The practical lesson is blunt: take the credits, but do not believe them. Use them to learn, ship and collect demand. Then model your unit economics as if the subsidy disappears, because it probably will.

The next AI moat may be less about who can call the smartest model, and more about who can afford to call it twice.

---