The AI CapEx Question Every PM Should Ask

The number is $700B, and it is aimed at your roadmap

Amazon, Google, and Meta have signaled combined AI infrastructure spending near $700 billion for 2026 — roughly two-thirds of the annual US defense budget. That capital is buying data centers, custom silicon, power contracts, and the model training runs that sit on top of them. If you build product, that number touches what you ship, what you charge, and what you can promise a customer. Most PMs treat it as macro news. It is not macro. It is your next pricing conversation.

Start with where the money actually goes. The headline figure hides three distinct buckets, and they pay off on different clocks. There is compute for training frontier models, which is a bet on capability that may or may not translate into products anyone pays for. There is compute for inference, which is the cost of serving every query your users make, and it scales with your success rather than your ambition. And there is the physical layer underneath both — land, transformers, cooling, and the multi-year power purchase agreements that Packy McCormick has written about at length in his work on the data center buildout. That physical layer is the slowest to move and the hardest to reverse. When a hyperscaler commits to a gigawatt of capacity in 2026, that commitment shapes what inference costs in 2028.

The reason this matters to a product person and not just a CFO: you are a tenant in this economy. You do not own the building. You rent capacity from the three companies making these bets, and the terms of your lease are set by their capital cycle, not your roadmap.

The second-order effects reach further than the API bill

The obvious first-order effect is price. Inference costs have fallen fast, and every PM who has watched token prices drop has quietly assumed they will keep dropping. Maybe. But the $700B is not being spent to make inference cheap for you. It is being spent to build capability moats and to capture the demand those moats create. Falling per-token prices and rising total spend can coexist, and they are coexisting right now. The unit economics improve while the strategic dependence deepens.

The second-order effects are where the real product consequences live.

First, capacity becomes a scheduling problem, not a spending problem. When training runs consume the same GPUs that serve inference, your latency and availability are hostage to someone else’s model release calendar. A frontier training run that soaks up capacity for six weeks is not on your roadmap, but it lands on your uptime.

Second, the economics push every provider toward vertical integration. Google has its own silicon. Amazon has Trainium and Inferentia. Meta is designing chips to escape its own dependence. Byrne Hobart has framed this AI CapEx cycle as a race where the winners are the ones who own the deepest part of the stack, because that is where margin accumulates when the buildout matures. For a PM, that vertical integration means the abstraction you build on is not neutral. The provider that owns the silicon has every incentive to make its own models the cheapest path and everyone else’s the expensive one.

Third, the spend concentrates power in exactly the way Ben Thompson’s aggregation theory predicts. In “Aggregators and AI,” he argued that the companies controlling demand aggregation will bend the AI supply chain toward themselves. The $700B is the physical expression of that argument. Three companies are buying the ground floor of the next computing platform, and they are doing it because owning the infrastructure lets them own the relationship with everyone who builds on top.

Build, buy, or route — and why most PMs get this backward

Every AI feature reduces to one of three postures. You train your own model. You buy access to someone else’s through an API. Or you route dynamically across several providers and models depending on the task. Most product teams pick one and defend it as identity. That is the mistake.

Training your own model made sense for almost no one in 2024 and makes sense for slightly more teams now, but only where the model is the product and the data moat is real. For the rest, training is a way to spend the $700B problem on yourself at a smaller scale and lose. The capital intensity that makes hyperscaler CapEx a headline is the same intensity that will drain a startup’s runway.

Buying through a single API is the default, and the default is a strategic risk disguised as convenience. When you build your entire product on one provider’s model, you have handed that provider your pricing power, your latency, and your roadmap dependencies. The moment they raise prices, deprecate a model, or throttle capacity during a training crunch, your product feels it and your customers hear about it.

Routing is the posture that treats the $700B as what it is — a competitive supply market you should shop rather than marry. Route by task, by cost, by latency, by the provider’s current capacity. Build the abstraction layer that lets you swap a model without rewriting your product. This is not hedging for its own sake. It is the recognition that when three companies are spending defense-budget money to lock in demand, the durable position is the one that stays liquid.

The counterargument is that routing adds complexity and that a single provider’s tight integration ships faster. True, early. A team racing to a first version should buy the simplest path and get to market. But treating that first choice as permanent is how you wake up in 2028 with a product whose margins belong to your model vendor. Cut scope to ship, then build the routing layer before the lock-in hardens, not after.

The infrastructure bets that actually pay off

Not all of the $700B pays off, and the parts that do will not pay off evenly. Training capability is the loudest line item and the least certain return. Frontier model capability is converging, and the gap between the best model and the third-best is narrowing to the point where, for most product tasks, it no longer decides the outcome. The teams betting everything on always being on the absolute frontier are betting on a gap that keeps closing.

The bets that compound are the ones lower in the stack. Custom silicon pays off because it lowers the marginal cost of inference for the owner and no one else. Power and data center capacity pays off because it is scarce, slow to build, and required by everyone. These are the parts of the spend that create a durable cost advantage rather than a temporary capability lead.

For a PM, the read-through is direct. The value you can capture sits in the layer above the model, not in the model itself. The model is becoming a commodity input. The product that wins is the one that turns that commodity into a workflow, a relationship, and a set of proprietary data that gets smarter with use. A self-learning system built on top of a routed model layer outlasts a clever prompt built on top of a single frontier bet. Own the relationship, not the token.

The question to bring to your platform team this quarter

Here is the strategic question every PM should put to their platform and infrastructure teams, and it is not “which model are we using.” It is: “If our primary provider doubled inference prices tomorrow, or throttled us during their next training run, what happens to our product and our margins, and how fast could we route around it?”

The answer reveals your real exposure to the $700B. If the answer is “we’d be fine, we can shift traffic in a day,” you have built the abstraction correctly and you have earned the right to move fast on top of a market you do not control. If the answer is a long silence, you have a dependency you have not priced, and the capital cycle that produced that $700B figure is going to price it for you.

The spend is not slowing. Three companies have decided that owning the infrastructure of the next computing platform is worth two-thirds of what the country spends defending itself. That decision flows downhill to every product built on their capacity. The PMs who understand where it flows will build the products that keep their margins. The ones who treat it as someone else’s balance sheet will find out whose it really was.