Benedict Evans published “Ways to think about token pricing” today — a detailed structural argument that AI foundation models will commoditize in the same pattern mobile network carriers did. The analogy is deliberate: mobile operators spent over $200 billion annually in infrastructure capex, carried orders of magnitude more data each year, and delivered miserable shareholder returns. All the value from the smartphone revolution was captured further up the stack — by Apple, Google, and the app developers — not by the carriers who built the infrastructure.
Evans is not the first to make this argument, but he makes it with more structural rigor than most. His conclusion: every current dynamic in AI — supply growth, inference efficiency gains, absence of network effects, capital intensity — points toward foundation model providers becoming low-margin commodity infrastructure. Value will accrue to those who build on top of them. For enterprise AI buyers, this reshapes how to think about vendor selection, architectural decisions, and where to invest organizational capability.
flowchart TD A[Enterprise AI investment] --> B[Bet on model lock-in] A --> C[Build above the commodity layer] B --> D[Tight API coupling<br/>single provider] D --> E[Pricing squeeze<br/>and migration costs] C --> F[Abstraction layer<br/>portable integration] F --> G[Workflow intelligence<br/>domain-specific capability] G --> H[Durable competitive advantage] class E bad class H good class G accent classDef good fill:#163a26,stroke:#44cc77,color:#d7ffe6; classDef bad fill:#3a1620,stroke:#ff5555,color:#ffd9d9; classDef warn fill:#3a2e16,stroke:#ffaa33,color:#ffe9c7; classDef accent fill:#15233b,stroke:#4488ff,color:#dce9ff;
The Rundown: The Telecom Analogy
Evans starts with the supply side. A trillion-plus dollars in data center capex is coming online. Inference efficiency is improving rapidly and continuously — models are becoming 50 to 200 times more efficient. The engineering push on inference cost is relentless across every major lab. There are no obvious network effects protecting any frontier model provider. Capital intensity is high; pricing power is not.
He draws the mobile operator comparison deliberately. Cellular data traffic rose by orders of magnitude over two decades. Carriers invested massively in infrastructure. Prices fell, usage exploded, and the value ended up with the companies that built on top of the carriers — not the carriers themselves. He asks directly whether AI will follow the same pattern.
His conclusion is careful: the situation is unstable, and all variables are in play as the market moves toward equilibrium. But the structural dynamics, as he maps them, point toward commoditization of the foundation model layer and value accumulation by those who build applications, workflows, and organizational capability on top of it.
For Engineers: The Abstraction Layer Is Not Optional
If Evans is right — and the structural case is credible — the engineering bet is clear: don’t hardwire your application to a single provider’s API, and don’t build integrations that make migration expensive.
The practical version of this is an abstraction layer between your application logic and the model APIs — a simple interface where the model is a configuration parameter, not a hardcoded dependency. This is not a complex architecture project. It is a one-time discipline decision that pays out over the next two to three years as model pricing shifts and the competitive landscape changes.
Engineers who build tightly to a single provider’s specific behaviors — taking dependencies on proprietary response formats, vendor-specific tool interfaces, or undocumented behaviors — accumulate switching costs that will be real when the economics move. Engineers who treat the model as interchangeable infrastructure will move easily. The time to build the abstraction is before the pressure to move arrives, not after.
For Business Owners and Operators: Vendor Durability Is Now Part of the Evaluation
Most enterprise AI vendor conversations are structured around capability: which model is most accurate, which platform has the best enterprise features, which vendor has the most integrations. Evans’ argument suggests that evaluating AI vendors on current capability alone is insufficient — durability matters too.
A vendor with genuine pricing power and a defensible moat is a different strategic bet than a vendor facing intense commoditization pressure in twenty-four months. The smart architectural response is to build your AI stack in a way that doesn’t require any single vendor to maintain its current competitive position.
Practically: design for provider portability, invest in the organizational capability to evaluate and migrate, and weight your architecture investments toward the application layer rather than the infrastructure layer. The application layer — where your business logic, domain knowledge, and process intelligence live — is where Evans’ framework says the value will compound. The infrastructure layer is where it will be competed away.
My Take
When I was at First American, the company had assembled arguably the most comprehensive property data infrastructure in the United States at the time: hundreds of millions of property records, fifteen subsidiaries, the largest SQL Server deployment I had encountered, roughly 770 applications across eight business units. They had spent a decade acquiring eighty-plus companies to build that data asset. What they kept learning, in the architectural work I was involved with, is that the infrastructure layer doesn’t create the moat — the applications built on top of it do. Competitors could access similar data. What differentiated First American was the underwriting intelligence, the workflow integrations, the institutional knowledge embedded in the applications. The data was the commodity substrate. The applications were the business.
Evans’ AI commoditization argument maps directly to that structure. The companies winning from AI in three to five years will not be the ones who locked in the best model rates in 2025. They will be the ones who invested in what runs on top of the models: the workflow intelligence, the quality of their evaluation and oversight processes, the domain-specific capability that is genuinely hard to replicate. That is the layer that compounds, because it is the layer that reflects real organizational knowledge rather than infrastructure access.
If your current AI strategy is primarily “we have access to the best models,” the second question worth adding is: what are we building on top of them that only we could build? That is the strategy question Evans’ framework puts in front of every enterprise AI buyer today.
