Redistributed

You don't run GPT-4. You call it. This architectural choice — AI delivered as a service through APIs — shapes who can build what, and who captures the value.

The Model

Instead of downloading and running a model yourself, you send requests to a server controlled by the model provider. Your prompt goes in, their response comes out. You pay per token.

This is how most developers interact with frontier AI:

OpenAI's API
Anthropic's API
Google's Vertex AI
Cohere, Mistral, and others

You get access to models you couldn't afford to train or run yourself. The provider gets control over access, pricing, and the relationship with your users.

What You Get

Simplicity. No infrastructure to manage. No GPUs to provision. Make an API call, get a response. The provider handles scaling, uptime, and updates.

Access to frontier capabilities. The best models require billions of dollars to train and specialized hardware to run. APIs let you use them for pennies per request.

Automatic improvements. When the provider updates the model, you get the improvements without changing your code (in theory — breaking changes happen).

Global reach. The provider handles serving infrastructure worldwide. You don't need data centers on every continent.

Control. The model runs on their servers, under their terms. They can change pricing, deprecate endpoints, or modify model behavior without your consent. If they decide to rate-limit you, you're rate-limited.

Privacy. Your prompts and responses traverse their systems. Enterprise agreements may limit data use, but the data flows through their infrastructure. For sensitive applications, this may be a non-starter.

Reliability guarantees. When the API goes down, your application goes down. You're dependent on their uptime, their capacity planning, their incident response.

Cost predictability. Per-token pricing means costs scale with usage in ways that can be hard to forecast. A feature that goes viral can generate surprising bills.

Differentiation. If your product is built on the same API as your competitor's, differentiation is hard. You're both wrapping the same capability.

The Power Dynamics

The API model creates a specific kind of dependency. Vipra and Korinek document how foundation model markets tend toward concentration, making these power dynamics structural rather than incidental (Vipra & Korinek, 2023):

The provider knows everything. They see every request, every use case, every edge case where their model fails. This information is enormously valuable for improving the model and identifying market opportunities.

You know only your slice. You see your application's behavior, but not how the model performs across other use cases. You can't benchmark against competitors' prompts.

Switching costs are real. Each API has quirks — prompting strategies that work on one model may not transfer to another. The more you optimize for one provider, the more locked in you become.

Platform risk is permanent. Build a business on someone else's API, and they can become your competitor. OpenAI launched ChatGPT Enterprise, competing with startups built on OpenAI's API.

Alternatives

Open-weight models — Llama, Mistral, and others can be self-hosted. You lose frontier capabilities but gain control. See Open Source vs. Open Weights.

Fine-tuned smaller models — For specific tasks, a smaller model fine-tuned on your data may match API capabilities at lower cost. Requires ML expertise.

Hybrid approaches — Use APIs for prototyping and demanding tasks, but move to self-hosted models for high-volume or sensitive applications.

The Infrastructure Question

APIs are infrastructure owned by a few companies. As Nick Srnicek argues in Platform Capitalism, platforms create new forms of economic power through control of infrastructure and data (Srnicek, 2017). Using them means participating in an economy where access, pricing, and capabilities are determined by private actors pursuing private interests.

This isn't inherently bad — commercial infrastructure enables all kinds of valuable building. APIs have dramatically lowered the barrier to incorporating AI into products: a developer today can access frontier capabilities with a few lines of code and a modest budget, which was unimaginable five years ago. That accessibility is real and valuable, even as the structural dynamics deserve scrutiny. But it's worth understanding the terms.

When you build on an API, you're not just choosing a technology. You're choosing a relationship, with all the power asymmetries that implies.

Sources

Srnicek, Nick. Platform Capitalism. Polity, 2017. Publisher
Andreessen Horowitz. "Who Owns the Generative AI Platform?" 2023. a16z
Benaich, Nathan & Hogarth, Ian. "State of AI Report." 2024. State of AI
Vipra, Jai & Korinek, Anton. "Market Concentration Implications of Foundation Models." Brookings, 2023. Brookings

General

The API Economy