DeepSeek at risk of disintermediation

The real power in AI is shifting to the interface that picks the model, not the lab that trained it. DeepSeek made itself easy to adopt by shipping open weights and OpenAI compatible endpoints, but those same choices let clouds, gateways, and app vendors swap DeepSeek in and out behind their own menus, billing, and workflows. That turns DeepSeek into the engine under someone else’s dashboard unless it builds stronger direct distribution.1 2 3 4 5

• In practice, enterprise buyers often never touch DeepSeek directly. Hebbia used Fireworks to add DeepSeek through the same API surface it used for other models, then sold customers on model choice inside Hebbia’s own product. Fireworks was abstracted away, and DeepSeek was even more abstracted away.2
• Broker layers are already capturing the traffic. Chinese open source labs grew to more than 50 percent of OpenRouter usage by mid 2026, with DeepSeek alone driving about 3T tokens per day there. OpenRouter charges the developer, owns routing logic, and decides when a cheaper or better model gets the call.3 4
• The main counterexample is Mistral, which responded to similar open model pressure by moving closer to the customer, selling private deployments, tooling, and sovereign compute as one package. That is harder to swap out than a bare model endpoint, because the customer is buying an operating setup, not just tokens.5 6

From here, model labs that win will look less like raw API vendors and more like productized infrastructure companies. If DeepSeek keeps feeding cloud catalogs and routing layers without owning workflow, billing, and security primitives itself, more of the margin and almost all of the customer loyalty will accumulate above the model layer.1 3 5