Ecosystem Subsidies Lock Out Standalone Labs
Thinking Machines
The real moat at the frontier model layer is no longer just model quality, it is the ability to hide model costs inside a much bigger product and cloud business. OpenAI can spread GPT investment across ChatGPT, Microsoft distribution, and Azure APIs. Anthropic can pair Claude with AWS infrastructure and Bedrock. Google can push Gemini through Workspace and Android. A standalone lab has to recover those same compute costs from model revenue alone.
-
Distribution is built in for the incumbents. Microsoft sells OpenAI through Azure OpenAI Service, Amazon distributes Claude through Bedrock, and Google keeps expanding Gemini across Workspace and its apps. That means they start with existing sales channels, billing relationships, and daily user touchpoints that an independent lab must buy or partner to access.
-
Subsidy shows up as pricing power and bundle economics. A cloud platform can accept thinner model margins because the real payoff may come from more cloud spend, higher software seat value, or better retention in a core product. That is why frontier model competition looks more like cloud platform competition than pure software licensing.
-
Thinking Machines enters this market with elite talent and a $10B valuation, but without a native cloud, office suite, or mass consumer surface. That pushes it toward a harder path, either building a product with its own daily usage loop, or finding a wedge where buyers will pay for control, customization, or deployment flexibility instead of lowest price.
The next phase favors labs that own both the model and the place where people already work. Standalone model companies will increasingly need to move up the stack into applications, agents, and workflow specific products, where they can earn margin from outcomes instead of competing token by token against subsidized giants.