Search as a Service Viability

This points to a likely squeeze on standalone AI search vendors, because the same companies that already sell models also control the compute, distribution, and developer relationships needed to bundle search into a broader API stack. That matters because search for agents is usually bought as plumbing, not as a destination product. When quality is close across vendors, the winner can be the one that makes search cheap, fast, and easy to plug into an existing model workflow.

• In practice, search as a service already works as a paid infrastructure layer. Ecosia routes 30 to 40 percent of queries, about 500,000 a day, to Exa and pays about $300,000 per month on a query based contract. The value is not direct revenue on each query, but better retention and a faster product launch.
  1 sacra
• The opening for model companies is straightforward. They already run large scale inference, already serve developers through APIs, and already absorb search costs inside broader products. A Cohere product leader argues the economics could make sense for a player like OpenAI, because replacing an external vendor could save millions once usage reaches large consumer scale.
  1 sacra 3 sacra
• Standalone vendors still have room because buyers often want a focused supplier that will tune latency, pricing, and integration details around a specific workflow. Ecosia chose Exa over Parallel partly for collaboration and custom pricing, and Cohere uses Tavily rather than building in house because search is useful plumbing, not its core product.
  1 sacra 3 sacra

Over time, search is likely to split in two. General web retrieval will get bundled into foundation model platforms, while independents that survive will move upmarket into higher value layers, such as deeper research workflows, domain specific data streams, and custom enterprise integrations that are harder to commoditize.