Search Costs Favor Enterprise AI
Product manager at Cohere on enterprise AI search infrastructure and deep research agents
This cost dynamic is what separates enterprise AI infrastructure from consumer AI products. Cohere keeps web search costs off its own P&L because North is privately deployed and each customer brings its own Tavily key, so usage is paid account by account. A company like OpenAI or Manus has to absorb search and browsing costs across a giant shared user base, which turns every extra query, page fetch, and multi step research run into a real margin decision.
-
In practice, the difference is seat level versus internet scale. A bank using North may have tens of thousands of employees, but its search bill is isolated to that one customer. A consumer product with hundreds of thousands or millions of monthly actives compounds the same unit cost across everyone at once.
-
Deep research makes the math worse because it is not one lookup. Parallel backed Manus reports can take 10 to 15 minutes and involve repeated searching, browsing, and reading across many sources, so the cost per finished answer is far higher than a simple cited response.
-
This is why vendor choice and pricing matter more for consumer facing AI than for Cohere. The interview describes Parallel as cheaper than Exa and good value, while Cohere is structured around private enterprise deployments that avoid the negative unit economics of serving free or low priced mass market traffic.
Going forward, the winners in agentic search will be the ones that either push search cost down sharply or reserve expensive deep research for high value workflows. That favors enterprise vendors with customer funded usage, and it pressures mass market AI products to bundle search more tightly, build more in house infrastructure, or gate premium research behind paid tiers.