A 100-billion-document index for rent, and retrieval turns into a thing labs buy

Web-scale retrieval has one economic property that decides the rest: the expensive part is scanning, not answering. Styskin puts it as a technical problem, saying that without index structures fine-tuned for a specific task the cost of serving and scanning the whole internet is enormous, and that the work is in narrowing the search space fast ^[4]. Read as a business statement instead, it says a crawl is a large fixed cost looking for more tenants. One index queried by several labs at training time and again at inference is the same corpus billed twice, which is the only way the arithmetic gets tolerable for anyone who is not Google.

The scarcity half of the argument comes from Accel's Zhenya Loginov, who led the investment and says AI players have very few options at web scale, particularly as Google and Microsoft move to close their existing search APIs rather than cannibalise themselves, favouring bundled deals with selected partners ^[6]. That is the load-bearing claim in the whole story, and it is made by the person who wrote the cheque. Keenable's own supporting evidence is similarly indirect: Styskin says he saw Cloudflare data showing AI crawlers taking a growing share of search volume ^[7], and the company will not name the labs and inference providers it says run its API in production ^[5].

Then there is a number the announcement leaves lying around. Twenty-six million dollars against more than 100 billion documents works out to roughly 26 cents per thousand documents already indexed ^[11], and Styskin's only comment on what the index cost was "Don't ask - it is painfully expensive" ^[10]. The round plainly did not pay for the crawl. It is going to headcount: 15 engineers across the U.S. and Europe, doubling by the end of the year, and the stated purpose is a go-to-market motion ^[9], which is roughly 30 people ^[12]. A company that already has the asset and is now hiring sellers is a company that thinks the buyers exist.

What would make that dependency sticky is the next product rather than the current one. WebQueryLanguage is meant to let AI systems assemble an answer from several web sources when no single one contains it ^[8]. An index is substitutable. A query language that agent code is written against is not, and Brave, Exa and a Google busy rebuilding its own search experience are all competing on the substitutable part ^[13].