Google and the costs of generative AI

One of the biggest obstacles preventing Google from incorporating generative artificial intelligence features into its main product, search — despite having the technology long before OpenAi launched its first products — is the unit cost.

The unit cost of a Google search is extremely low. Although the database on which the company performs searches is huge (because if you are reading this, I assume you didn’t think that searches were performed directly on the web, which is technically impossible if you want to deliver instant results) and instead is done by creating a copy of the entire accessible Internet, the technology that allows searches on this database has been refined again and again by the company throughout its history, with database technologies such as MapReduce, Bigtable and many others. When Google searches show not only the number of results returned, but also the number of seconds it took to get them, which users don’t care about, we’re simply talking about bragging rights, a way for engineers to boast about the level of optimization they are able to achieve.

Thanks to this approach, the unit cost of a conventional search is extremely low, which is essential for a company that performs 320,000 searches per second. The entire search results account is organized on the basis of this cost and its optimization. So when a technology like…