Databricks Publishes Analysis of AI Cost Reductions

Patrick Wendell announced that Databricks published an analysis of techniques used to reduce its internal AI inference spend while growing adoption. The post lists several layered changes, including shifting defaults to more efficient models such as open-source GLM, that together produced unit-cost reductions reaching 90 percent in some scenarios. Wendell and colleagues also noted the use of an AI Gateway for centralized routing and analysis, along with configuration adjustments on models like Claude. Replies from other Databricks leaders and engineers echoed the same findings and pointed readers to the company blog post.