Matthew Kayser
\n\u00a0|\u00a0 Contributor<\/p>\n
For many companies, the move to the cloud promised a future of flexibility and scalability. But as data needs have grown, so too have the bills. Snowflake, one of the most popular cloud data warehouse platforms, often ranks as a company\u2019s second-largest IT expense after its primary cloud provider. For finance chiefs, the challenge isn\u2019t whether to use data warehouses, but how to use them strategically: Are you optimizing for both performance and cost, or just convenience?\u00a0<\/p>\n
Ben Lerner<\/a> knows this problem well. A veteran of Google Search and DeepMind, Lerner co-founded Espresso AI<\/a> to address what he calls the \u201crunaway growth\u201d of cloud costs. \u201cIt\u2019s very common for a team to think they\u2019ve budgeted two years of runway with cloud data warehouse platforms, only to realize they\u2019ve burned through the contract in 14 months,\u201d he explains. \u201cCosts can grow faster than the revenue of the business, and teams are often told to cut the bill in half without any extra headcount. That is an almost impossible ask.\u201d\u00a0<\/p>\n A Different Approach to FinOps<\/strong>\u00a0<\/p>\n The field of cloud cost management, often called FinOps, has grown quickly as organizations look for ways to control spending. Traditional approaches focus on financial engineering such as negotiating discounts for long-term commitments, or on observability tools that track who is running what. Those methods provide visibility, but Lerner argues they leave most of the work to already-stretched engineers.\u00a0<\/p>\n \u201cOther startups in this space are essentially giving you dashboards and graphs,\u201d he says. \u201cThat is valuable, but it also creates more work for the team. Our approach is different. We take the workloads you are already running, and Espresso\u2019s software makes them more efficient. You get the same results, faster, on less hardware.\u201d\u00a0<\/p>\n Espresso\u2019s software sits between a company\u2019s analytics tools and Snowflake. Every query passes through its system, where it is evaluated, rewritten, and scheduled in real time. The company uses large language models combined with reinforcement learning and formal verification. In practice, that can mean running the same queries more efficiently in less time and with lower compute costs.\u00a0<\/p>\n An Air Traffic Controller for Data<\/strong>\u00a0<\/p>\n In addition to optimizing queries, Espresso routes queries between different compute clusters to maximize efficiency. Lerner often reaches for analogies to explain what Espresso does, at one point comparing Espresso\u2019s software to an air traffic controller: \u00a0\u00a0\u201cImagine an airport where every flight had its own dedicated runway all day. It would work, but it would be horribly inefficient,\u201d says Lerner. \u201cInstead, you have controllers directing planes to land and take off in sequence. That is what our software does with queries.\u201d\u00a0<\/p>\n Another metaphor is packing a moving truck. Instead of just throwing things into the truck without thinking about how it all fits \u2013 which leads to wasted space and more trips than necessary \u2013 it\u2019s preferable to pack the truck in a thoughtful manner that maximizes efficiency. Similarly, by managing workloads intelligently, Espresso reduces the \u201cdead space\u201d in how data warehouses allocate compute resources.\u00a0<\/p>\n