Espresso AI CEO Ben Lerner on using ML to cut Snowflake and Databricks SQL bills in the background, and why AI coding tools still fall short on back end work.
Espresso AI cuts Snowflake and Databricks SQL bills by using machine learning models to rewrite queries, route workloads between clusters and resize warehouses in real time, all in the background without customers changing their code. Ben Lerner, co-founder and CEO of Espresso AI, explained how it works on Episode 27 of Built This Week. He also gave a frank view of where AI coding tools still fall short for serious back end work.
Before the interview, Sam demoed a quiz we built for Espresso AI that asks data teams about habits like SELECT * in production, dashboard refresh rates and teams querying the same raw data. You can read about it on our data warehouse waste quiz build page.
Ben liked it, and pointed out the irony. The quiz touches on FinOps: review your budget, figure out what you run and why, and stop refreshing a dashboard ten times a day if someone checks it once a week. He compared it to noticing you haven't watched Netflix in six months and cancelling. Espresso does the opposite. It plugs into the data warehouse and shifts things around in the background without being opinionated about what you run.
His pitch to overworked data teams is a button that cuts the bill in half without anyone touching anything. He says teams can and should do FinOps work, and at scale they will need to, but they can also let Espresso lower the bill for them. Jordan mentioned a GPU cluster we left running that caused an AWS overage. Ben said Espresso can't stop engineers from running up GPU clusters, but it can catch underused capacity someone forgot to shut off.
Watch Ben Lerner explain Espresso AI in Episode 27 (starts at 6:04) · Episode 27 notes and transcript
The goal of Espresso is to use machine learning to optimize compute. So we use ML models to rewrite code and make it faster.
— Ben Lerner, Episode 27
Beyond rewriting code, Ben says Espresso shifts workloads between clusters and makes sure the right machines run at the right time, so customers get the performance they want without paying for unused capacity. For technical listeners, he described it as Kubernetes powered by ML. Today it works with Snowflake and Databricks SQL, with plans to expand to other data warehouses and compute.
Jordan asked whether customers were simply overpaying. Ben pushed back. Everything Espresso does, customers could in principle do themselves. People have tried to make computers run faster since there have been computers, and ML is the next step. It's not that teams are doing something wrong, he says. The technology got better.
Not through the marketplaces, which Ben says are built for apps that add dashboards or features on top of data, not for rewriting and redirecting all of a customer's traffic in real time. Espresso sits on the customer side and integrates in two ways.
Jordan suggested the same idea could apply to GPUs. Ben agreed and said GPU inference is something Espresso plans to look at. Any large compute where you don't know exactly how much you'll need, and where the question is how to pack work onto physical machines, can benefit. He described pairing workloads that would overload a machine if both burst at once, but whose demand isn't correlated, so they can safely share it with an occasional capacity spike. We went deeper on inference costs with Positron in Episode 37.
Ben says the company exists because of the newer generation of models; Espresso started about six months after ChatGPT came out. Transformers understand code far more deeply than earlier models, which is what powers tools like Cursor, Claude Code and Codex. Espresso uses that understanding differently.
Its models read each incoming SQL query and compare it to previous runs. If seven queries are running in parallel on a machine, what happens when you add an eighth? Will it queue, trigger more capacity, or slow the queries already running? Ben calls these really difficult, unsolved problems. Earlier research relied on coarse heuristics, and he says modern models get much better results.
Not yet, in Ben's view. A lot of Espresso's dashboards and throwaway pandas scripts come from prompts to Codex, and for working with an unfamiliar API he calls it transformational. He hasn't visited Stack Overflow in over a year. But the core product is different:
But, like, I can't just, prompt at something and then throw it out into production and expect to run, like, tens of millions of dollars of Snowflake workloads through it.
— Ben Lerner, Episode 27
He still writes a lot of code by hand, with plenty of autocomplete. Jordan's take was that everybody still gets a superpower, like a back end engineer who can now clear front end tasks quickly. Ben agreed:
I would say everyone is probably ten, twenty, 30% more productive, which is huge.
— Ben Lerner, Episode 27
He added that Espresso pays far less for Claude Code than a third of an engineer per engineer. The team has no dedicated front end engineer; dashboard work goes into the on-call rotation. Design is the bigger bottleneck, since customers saving millions want something that looks polished.
Ben used to work on natural language processing at Google, and some of the people and code he worked with eventually became part of Gemini. He says Siri has huge mind share and usage but hasn't gotten substantially better in years, so the upgrade to Gemini is exciting. He doesn't have a strong view on Gemini versus Anthropic or OpenAI, since he thinks the models keep pace with each other. He also noted, as public information, that many people from that part of Google moved to Apple.
Jordan agreed that speech to text on the iPhone lags most LLM apps, and compared Siri's slide to Alexa's. We also discussed Meta building its own massive compute. Ben, whose investor Daniel Gross is involved in that effort, suspects Meta will keep it internal rather than sell it as a cloud service.
Ben goes into more detail on query routing, warehouse resizing and running GPUs in Espresso's office in the early days. Learn more about Ben on his guest page or at Espresso AI, and watch Episode 27 of Built This Week, where the full transcript is also on the page.
