The neural optimizer that reads your SQL, argues with your data warehouse, and hands back a smaller bill.
Here is a truth that every data team knows and no CFO enjoys: the cloud data warehouse is a meter that never stops running. Espresso AI's proposition is that a large language model, pointed at that meter, can turn it down without anyone noticing a difference in the output.
Espresso AI - legally Espresso Computing, Inc. - was founded in 2023 by three former Googlers: Ben Lerner, Alex Kouzemtchenko and Juri Ganitkevitch. Between them they came out of Google Search, Google Cloud and DeepMind, which is to say they spent years making other people's systems fast and then decided to make that a company. It is headquartered at 25 Kent Avenue in Brooklyn, which is notable mostly because it is not Silicon Valley.
The product category the company invented for itself is the “neural optimizer.” The unglamorous version: modern companies run enormous SQL workloads on platforms like Snowflake and Databricks, those workloads are provisioned generously and tuned rarely, and the gap between what you pay for and what you use is where the money leaks out. Espresso AI's bet is that this gap is not a management problem or a dashboard problem. It is a math problem, and math problems are what machine learning is for.
What makes the pitch land with the right people is who signed the check. When the company emerged from stealth in May 2024, the $11 million seed round was led by Nat Friedman and Daniel Gross - two investors who understand infrastructure well enough to know that the next act of AI is not another chatbot, it is making the machines underneath everything cost less. FirstMark's Matt Turck led the earlier pre-seed, and dbt Labs founder Tristan Handy is in the cap table too. That is a specific kind of investor, and it tells you the thesis is aimed at people who read query plans for a living.
The thesis, stated plainly by the company, is that “AI will accelerate compute by 1000x.” That is a large number and the sort of claim that would normally invite an eye-roll, except the company is starting somewhere deliberately small and checkable: the Snowflake invoice. It is easier to believe a 1000x future when the present deliverable is a line item your finance team already hates.
“Espresso AI is building the world's first neural optimizer, starting with data warehousing.”
Figures per Espresso AI and press coverage (VentureBeat, FirstMark). Savings vary by workload and should be read as company claims, not guarantees.
Espresso AI is not a single model but a system of agents, each aimed at a different layer of where money leaks in a data warehouse: scaling, routing, and the queries themselves.
Snowflake scales with a one-size-fits-all algorithm. Espresso analyzes each workload independently, learns from your metadata logs, and treats scaling as a constrained optimization problem - minimize cost, don't touch latency. Reported result: 25%+ utilization gains.
The average Snowflake warehouse runs 40-60% idle. This agent - “Kubernetes for Snowflake” - routes queries across existing warehouses in real time using LLM-powered logic borrowed from how hyperscalers run datacenters. It aims to push idle time below 10%.
In closed beta: an LLM that rewrites your SQL for efficiency, like an expert engineer reviewing every query before it runs. The twist that matters - it formally verifies each rewrite is mathematically equivalent to the original, so it can't quietly change your numbers.
Bars illustrate company-reported ranges; actual figures depend on workload.
If you run a data platform, the appeal is that adoption doesn't require a migration. Espresso plugs into an existing Snowflake environment, learns from its metadata, and starts trimming spend - no re-platforming, no six-month project, no rewriting your pipelines. There is a free observability layer and a savings estimator, so you can see the leak before you commit to plugging it.
For the analytics engineer, the promise is subtler: it scales the judgment of your best query-tuner to every query that comes through the door, at three in the morning, without a ticket. For the CFO, it is a rare thing - an AI product whose value shows up directly on a bill you already receive. And for teams on Databricks, the company has extended the same idea into an “agentic lakehouse,” so the approach isn't locked to a single vendor.
“Like having an expert data engineer look at every piece of SQL before it hits your data warehouse.”
Former Googler; leads the company from Brooklyn. Public contact and voice of the founding team.
Systems and performance engineering background out of Google Cloud.
Machine learning and deep-learning research pedigree from Google Search and DeepMind.
The broader team is described as a small, talent-dense group of AI and performance engineers recruited from Google, Apple and MIT - roughly 16 to 19 people at last count. Concentrated, like the drink.
Ben Lerner, Alex Kouzemtchenko and Juri Ganitkevitch leave Google to build a neural optimizer. FirstMark's Matt Turck leads an early pre-seed.
Nat Friedman and Daniel Gross lead the round. VentureBeat covers the launch; the initial product targets Snowflake SQL optimization.
Launches the scheduling agent that routes queries in real time to renovate warehouse utilization.
Extends the neural optimizer beyond Snowflake, signaling the technology is platform-agnostic.
Search links - Espresso AI does not host a central video channel, so these open curated results.