There's a moment in every startup's life cycle when the AWS bill arrives and someone in finance has a quiet panic attack. For companies running on Snowflake or Databricks, that moment often comes earlier—and the numbers tend to be worse.
Espresso AI thinks it has spotted an opening in that gap between sticker shock and actual optimization. The startup emerged from stealth last May with just over $11 million in combined pre-seed and seed funding, making a straightforward pitch: let us use AI to shrink your data warehouse bill by up to 70%, and we'll only charge you if it actually works.
It's the kind of value proposition that sounds almost too convenient. But in a market where both Snowflake and Databricks have crossed $4 billion in annual revenue—much of it extracted from customers who struggle to optimize their own workloads—the thesis isn't as far-fetched as it might seem.
Google Engineers, Venture Capital Believers
The seed round drew Daniel Gross and Nat Friedman as leads, with FirstMark's Matt Turck anchoring the earlier pre-seed. The investor lineup reads like a who's who of data infrastructure: Tasso Argyros, Spencer Kimball, Tristan Handy. People who've built these systems before, in other words, and presumably know where the inefficiencies hide.
Co-founders Ben Lerner, Alex Kouzemtchenko, and Juri Ganitkevitch all spent time at Google—across Search, Cloud, and DeepMind—before decamping to build what they're calling "the first neural compute optimizer." The product, as Lerner describes it, essentially runs your data warehouse "like a team of world-class data engineers working 24/7." Setup, the company claims, requires one SQL command and a config file.
Whether that's marketing poetry or operational reality depends somewhat on which customer you ask.
The Savings Spread
Minerva cut its bill by 61%. Accredible hit 62%. Black Crow logged 57%, Authentic came in at 44%. Then there's Koalafi at 33% and Goldbelly at 19%—still money back in the budget, but nowhere near that headline 70% figure.
The variance matters. It suggests that Espresso AI's algorithms work better for some workload patterns than others, or that certain customers had more low-hanging optimization fruit to begin with. The company doesn't hide these numbers, which is perhaps to its credit. But it does raise the question of how predictable the savings really are for a new customer signing on today.
The pricing model attempts to sidestep that uncertainty entirely. Customers on the "Single-Shot" plan pay 40% of monthly measured savings. The "Double-Shot" plan bills 36% of estimated annual savings upfront—a slight discount for committing, though with the same underlying risk profile. No savings, no fee.
It's clever. It shifts the performance risk entirely to Espresso AI while giving customers a reason to at least try the product. Whether it's sustainable long-term depends on how consistently the startup can deliver.
Kubernetes for Data, Sort Of
In August of this year, the company rolled out what it's calling "Kubernetes for Snowflake"—a dynamic scheduler that routes queries across warehouses based on workload characteristics. The promise is fewer manually configured warehouses, less idle capacity, lower bills.
Two months later came the Databricks expansion: an "agentic lakehouse" with autoscaling, scheduling, and query optimization agents baked in. The timing is worth noting. Both Snowflake and Databricks are racing to embed their own AI-powered optimization features, which raises an uncomfortable question for any third-party optimizer: how long before the platforms just solve this themselves?
Snowflake's AI_FILTER optimization, launched in September, already promises 2-10x speedups and up to 60% token cost reduction for AI queries. That's not quite the same workload coverage as Espresso AI, but the trajectory is clear.
The Crowding Problem
Espresso AI isn't operating in a vacuum. Sundeck pulled in $20 million from NEA and Coatue to chase similar Snowflake query engineering opportunities. Keebo markets "autonomous warehouse optimization." Capital One spun out Slingshot, which claims up to 40% savings on both Snowflake and Databricks.
The category is filling up fast, perhaps faster than the total addressable market is growing. And lurking in the background is the bigger strategic question: do enterprises really want yet another vendor in the stack, or would they rather pay Snowflake or Databricks a bit more for native tools that integrate seamlessly?
Maybe that's why the pay-for-performance model makes sense—not just as a customer acquisition tactic, but as insurance. If the platforms close the optimization gap, Espresso AI's customers walk away without eating a sunk cost.
For now, the savings appear real enough to keep the lights on. Whether that's still true in 18 months, when Snowflake and Databricks have shipped another dozen AI-powered features, is the bet Gross and Friedman are underwriting.
It's a race between third-party innovation and platform consolidation. In enterprise software, the platforms usually win eventually. But "eventually" can leave room for a lucrative exit in the meantime.
