Starburst recorded three-to-seven-times faster SQL query performance on Nvidia GPUs versus x86 CPUs, CEO Justin Borgman disclosed in a briefing to Chip Dispatch this month.

The speedup depends on the query. Larger, complex workloads hit the seven-times ceiling. Everyday jobs cluster around three to four times, Borgman said.

How did the partnership begin?

Nvidia approached Starburst earlier this year to benchmark its Vera processor. The Trino engine ran faster on Vera than on Intel silicon. Nvidia featured those numbers in a Jensen Huang keynote.

That began the working relationship, Borgman said. Nvidia then asked Starburst to test its cuDF framework. Starburst agreed. A first release of the GPU-capable engine ships now.

What does cuDF actually do for customers?

cuDF is an open-source piece of Nvidia's CUDA-X data processing toolkit. It runs on Nvidia's CUDA primitives. It accelerates structured-data jobs in parallel across the GPU. Starburst layered its Trino distributed SQL engine on top, with GPU support added this year.

Borgman named the target use case. Large banks bought GPUs for AI training. Utilization sits around sixty percent. They could lift it to ninety percent overnight, when jobs are idle. The silicon is paid for.

"We now also support GPUs in addition to CPUs," Borgman said. "You can use existing GPUs that you bought… leverage those GPUs as well."

The start of the GPU work dates to the beginning of this year, Borgman stated. Two scheduling modes are available. Customers can assign fixed night-time slots, or Starburst can measure live utilization and slot jobs in dynamically.

How does this change Starburst's position against Databricks and Snowflake?

Borgman labeled the relationship with the two cloud giants "classic co-opetition." Customers often operate Starburst alongside Databricks, Snowflake, Teradata, IBM and Oracle simultaneously. The engine queries all of them, Borgman said.

The structural contrast is deployment. Databricks and Snowflake are cloud-only. Starburst runs on-premises and in hybrid mode. Regulated banks favor the latter, per Borgman.

Cost flows differently too. Cloud platforms carry processing cost inside their margins. Starburst does not. Customers pick compute. They can pick Nvidia silicon.

"Databricks and Snowflake, as you know, they're cloud-hosted offerings… carry the cost of processing within their cost structure," Borgman said. The pricing pressure on hyperscalers, in his view, points them away from Nvidia silicon.

What are the caveats?

GPUs cost more than CPUs per unit of silicon. Borgman acknowledged this directly. Idle capacity changes the math. Banks are not buying new hardware for analytics, only reusing what they already own.

The seven-times figure is the ceiling, not the mean. Smaller, simpler queries sit at three to four times, per Borgman's briefing. Whether the economics hold at enterprise scale is untested.

Dell already named Starburst in its May AI Factory update. The platform claims six-times faster SQL analytics on Nvidia Blackwell GPUs, "with support designed for future Vera systems." That six-times figure is Dell's, attributed in Dell's AI Factory material, not Starburst's own benchmark.

Nvidia is "a $5 trillion company," Borgman added. "TAM expansion" is how he frames the cuDF engagement on the Nvidia side, though Nvidia did not confirm that figure on the record in this briefing.

What should we watch next?

Vera and Rubin hardware availability will be the next test. More banks running overnight analytics on idle Nvidia silicon will tell whether Starburst's on-prem margin story holds against cloud-only rivals, and whether the TAM Borgman keeps returning to materializes.