What Spark accelerators cost: what each vendor publishes about price

These are the terms as published on the pages we reviewed on 27 September 2026. They change, so check the vendor or project page before you budget.

Three models cover the seven products. The first is no license fee: Apache Gluten, Apache DataFusion Comet and Apache Auron are Apache License 2.0 projects that run on the standard CPU instances you already pay for. The second is a change of hardware: the RAPIDS Accelerator plugin is also free, but it runs on NVIDIA GPU instances, so the cost question becomes GPU instance price against runtime saved, which you can model from public cloud prices. The third is a commercial product: DualBird and Flarion are licensed commercially and do not publish prices on the pages we reviewed (Flarion is listed on AWS Marketplace), and Photon is part of the Databricks Runtime, billed through Databricks Units.

A plugin with no license fee still costs compute and people. The open-source plugins need off-heap memory sized, a shuffle manager switched and spill limits set, and they are supported through community channels. Put a number on the engineering hours for the trial, the first months in production and each Spark upgrade, and add it to the cost of running the job. If the compute saving is smaller than the cost of those hours, the free option is the expensive one.

When an accelerator needs different instances, as the RAPIDS Accelerator needs GPU instances and DualBird runs on Amazon EC2 F2 instances, compare cost per run rather than hourly price. The break-even speedup is the new hourly cluster cost divided by the old one: a cluster that costs twice as much per hour has to finish in less than half the time to save money. Include every fee in the hourly figure, including platform fees and any accelerator fee. The cost per job guide has a worked example.

Databricks publishes its DBU rates, and its documentation says Photon compute consumes DBUs at a different rate from non-Photon compute. Whether Photon lowers a bill depends on whether a job finishes enough faster to offset that rate. Databricks also documents where Photon does not help, such as queries that run in under two seconds and code that relies on UDFs. Model it per job, with Photon on and off, on the same cluster.

• The unit of pricing: per node, per core hour, per job, per cluster or a flat fee
• Whether the fee includes support, and at what response times
• Any minimum term or minimum spend
• What the trial costs, and whether trial terms carry into the contract
• How the price changes as data volume or cluster count grows

Get the quote in writing before a trial ends, so the pass or fail decision in your proof of concept is made on cost per run, not on speed alone.

Cost model transparency is one of seven criteria and carries 10% of the weight on the method page. It measures whether a buyer can work out the cost before a sales call, not whether a product is cheap. If a public price matters more to your procurement process, raise that weight in the calculator.