Is Photon worth it? The Databricks Photon cost break-even, explained

By DBU Watch

Photon, the vectorized query engine in Databricks, makes many workloads faster. On classic compute it also makes every node-hour cost more. So the cost question is simple: does your job speed up enough to cover the higher hourly rate? At current list prices (as of October 16, 2026), the bar for common 8-vCPU instances sits at roughly 1.5x to 1.9x faster. Below that, Photon costs you more.

How Photon is priced

You won’t find a separate Photon line on the price list. According to the Databricks documentation, Photon instance types consume DBUs at a different rate than the same instance on the standard runtime. The price per DBU doesn’t change (AWS Jobs Compute on Premium is $0.15 either way). What changes is how many DBUs each node burns per hour. An AWS m5d.2xlarge on Jobs Compute goes from 1.37 to 3.97 DBU/hour, which is 2.90x. An Azure D8ds v5 goes from 2 to 5 (2.50x). All-Purpose Compute uses a smaller multiplier than Jobs on all three clouds.

The cloud VM bill is untouched. Photon on or off, you pay the same VM rate.

The break-even formula

Photon pays for itself when the job gets faster by at least this factor:

break-even speed-up = (Photon DBU/h x $/DBU + VM $/h) / (standard DBU/h x $/DBU + VM $/h)

Because the VM cost sits on both sides of the formula, the break-even ends up well below the raw DBU multiplier. The bigger the VM share of your bill, the easier the bar is to clear. Some examples per node, using Premium tier and on-demand list prices in us-east-1, Azure East US and GCP us-central1:

Instance choice matters too. Across the 271 AWS instance types with Photon ratings in us-east-1, the Jobs Compute break-even runs from 1.50x to 1.90x, with a median of 1.71x. If you run spot or reserved VMs, your VM cost is lower, so the break-even climbs toward the raw DBU multiplier.

When it tends to pay off, and when it doesn’t

Databricks says Photon helps most on longer-running queries over large datasets, while queries that normally finish in under two seconds see no meaningful gain. Photon also doesn’t support UDFs, RDD or Dataset APIs, or stateful streaming. Those operations fall back to the regular Spark engine, and you keep paying the Photon DBU rate for them with no speed-up.

• Usually worth keeping: heavy SQL and DataFrame ETL, large joins and aggregations, Delta writes and MERGEs.
• Worth testing with Photon off: Python UDF or RDD-heavy jobs, jobs that finish in seconds, and stateful streaming.

The default-setting gotcha

SQL warehouses and serverless compute always run Photon, so there’s nothing to decide there. On classic jobs, all-purpose and pipeline compute, the Databricks UI turns Photon on by default. The Clusters and Jobs APIs only use it when runtime_engine is set to PHOTON. If your teams build clusters by clicking through the UI, you are probably paying Photon rates already, checked or not.

How to measure it

• Pick a job that runs regularly and takes more than a few minutes. Databricks’ own cost optimization guidance recommends checking recurring jobs to see whether Photon makes them cheaper, not just faster.
• Run it several times with Photon on and several times with it off, on the same cluster and data volume. Compare median runtimes.
• Check how much of the job actually ran in Photon. Photon operators show up in orange in the Spark UI. A lot of fallback means you’re paying for little.
• Compare DBUs in the system.billing.usage system table, where product_features.is_photon flags Photon usage. Then add the VM hours from your cloud bill.
• Divide the old runtime by the new one and compare that speed-up with your break-even.

Full break-even numbers for every cloud, plus a calculator: https://dbuwatch.com/guides/is-photon-worth-it/. DBU Watch is a free, independent tracker of Databricks list prices.