// data platforms · 27 October 2026 · 5 min read
What Databricks and Snowflake actually charge you for
Two meters, a premium for convenience, some charges on a separate line, and only one real off-switch.
Databricks and Snowflake both charge you mostly for one thing: computer time. Not the data you store. The time machines spend working on it.
Almost every surprise on either bill comes from three places. You paid extra for convenience without noticing, some charges showed up on a line you weren’t watching, or nothing stopped the meter when you expected it to.
I ran the same data project on both platforms (the full comparison is in Databricks vs Snowflake: what actually happens when you build the same pipeline on both). This piece is just the money part, in plain English.
First, five words you’ll need
- Compute is the computers doing the work when you run a query or a data job. Their meter runs while they’re switched on.
- Storage is where your data sits when nothing is happening. On both platforms it’s cheap next to compute.
- A DBU (Databricks Unit) is how Databricks measures computer time. Its price depends on what kind of computer you’re using.
- A credit is Snowflake’s version of the same idea: one credit buys a slice of computer time.
- Serverless means the platform runs the machines for you. You send the work; you never pick, start, or stop a computer yourself.

Snowflake: one meter, sizes that double
In Snowflake, the machines that run your queries are called a warehouse. Not a building, just a group of computers you switch on and off.
Warehouses come in T-shirt sizes, from X-Small upwards, and each step up roughly doubles the credits you burn per hour. A Large costs about eight times an X-Small for every hour it runs.
The meter is fair in one way and sneaky in another. It bills by the second, and a paused warehouse costs nothing. But every time a warehouse starts or wakes up, you pay for at least a full minute, according to Snowflake’s own explanation of compute costs. Like a taxi that charges a minimum fare every time you get in, however short the trip.
Databricks: the cheap-looking price and the all-inclusive one
Databricks gives you two main ways to run SQL work, and their price tags aren’t measuring the same thing.
With classic compute, you rent the machines from your cloud provider (AWS, Azure or Google) and pay Databricks a fee on top. Two bills. With serverless, Databricks runs the machines and the price includes them. One bill.
Third-party guides list the SQL prices on AWS (US East) at about $0.22 per DBU for classic, $0.55 for “pro” and $0.70 for serverless (one 2026 breakdown; always check Databricks’ own pricing page for your region, since rates move).
Here’s the catch most comparisons skip. The $0.22 doesn’t include the machines. You still pay AWS for those, on a separate bill. The $0.70 does. So serverless looks three times more expensive on the price list, but the real gap is smaller, and how much smaller depends on what your cloud bill adds.
It’s renting a car versus taking a taxi. The rental’s daily rate looks cheap until you add fuel and parking. The taxi fare is everything, all in.
When the taxi wins, and when it doesn’t
Serverless wins when your work comes in bursts. It starts in seconds and you pay nothing while you’re idle. With classic compute you often keep machines running “just in case,” and an idle machine still bills.
Classic wins when something runs all day, every day. At that point you’re paying the convenience premium around the clock, and owning the car (classic machines on committed, discounted pricing) usually comes out cheaper.
So start with serverless. Then check, after a month or two, whether your work has turned into something that never really stops.
Snowflake’s charges on a separate line
Most people budget Snowflake by warehouse size. That misses a set of background helpers that run on Snowflake’s own machines and bill separately:
Snowpipe loads new files automatically as they land. Automatic clustering keeps big tables sorted so queries stay fast. Materialized view maintenance keeps pre-calculated summary tables up to date. Search optimization speeds up lookups.
None of these use your warehouse, so none of them show up in the “what size warehouse do we need?” conversation. They each get their own line on the bill. Like roaming charges on a phone plan: the monthly price is clear, the extras sit on a line nobody reads until it’s large.
The off-switch: only one platform has a real one
This was the most practical difference I found.
On a plain Snowflake trial account, I set a hard spending limit with a single command, and it worked straight away. A resource monitor is Snowflake’s spending cap: you give it a number of credits, and at the thresholds you choose it can warn you, pause your warehouses once running queries finish, or pause them immediately (Snowflake’s resource monitor docs).
One honest catch: resource monitors only cover warehouses. The background helpers above aren’t included, and Snowflake points you to a separate budget feature to watch those.
On Databricks, a budget is an alert, not a cap. It emails you when spending crosses a line, and Databricks’ budget docs say that email can arrive up to 24 hours after the spending happened. It doesn’t stop anything that’s running. Your real protection there is setting machines to shut themselves off quickly when idle.
And on Databricks’ free version, there’s no bill to look at at all. I went looking for the billing data and the table that normally holds it simply doesn’t exist there. Fine for learning. Useless for working out what something would cost you.
What I’d actually do
Start on serverless on either platform, and look again once your workload settles. On Snowflake, set a resource monitor on day one, plus a budget for the background features. On Databricks, set short idle shut-off times and a budget alert, and remember the alert is a smoke detector, not a sprinkler.
Read the line items, not just the total. Years of building payment systems taught me to read a bill like a ledger: the total tells you something is wrong, the lines tell you what.
Both platforms charge you for time. The difference is who’s holding the off-switch.
Muhammad Mairaj is a Senior Backend Engineer with 8 years building payment and data systems. He writes about data engineering, measured. More about him.