The 2026 Warehouse Pricing Playbook

Last updated: 25 September 2026. Prices verified against vendor pricing pages on that date. Vendor prices change without notice; always check the source before signing.

Quick answer:

Every cloud warehouse charges compute and storage on different meters, and each vendor picked a different unit for compute. Snowflake sells credits per second. BigQuery sells bytes scanned or slot-hours. Databricks sells DBUs. Redshift sells node-hours or RPU-hours. Understanding the meter is the difference between a $500 bill and a $50,000 one.

This chapter explains what each meter actually measures, where the surprises hide, and roughly what you should expect to pay at each team size.

The 4 pricing meters, plain English

Snowflake: credits per second

Compute runs on virtual warehouses (XS, S, M, L, XL, 2XL, 3XL, 4XL, 5XL, 6XL for Gen1). Each size doubles the compute of the one below and burns credits at 2x the rate: XS = 1 credit per hour, S = 2, M = 4, up to 6XL = 512.

Billing is per second (with a 60-second minimum every time a warehouse resumes). Warehouses auto-suspend after an idle period you set (60 seconds is typical). Credits are priced per edition (Standard, Enterprise, Business Critical), per region, and are typically in the $2 to $4 range for Standard, more for higher editions.

Gen2 warehouses launched in 2024-2025 with faster compute at their own credit rates, capped at 4XL, and became the default in Snowsight where available.

Storage is billed per TB per month on top of compute, at rates similar to raw cloud object storage.

Where the invoice surprises hide:

• A warehouse left running because auto-suspend was set to 60 minutes instead of 60 seconds.
• Multi-cluster warehouses scaling out under load and staying up.
• Snowpipe streaming with high change volume.
• Cortex LLM calls billed per token, easy to forget as a line item.

BigQuery: bytes scanned or slots

Two pricing models. Pick one per project (or mix at the reservation level).

On-demand: billed per TB of data scanned by queries. A query that reads 1 TB of source data costs the same whether it returns 1 row or 1 million. Storage is billed separately per TB per month, with a distinction between logical and physical billing that matters at scale (physical is often 40 to 60% cheaper, but there’s a 14-day lock once you switch).

Capacity (reservations): commit to a number of slots (units of parallelism) and pay per slot-hour. Predictable, and often cheaper if you have steady query volume.

Where the invoice surprises hide:

• SELECT * on partitioned tables that aren’t actually filtered by partition.
• Streaming inserts on high-volume events.
• BI tools running unbounded dashboards on huge fact tables.
• Materialized views recomputing more often than expected.

Databricks: DBUs

A DBU (Databricks Unit) is a unit of processing capability per hour. Different workload types burn DBUs at different rates: SQL Warehouses, All-Purpose Compute (interactive notebooks), Jobs Compute (scheduled), and Model Serving each have their own DBU tier. Serverless SQL is priced separately again.

You also pay the underlying cloud (AWS, Azure, GCP) for the VMs your clusters run on. So a Databricks invoice is Databricks DBUs + cloud VM cost, side by side.

Where the invoice surprises hide:

• All-Purpose clusters left running by data scientists (the interactive tier is the most expensive).
• Undersized clusters that stay up longer to finish the same work.
• Photon-enabled workloads (fast but higher DBU rate).
• Model Serving endpoints kept warm outside working hours.

Amazon Redshift: node-hours or RPU-hours

Two flavours.

Provisioned (RA3): pay per node-hour. Storage is billed separately (RA3 nodes separate storage and compute, unlike the old DC2). Reserved instances give 1-year or 3-year discounts.

Serverless: pay per RPU-hour (Redshift Processing Unit). The cluster scales automatically; you set a base capacity floor and a spend limit.

Where the invoice surprises hide:

• Concurrency Scaling cost when many users hit the cluster at once.
• Data sharing across regions.
• Managed storage for RA3 clusters growing quietly.

Cost by team size, honest ranges

All-in monthly cost for a full analytics stack (warehouse + ingestion + BI). Ranges are for typical usage, not corner cases.

Add ingestion (Fivetran/Airbyte) and BI (Metabase/Looker/Sigma) on top. Full breakdown in the main guide. For per-tool pricing pages, see the reviews for Snowflake, BigQuery, Databricks, Fivetran, and Airbyte.

The 6 hidden costs almost every team hits

1. Data egress out of your cloud. Reading data across regions or clouds costs real money. Keep sources, warehouse and destinations in the same region where possible.
2. Ingestion pricing that scales with change, not volume. Fivetran bills per monthly active row. A bulk update that touches 50M rows upstream ships 50M change events downstream, whether you needed them or not.
3. BI tools running long dashboards. A slow dashboard querying a 10 TB table 200 times a day is a $10k a month problem hiding in your BigQuery invoice.
4. Cortex, Bedrock, Vertex AI calls. LLM-powered queries against the warehouse are convenient and per-token. Track spend from day one.
5. Reserved capacity you don’t use. Committed slots or credits look cheap on paper. If your workload dropped by 40%, so did your effective rate.
6. Storage growth from raw layers. ELT means keeping raw copies of everything, which is cheap until it’s not. Set retention on raw data you’ll never re-query.

Can you negotiate?

Yes, on 2 dimensions: credit price and committed spend.

Credit price: published rates are list. Committed contracts of $50k+ per year typically get 15 to 30% off list, and multi-year commits get more. Getting a lower rate is largely about walking into the conversation with a competitor’s quote (Snowflake vs Databricks, Snowflake vs BigQuery). Reps expect this.

Committed spend: committing to $X of usage over 12 months gets a discount, but committing to more than you’ll actually use costs more than list. The sweet spot is committing to 70 to 80% of your realistic usage and paying on-demand for the rest.

Best months to negotiate: end of quarter (Snowflake, Databricks, Google), especially Q4. AWS is less quarter-driven but has similar patterns around re:Invent (December) and Summit season.

What’s rarely negotiable: storage rates, egress rates, list prices for compute in small-dollar contracts. Auto-renewal clauses are standard; ask for a 60-day exit window if you can.

The one calculation you should run before you commit

For any warehouse decision, calculate cost per active user per month.

Cost per active user = (monthly warehouse spend + BI spend + ingestion spend) / (number of people whose decisions those tools inform)

If it’s under $50, you’re fine. If it’s $50-200, you’re in the normal band and should audit for waste. If it’s over $200 per person and you’re not at enterprise scale, something is expensively wrong. Usually it’s a dashboard that shouldn’t exist, a Fivetran connector nobody uses, or an all-purpose Databricks cluster left running.

Watching for changes

Pricing pages change without notice. Snowflake introduced Gen2 credits quietly in 2024-2025. BigQuery moved to editions (Standard, Enterprise, Enterprise Plus) in 2023, changing how many features you get at each price. Databricks periodically updates DBU rates by workload.

We’re building a monthly pricing changelog for the tools in our directory. Until it ships, the best practice is to snapshot vendor pricing pages when you sign, review them quarterly, and compare against your actual invoice line by line at least twice a year.

Common pricing questions

What’s the cheapest warehouse for a solo founder?

BigQuery on-demand for anything under a few hundred GB of query volume. The first 1 TB scanned per month is on the free tier. Snowflake is a close second and easier to use. If your data fits on a laptop, honestly, MotherDuck or DuckDB on a $5 VM is even cheaper.

Is Snowflake actually more expensive than the others?

Not inherently. Snowflake bills you honestly per second of compute, which makes waste visible. BigQuery bills bytes scanned, which hides waste inside the query. Databricks and Redshift bill nodes or DBUs, which hides waste inside “we needed this cluster running anyway.” Snowflake’s meter surfaces overprovisioning fastest, which is why the horror stories cluster around it.

How do I estimate cost before I’ve built anything?

Take your source data sizes, estimate query volume per day (be honest: BI tools + data science + ML training), and use the vendor’s pricing calculator with the workload described. Then double it. Real workloads always include exploration and mistakes.

Do reserved slots or committed credits ever pay off?

Yes, if your workload is stable and above a real threshold. As a rule of thumb, if your on-demand spend is above $5,000 a month and you can predict 70% of it, committed capacity pays back. Below that, on-demand keeps you optionable.

Where to next

For the market context around these prices, see Data Warehouses in 2026. For the tools that plug into the warehouse and the fees they add on top, see Data Warehouse Tools and ETL Tools. For per-tool reviews with dedicated pricing pages, browse the directory.