DBUs vs Credits: The Two Pricing Engines
AI data pipeline licensing on the two dominant platforms rests on different units, and conflating them is the first mistake buyers make. Databricks bills per DBU (Databricks Unit), with list rates spanning roughly $0.07 for Jobs Light to $1.40 for SQL Serverless on the Azure Enterprise tier — a 20× spread. Jobs Compute for automated pipelines sits around $0.15–$0.30, All-Purpose interactive clusters at $0.40–$0.55, and SQL serverless near $0.22. The Standard tier has been sunset on AWS and GCP (October 2025) with Azure following by October 2026, so Premium is now the baseline and the Enterprise tier runs about 15–25% higher for equivalent compute.
Snowflake bills per credit: roughly $2 on Standard, $3 on Enterprise and $4 on Business Critical, with storage charged separately at $23–$40 per TB per month and warehouse sizes consuming 1–8+ credits per hour. For pure SQL analytics the two platforms are broadly cost-comparable at scale; for data-engineering pipelines, Databricks Jobs Compute is often materially cheaper. Choosing the wrong engine for the workload is a structural overpayment that no discount fixes — a theme we return to throughout the AI contract negotiation deep dive.
The Dual-Billing Trap
The most dangerous misread on Databricks is treating the DBU rate as the bill. DBU charges cover the platform only — you also pay your cloud provider separately for the underlying VMs, storage and networking, and that infrastructure can add 50–200% on top of the DBU charge. A pipeline quoted at a DBU rate can cost two to three times that figure once compute is counted.
On Databricks, the DBU is half the invoice. Model the cloud infrastructure alongside it — 50–200% on top — or the business case is wrong before the contract is signed.
Snowflake avoids the split by bundling compute and infrastructure into the credit, which makes forecasting cleaner — but it then layers its own additions on top, covered below. Either way, the right comparison is all-in cost per workload, not the headline unit rate, the same total-cost discipline we apply in AI fine-tuning costs and contracts and AI model hosting contracts.
Committed-Use Discount Bands
Both platforms reward annual commitment, with committed-use discounts of roughly 30–50% on upfront contracts. The structure is tiered and worth knowing precisely before you negotiate.
| Annual commit | Databricks discount vs list | Mechanism |
|---|---|---|
| $1M | 18–28% | Committed-use / DBCU prepay |
| $3M | 25–38% | Committed-use / DBCU prepay |
| $10M | 35–48% | Committed-use / DBCU prepay |
Pre-purchasing DBU as Commit Units (DBCU) for one or three years can save up to about 37% versus pay-as-you-go. On Snowflake, $1M+ ARR deals routinely carry custom pricing well below list. But the discount only pays off if the committed volume matches reality: over-buying locks in unused capacity, and under-buying forfeits the discount on overage. The discount band is a starting point for negotiation, not a generous gift — and the leverage maths mirrors the consumption-commit logic in negotiating AI compute costs.
The AI Surcharge Nobody Budgets For
The newest cost line is native AI. Snowflake's Cortex AI services — hosted text completion, classification, summarisation, embeddings and document parsing — bill per token consumed, on top of warehouse credits, and newer or larger models cost meaningfully more. Databricks has expanded the other direction, adding Lakebase (a serverless PostgreSQL offering from its May 2025 Neon acquisition) so it can serve transactional workloads alongside analytics and ML on one platform — broadening the surface area on which DBUs accrue. Both moves mean the AI features that justified the platform also expand the bill, so model token-level AI consumption explicitly rather than assuming it sits inside the credit or DBU you already committed. The data-rights implications of running models over your pipelines are covered in AI training data licensing.
Right-Sizing the Commit
The discount bands only pay off if the committed volume matches actual consumption, and this is where most data-platform deals leak money. Over-buying locks in unused capacity at a discount you never fully draw; under-buying forfeits the discount on the overage, which is billed at on-demand rates well above the prepaid level. Both errors are common precisely because DBU and credit consumption are volatile — a new pipeline, a model rollout or a seasonal spike can shift monthly usage sharply.
The discipline is to forecast before committing, not after. Run a representative measurement period on pay-as-you-go to establish a real consumption baseline, segment it by workload type — Jobs Compute pipelines behave very differently from interactive All-Purpose clusters — and size the commit to the defensible floor of expected usage rather than an optimistic mid-point, leaving headroom to be absorbed by a negotiated true-down. Aim the commit at the band you will reliably consume (for example the $1M tier at 18–28% off on Databricks) rather than stretching to a higher band you cannot fill, since an unfilled $3M commit is more expensive than a fully-used $1M one. Treat the FinOps forecast as the foundation of the negotiation, supported by ongoing tagging and monitoring so the next renewal is sized on evidence.
The Four Clauses That Protect the Commit
Because the commit is the real exposure, four contract terms decide whether a data-platform deal is defensible. First, a price-protection clause that locks DBU or credit rates for the contract term, so a list-price rise cannot quietly erode your negotiated discount. Second, a true-down right allowing you to reduce the annual commit by up to about 20% at each anniversary, so a usage misforecast is not punished for the full term.
Third, termination for convenience after year one with a pro-rated refund of unused commit — your exit, not just theirs. Fourth, favourable overage terms, because consumption above a prepaid threshold is billed at on-demand rates that are significantly higher than prepaid; negotiate a blended or capped overage rate before you sign, not after you breach. For the full clause set, work through the AI Procurement Checklist and the AI Contract Red Flags brief, benchmark the host clouds via the AWS and Microsoft hubs, and request a confidential briefing before committing to a multi-year DBU or credit spend.