On demand compute sprawl rather than storage drove the bill, and ad hoc query spend ran 30 to 60 percent above a slot reservation for the same workload
The pricing is published openly, which is exactly why the edition choice is the most common error. Nobody negotiates a number they can already read.
Prepared by Redress Compliance · August 18, 2026 · Google Cloud engagements in banking. 15 to 25 engagements reviewed, 2024 to 2025.
Executive summary
Ad hoc on demand query spend ran 30 to 60 percent above a slot reservation for the same workload. The overrun is a routing decision rather than a rate.
The top edition was bought where a lower one would have served, a 25 to 45 percent premium. Edition choice was the most common error in the estates reviewed.
Cold partitions stayed in active storage, costing 2x the long term rate. The long term discount applies automatically at 90 days and only to data nobody has touched.
The wrong order is to accept the proposed slot count and reverse engineer the workload to fit it. Model both routes first, then negotiate the commitment and the edition as one thing.
What is actually being priced?
Three axes. Compute as slots or per terabyte scanned, storage as logical or physical bytes, and editions as feature tiers. The slot model dominates large workloads, and the rates are published on the BigQuery pricing page.
| Route | Unit | Best fit | Typical banking workload |
|---|---|---|---|
| On demand | $6.25 per terabyte scanned | Low and variable volumes | Ad hoc reporting and sandboxes |
| Standard slots | $0.04 per slot per hour | Production workloads with a predictable shape | Daily pipelines and recurring dashboards |
| Enterprise slots | $0.06 per slot per hour | Workloads needing customer managed keys and network controls | Regulatory reporting and fraud analytics |
| Enterprise Plus slots | $0.10 per slot per hour | Multi region with advanced security | Cross border data and residency bound workloads |
The break even sits around $30k a month
Roughly 5,000 terabytes scanned. Below it on demand is reasonable, above it a reservation wins, and a one year commitment carries 20 percent against 40 percent on three years.
Where does the storage bill go wrong?
In the tier rather than the rate. Active storage bills at $0.02 per gigabyte per month on logical bytes and $0.04 on physical, and the long term tier halves both after 90 days without modification.
Cold partitions left in active storage therefore cost twice the long term rate, which is the cheapest correction available on the whole bill and the one nobody schedules.
Logical bytes count uncompressed data and physical bytes count what is actually stored. Physical billing can be cheaper for compressed columnar formats, which is a modelling exercise rather than a negotiation.
The BigQuery cost governance brief
Commit the baseline, meter the variable top, and stop paying on demand rates for predictable work.
Get the brief →What 15 to 25 banking engagements showed
Across roughly 15 to 25 Google Cloud engagements in banking reviewed between 2024 and 2025, on demand compute sprawl rather than storage drove the bill. Three patterns recur.
- On demand overruns: ad hoc query spend ran 30 to 60 percent above a slot reservation for the same workload.
- Edition mismatch: the top edition was bought where a lower one would have served, a 25 to 45 percent premium.
- Storage left hot: cold partitions stayed in active storage, costing 2x the long term rate.
The pricing is published openly and the edition choice was still the most common error. Transparency does not help a buyer who never modelled the alternative.
- Every risky clause flagged with the verbatim quote and page anchor
- Entitlements, caps and protections verified across your whole contract portfolio
- Paste ready replacement language and an evidence trail for the response
Which edition does a regulated estate actually need?
The one whose controls the regulator requires, and no higher. Customer managed encryption keys and network service controls sit in the middle edition, so the entry tier is genuinely unsuitable for regulated data.
The top edition adds multi region and advanced security. Where the workload is single region, buying it is a 25 to 45 percent premium against a feature set nobody switched on.
Materialized views are a cost lever inside compute
They absorb peak queries without changing the commercial terms at all, which makes them the one lever available between renewals rather than at one.
Watch the briefing · 4:10The Cloud Bill and the AI BillHow commitments stack on private rates, and the clauses that decide the contract.
What is the right order of operations?
Model the workload against both compute routes, then negotiate the slot commitment and the edition choice as a single committed use discount. One conversation, two decisions, both made from your own telemetry.
Two decisions, one conversation
The wrong order is to accept the proposed slot count and reverse engineer the workload to fit it. That produces a commitment shaped to the seller's forecast and an edition chosen for the seller's comfort.
The capacity model and the tier boundaries are documented in the reservations documentation and the editions overview.
Reservation sharing releases idle slots to other reservations, which reduces stranded capacity inside a commitment you have already made. The wider commitment mechanics sit in the committed use discount guide, the commitment negotiation guide and the leverage framework.
The governance layer that keeps the variable top honest sits in the BigQuery cost governance guide and the cloud financial operations playbook, alongside the wider Google negotiation guide.
What the engagements measured, 2024 to 2025
Two cuts of the engagement file, both on the compute side of the bill.
For the same workload, where ad hoc querying was never moved onto committed capacity.
Where the top tier was bought and the controls that justify it were never required by the workload.
Both are modelling failures rather than pricing ones, which is why the published rate card did nothing to prevent either.
Your first five moves
- Model the workload against both compute routes before anybody proposes a slot count, because the reservation beat on demand by 30 to 60 percent on the same work.
- Check the break even honestly at around $30k a month, roughly 5,000 terabytes scanned, rather than assuming reservations always win.
- Choose the edition on the controls the regulator requires, since the top tier carried a 25 to 45 percent premium wherever its features went unused.
- Move cold partitions out of active storage, which is a 2x rate correction that the 90 day long term discount applies automatically once they are left alone.
- Negotiate the slot commitment and the edition as one committed use discount. The Google practice builds the model before the proposal arrives, and the commitment brief carries the levers.
Frequently asked questions
How does BigQuery price?
On three axes: compute as slots or per terabyte scanned, storage as logical or physical bytes, and editions as feature tiers. The slot model dominates large workloads.
When do reservations beat on demand?
Above roughly $30k a month, which is about 5,000 terabytes scanned. In the engagements reviewed, ad hoc on demand spend ran 30 to 60 percent above a reservation for the same workload.
What do the commitments buy?
A one year commitment carries 20 percent and three years carries 40 percent on slots. Reservation sharing releases idle slots to other reservations, which reduces stranded capacity.
Which edition does a bank need?
The one carrying the controls the regulator requires. Customer managed keys and network service controls sit in the middle tier, so the entry edition is unsuitable for regulated data.
What does the top edition cost?
A 25 to 45 percent premium where a lower edition would have served. Edition mismatch was the most common error across the estates reviewed.
What is the difference between logical and physical billing?
Logical bytes count uncompressed data and physical bytes count what is actually stored. Physical billing can be cheaper for compressed columnar formats, which is worth modelling.
How does the long term storage discount work?
It applies automatically at 90 days on data that has not been modified, halving the rate. Cold partitions left in active storage therefore pay twice what they need to.
What drove the bill, compute or storage?
Compute, in every engagement reviewed. On demand sprawl rather than storage volume was the cost driver, which is the opposite of where most reviews start looking.
Are materialized views a commercial lever?
No, they are a technical one that lowers the bill without touching the contract. They absorb peak queries, which makes them usable between renewals rather than only at one.
What is the wrong way to run this?
Accepting the proposed slot count and reverse engineering the workload to fit it. That produces a commitment shaped by the seller's forecast rather than by your telemetry.
