When Does Azure SQL Serverless Beat Provisioned? The Maths

When Does Azure SQL Serverless Beat Provisioned? The Maths

Published by Rushikesh Dhule, LessBill · November 16, 2026 · Prices checked on this date

Azure SQL Database serverless sounds like the obvious choice for a small team: it pauses when nobody is using it and you pay by the second. But a serverless vCore-hour costs about twice as much as a provisioned one. So the real question is simple: how busy does your database have to be before provisioned is cheaper? With today's East US prices, the answer is when your average billed compute is above roughly 48% of the vCores you would otherwise provision.

This guide shows the prices, the formula Azure uses to bill serverless, two worked examples, and the commands to check which side of the line your database is on.

The two prices that decide everything

All prices below are pay-as-you-go, General Purpose tier, standard-series (Gen5) hardware, East US, taken from the Azure Retail Prices API on November 16, 2026. You can see the same numbers on the Azure SQL Database pricing page.

ItemServerlessProvisioned
Compute, per vCore-hour$0.521758 (license included)$0.152217 compute + $0.099966 SQL license = $0.252183
One vCore for a full month (730 h)$380.88$184.09
Billing unitPer second, on what you use (with a floor)Per hour, on what you reserve
When idleCan auto-pause; compute drops to $0Full price, always on
Reservations / Azure Hybrid BenefitNot availableAvailable
Data storage$0.115 per GB-month either way

Divide the two compute prices: $0.521758 ÷ $0.252183 = 2.07. A serverless vCore-hour costs 2.07 times a provisioned one. Storage is the same on both, so it does not change the decision.

How serverless is actually billed

This is where most estimates go wrong. Microsoft's serverless billing documentation says that every second you are charged for the largest of four numbers: your minimum vCores, the vCores you used, your minimum memory divided by 3, and the memory you used divided by 3. Memory is converted at 3 GB per vCore.

Two things follow from that:

  • Memory can bill you even when CPU is quiet. SQL Server keeps data in its cache. A database using 9 GB of memory is billed as 3 vCores, even if its CPU sits at 1 vCore.
  • Idle is not free until the database pauses. While it is online but quiet, you still pay for the minimum vCores (or minimum memory). Only a paused database bills zero compute, and pausing only happens after the auto-pause delay, which is at least 15 minutes on General Purpose (default 60, maximum 7 days), according to the auto-pause documentation.

The break-even rule

Since serverless costs 2.07 times more per vCore-hour, it only wins if you are billed for fewer than 1 ÷ 2.07 = 48.3% of the vCore-hours a provisioned database would charge.

Put another way: take the provisioned size you would need for your peak (say 4 vCores). If your database's average billed vCores across the whole month, including paused time as zero, stays under 48% of that (under 1.93 vCores here), serverless is cheaper. Above it, provisioned is cheaper.

For a database that runs flat out at 4 vCores when active and is paused the rest of the time, the break-even is $736.37 ÷ (4 × $0.521758) = 353 active hours a month, about 11.6 hours a day. Run it more than that and provisioned wins.

Worked example 1: an internal tool used in office hours

An internal admin app. Max 4 vCores, min 1 vCore, auto-pause delay 60 minutes. It is used 10 hours a day on 22 working days, and during those hours it is billed an average of 2.5 vCores (mostly memory). Nights and weekends it is paused.

  • Active time: 22 days × 10 h × 2.5 vCores = 550 vCore-hours
  • The 60-minute wait before each pause, billed at the 1 vCore minimum: 22 × 1 h × 1 = 22 vCore-hours
  • Total: 572 vCore-hours × $0.521758 = $298.45 a month

A provisioned database big enough for the 4-vCore peak would cost 4 × $184.09 = $736.37. Even a 2-vCore provisioned database ($368.19) costs more and would be too small at peak. Serverless saves about $438 a month here. The average billed compute is 572 ÷ 730 = 0.78 vCores, well under the 1.93 break-even.

Worked example 2: a production API that never sleeps

A customer-facing API with traffic around the clock. Same 4 vCore max, but connections never drop to zero, so it never pauses. It is billed an average of 2.5 vCores all month.

  • Serverless: 730 h × 2.5 vCores × $0.521758 = $952.21 a month
  • Provisioned 4 vCores: $736.37 a month

Provisioned is $215.84 a month cheaper, and that is before any reservation or Azure Hybrid Benefit discount, which the billing documentation says do not apply to serverless at all. Average billed compute is 2.5 vCores, above the 1.93 break-even.

Check your own database in five minutes

If the database is already serverless, Azure Monitor has a metric called app_cpu_billed that reports exactly what you were billed for, in vCore-seconds (see the serverless monitoring page). Pull a month of it with the Azure CLI:

# Current settings: compute model, min vCores, auto-pause delay
az sql db show -g my-rg -s my-server -n my-db \
  --query "{sku:currentSku, minCapacity:minCapacity, autoPauseDelay:autoPauseDelay}"

# vCore-seconds billed per day over the last 30 days
DB_ID=$(az sql db show -g my-rg -s my-server -n my-db --query id -o tsv)
az monitor metrics list --resource "$DB_ID" \
  --metric app_cpu_billed --aggregation Total \
  --interval P1D --offset 30d -o table

Add up the daily totals and divide by 3,600 to get vCore-hours. Divide that by 730 for your average billed vCores, then compare it with 48% of the provisioned size you would pick. To switch models, the az sql db reference documents the flags:

# Move to serverless: max 4 vCores, min 1, pause after 60 idle minutes
az sql db update -g my-rg -s my-server -n my-db \
  --edition GeneralPurpose --family Gen5 --capacity 4 \
  --compute-model Serverless --min-capacity 1 --auto-pause-delay 60

# Move back to provisioned 4 vCores
az sql db update -g my-rg -s my-server -n my-db \
  --edition GeneralPurpose --family Gen5 --capacity 4 \
  --compute-model Provisioned

If the database is provisioned today, look at its CPU and memory percentages in the portal (Monitoring → Metrics) for a normal month and estimate billed vCores the same way, using whichever of CPU or memory is higher.

When this advice does not apply

  • Things that stop auto-pause. Active geo-replication, failover groups, long-term backup retention and a DNS alias on the server all prevent pausing. So does any app that keeps a connection open, such as a health check or a connection pool. Without pausing, serverless is just the more expensive per-hour option.
  • Cold starts. Resuming usually takes about a minute, and the first connection fails with error 40613 until it is back. Fine for an internal tool, not for a checkout page.
  • Discounts on provisioned. If you have a reservation or SQL Server licenses for Azure Hybrid Benefit, provisioned gets cheaper and the break-even moves lower. Our guide to Savings Plans and Reserved Instances explains how commitment discounts change this kind of maths.
  • Very small databases. Microsoft's free offer gives up to 10 General Purpose serverless databases per subscription, each with 100,000 vCore-seconds (about 27.8 vCore-hours, worth $14.49 at the price above) and 32 GB of storage a month. A hobby or test database may cost nothing.

Dev and staging databases are usually the easiest serverless wins, for the same reason idle environments waste money in general. We covered that in stopping idle dev and staging environments.

The bottom line

Serverless costs 2.07 times more per vCore-hour than provisioned in East US today, so it only pays off when the database is billed for less than about 48% of the provisioned capacity you would otherwise buy. Office-hours tools, dev databases and spiky low-traffic apps usually land under that line. Busy production databases that never pause usually land over it. Check app_cpu_billed for a month before deciding.

How LessBill can help

LessBill.org shows idle and oversized databases across AWS, Azure and GCP in one view, so you can see which ones to pause, resize or move. It is built for teams who want the savings without a full-time FinOps person.

Sources

Comments

Popular posts from this blog

Stop Paying On-Demand Prices: Savings Plans, Reserved Instances and Spot Explained

7 Zombie Cloud Resources Quietly Draining Your Bill (and How to Kill Them)