DynamoDB On-Demand vs Provisioned: Where the Break-Even Sits

Published by Rushikesh Dhule, LessBill · November 20, 2026 · Prices checked on this date
Every DynamoDB table has to pick a capacity mode: on-demand (pay per request) or provisioned (pay per hour for capacity you reserve). The short answer: provisioned is cheaper once your table uses more than about 29% of the capacity you provision, averaged over the month. Below that, on-demand wins. This guide shows where that number comes from, three worked examples, and the commands to measure your own tables and switch.
All prices below are for the DynamoDB Standard table class in US East (N. Virginia), taken from the official DynamoDB pricing page on November 20, 2026. Other regions cost more, but the ratio between the two modes is what matters, and it is broadly similar.
The two prices you are comparing
On-demand charges per request unit. A write request unit (WRU) covers a write of up to 1 KB; a read request unit (RRU) covers a strongly consistent read of up to 4 KB, and an eventually consistent read uses half a unit. Provisioned charges per capacity unit per hour, and one capacity unit means one such operation per second, whether you use it or not.
| Item (us-east-1, Standard) | On-demand | Provisioned |
|---|---|---|
| Writes | $0.625 per million WRU | $0.00065 per WCU-hour |
| Reads | $0.125 per million RRU | $0.00013 per RCU-hour |
| Free tier | None for requests | 25 WCU + 25 RCU per region |
| Commitment discounts | Database Savings Plans, up to 18% | Reserved capacity, up to 54% (1 yr) / 77% (3 yr); or Savings Plans, up to 12% |
The break-even maths
Take one write capacity unit for one hour. If you fully use it, it handles 3,600 writes (one per second). The same 3,600 writes on-demand cost:
3,600 × $0.625 / 1,000,000 = $0.00225
Provisioned cost for that hour = $0.00065
Break-even utilization = 0.00065 / 0.00225 = 28.9%
Reads work out the same: 3,600 × $0.125 / 1,000,000 = $0.00045 against $0.00013 per RCU-hour, which is also 28.9%. So the rule is one number for both:
Decision rule: if (consumed capacity ÷ provisioned capacity), averaged across the whole month, is above roughly 29%, provisioned is cheaper. If it is below, on-demand is cheaper. On-demand costs about 3.5 times as much per unit of work you actually do; provisioned costs you for the idle seconds.
Worked example 1: a steady SaaS API table
An orders table handles an average of 200 writes per second and 1,000 eventually consistent reads per second, all items under 1 KB. Each eventually consistent read uses 0.5 units, so that is 500 read units per second. A month is 730 hours.
On-demand:
Writes: 200 × 3,600 × 730 = 525.6M WRU × $0.625/M = $328.50
Reads: 500 × 3,600 × 730 = 1,314M RRU × $0.125/M = $164.25
Total = $492.75
Provisioned with auto scaling at the default 70% target: auto scaling keeps provisioned capacity at roughly consumed ÷ 0.7.
WCU: 200 / 0.7 = 286 × 730 h × $0.00065 = $135.57
RCU: 500 / 0.7 = 714 × 730 h × $0.00013 = $67.79
Total = $203.36
That is about 59% less. The real figure will be a bit higher, because auto scaling reacts over minutes and keeps headroom after a spike, but the gap is wide.
Provisioned with no auto scaling, sized for peak: say the peak is 600 writes and 1,500 read units per second, and you set those as fixed values. Cost: 600 × 730 × $0.00065 = $284.70 plus 1,500 × 730 × $0.00013 = $142.35, total $427.05. Average utilization is now 200 ÷ 600 = 33%, just above break-even, so you only save about 13%. Sizing for peak eats most of the saving.
Worked example 2: a nightly batch job
A table receives 5,000 writes per second for one hour a night and almost nothing for the other 23 hours. On-demand: 5,000 × 3,600 × 30 nights = 540M WRU × $0.625/M = $337.50. Provisioned at a flat 5,000 WCU: 5,000 × 730 × $0.00065 = $2,372.50. Utilization is about 4%, far below 29%, so on-demand is seven times cheaper. Auto scaling helps a little, but it cannot add thousands of units the second the job starts, so you either throttle or pre-provision.
Worked example 3: what discounts do to the line
Commitments shift the break-even. Using the maximum discounts on the pricing page:
- Database Savings Plan on on-demand (up to 18% off): on-demand becomes $0.001845 per WCU-hour equivalent, and break-even moves up to about 35%.
- Savings Plan covering both modes (18% on-demand, 12% provisioned): about 31%.
- One-year reserved capacity on provisioned (up to 54% off): about 13%. Reserved capacity is sold in blocks of 100 units, only for Standard tables in a single region, and you pay for it even when unused.
These are "up to" rates. Check the actual effective rate in your Billing console before you buy.
Measure your own utilization
CloudWatch publishes consumed capacity for every table. Pull a week of hourly sums:
aws cloudwatch get-metric-statistics \
--namespace AWS/DynamoDB \
--metric-name ConsumedWriteCapacityUnits \
--dimensions Name=TableName,Value=orders \
--start-time 2026-11-13T00:00:00Z \
--end-time 2026-11-20T00:00:00Z \
--period 3600 --statistics Sum
Divide each hourly Sum by 3,600 to get average units per second for that hour. Repeat with ConsumedReadCapacityUnits, and for provisioned tables compare against ProvisionedWriteCapacityUnits. If you are on on-demand, compare the weekly average to the peak hour: if the average is above about 29% of what you would need to provision, test provisioned.
Switching modes safely
# provisioned -> on-demand
aws dynamodb update-table --table-name orders \
--billing-mode PAY_PER_REQUEST
# on-demand -> provisioned
aws dynamodb update-table --table-name orders \
--billing-mode PROVISIONED \
--provisioned-throughput ReadCapacityUnits=720,WriteCapacityUnits=290
Per the AWS guidance on switching capacity modes, you can move a table from provisioned to on-demand up to four times in a rolling 24 hours, and back to provisioned at any time. If the table has global secondary indexes, give them capacity too when moving to provisioned, and set up auto scaling straight after. Switching in the console removes auto scaling settings; the CLI keeps them.
If you stay on on-demand, you can cap runaway spend with a maximum throughput per table or index; requests above it are throttled rather than billed. The on-demand mode documentation also notes that a table instantly handles up to double its previous peak, and traffic that more than doubles within 30 minutes can be throttled.
When this advice does not apply
- Small tables. Below a few dollars a month, the engineering time to tune capacity costs more than the saving. The provisioned free tier (25 WCU and 25 RCU, worth about $14 a month) is only a reason to pick provisioned for hobby tables.
- Storage-heavy tables. If storage is most of the bill ($0.25 per GB-month on Standard), the bigger lever is the Standard-IA table class at $0.10 per GB-month, not the capacity mode.
- Global tables. Replicated writes are billed in every region, so do the maths per replica. We covered the same trade-off for SQL in when Azure SQL serverless beats provisioned.
- Before buying commitments. Get the mode right first, then commit. Our guide to Savings Plans, Reserved Instances and Spot explains how to size a commitment without overbuying.
The bottom line
Look at one number per table: average consumed capacity divided by what you would provision. Above about 29%, provisioned with auto scaling usually saves 40-60%. Below it, or for spiky and batch tables, stay on on-demand and set a maximum throughput as a spend guardrail. Re-check every quarter, because traffic shapes change.
How LessBill can help
LessBill.org watches usage across AWS, Azure and GCP and flags resources priced on the wrong model. It points out tables like the ones above so you can decide with real numbers.
Comments
Post a Comment