Rates and break-even

Standard table class, in US East (N. Virginia). Break-even is the average utilization of provisioned capacity at which both modes cost the same.

OperationOn-demandProvisionedBreak-even utilization
WritesPer 1 KB write$0.625 per million$0.00065 per WCU-hour29%
ReadsPer 4 KB strongly consistent read$0.125 per million$0.00013 per RCU-hour29%

Monthly cost for 100 provisioned units at different utilization

100 write capacity units and 100 read capacity units, running 730 hours a month. "Busy" is the share of provisioned capacity actually used; on-demand is billed on that traffic. The cheaper mode is highlighted.

UtilizationWrites, on-demandWrites, provisionedReads, on-demandReads, provisioned
10% busy$16.43Lowest$47.45$3.29Lowest$9.49
25% busy$41.06Lowest$47.45$8.21Lowest$9.49
50% busy$82.13$47.45Lowest$16.43$9.49Lowest
75% busy$123.19$47.45Lowest$24.64$9.49Lowest
100% busy$164.25$47.45Lowest$32.85$9.49Lowest

Which one to pick

Each recommendation links to the page for that option.

Recommended:On-demand
When: A new table, or traffic you cannot forecast
Why: No capacity to set. You pay $0.625 per million writes and $0.125 per million reads, and nothing while idle.
When: Steady or daily-cycle traffic that keeps capacity busy more than about 29% of the time
Why: A write capacity unit costs $0.00065 an hour. At full use, the same writes on-demand cost 3.5 times as much.
Recommended:On-demand
When: Spiky or mostly idle tables, such as development, batch jobs, or event-driven workloads
Why: Average utilization stays under the break-even, and provisioning for the peak would sit idle most of the day.
When: A new product whose traffic will settle into a pattern
Why: Run on-demand while you learn the pattern. When a right-sized provisioned capacity would stay busy above the break-even, move to provisioned.

What to watch for

Gaps and gotchas that the price table does not show.

The break-even assumes perfect sizing

The 29% figure holds only if provisioned capacity matches the load. Real tables keep headroom for peaks and autoscaling reacts with a delay, so average utilization needs to sit comfortably above the break-even before provisioned wins in practice.

Reads and writes break even at the same point

Both modes bill in the same request units, and reads and writes are priced in the same ratio, so each breaks even at 29%. Item size, transactions and eventually consistent reads change how many units a request uses, not where the line sits.

You can switch modes, but not constantly

A table can switch between on-demand and provisioned once every 24 hours, so you can start on-demand and move to provisioned when traffic settles.

Storage, indexes and replication are separate

Neither mode changes storage at $0.25 per GB-month. Global secondary indexes consume capacity of their own in either mode.

Reserved capacity is not in the dataset

Provisioned tables can buy reserved capacity for a further discount. That rate is not in the processed data, so the table above uses standard provisioned prices.

Related comparisons