DynamoDB Pricing Explained: On-Demand vs Provisioned (2026)

Every DynamoDB table runs in one of two capacity modes, chosen per table (and switchable up to once per 24 hours). On-Demand bills per request with no capacity planning required — DynamoDB scales instantly to whatever throughput you send it and charges for exactly that. Provisioned requires you to set a WCU (write capacity unit) and RCU (read capacity unit) ceiling ahead of time, billed hourly whether you use the capacity or not, but at a per-unit rate that's dramatically cheaper than the On-Demand equivalent once utilization is reasonably steady.

Neither mode is universally cheaper — it depends entirely on how predictable and how steady your traffic is. Spiky, unpredictable, or low-volume tables are almost always cheaper On-Demand once you account for the capacity headroom Provisioned mode needs to avoid throttling. Steady, high-volume tables are almost always cheaper Provisioned, especially combined with Auto Scaling or Reserved Capacity. The worked examples further down quantify exactly how large that gap gets.

On-Demand bills in Write Request Units (WRU) and Read Request Units (RRU) — the same underlying unit of throughput as Provisioned's WCU/RCU, just billed per request instead of per hour reserved (AWS DynamoDB on-demand pricing, 2026):

DynamoDB also has an always-on free tier, not a 12-month promotion: 25 GB of storage, 25 WCU/25 RCU of provisioned throughput (or the On-Demand equivalent), and 2.5 million Streams reads per account per Region, every month, with no expiration.

Provisioned capacity is billed hourly per unit reserved, at a per-unit rate far below the equivalent On-Demand cost — the tradeoff is that you set (or Auto Scale) a ceiling ahead of time, and requests beyond it throttle unless Auto Scaling reacts in time (AWS DynamoDB provisioned pricing, 2026):

Auto Scaling adjusts provisioned WCU/RCU within a min/max range you set, targeting a utilization percentage (commonly 70%), at no extra charge beyond the capacity actually provisioned at any given hour. It reacts to sustained changes in traffic, not instantaneous spikes — a workload with genuinely unpredictable, second-to-second bursts is still better served by On-Demand than by Auto Scaling trying to chase it.

For tables with genuinely steady, forecastable throughput, Reserved Capacity discounts the Provisioned hourly rate further in exchange for an upfront commitment (AWS DynamoDB reserved capacity pricing, 2026):

• 1-year term: up to 54% savings versus standard Provisioned rates
• 3-year term: up to 77% savings versus standard Provisioned rates
• Purchased in increments of 100 RCU or 100 WCU, per Region
• Requires an upfront fee plus a discounted hourly rate for the remaining commitment
• Non-refundable and non-transferable — cannot move to another Region or account once purchased

Reserved Capacity only pays off if the reserved units stay genuinely utilized for the full term — unlike Savings Plans for EC2, there's no automatic application across differently-sized workloads, so over-committing on a table whose traffic later drops locks in a cost you can't shed. Model the break-even against Auto Scaling on Provisioned (no commitment, still far cheaper than On-Demand for steady traffic) before committing to a 1- or 3-year term.

Global Tables replicate a table across multiple Regions for low-latency local reads and multi-Region resilience — but every write is billed again, in full, in each additional replica Region:

• Each write replicated to another Region consumes a full-price replicated Write Request Unit (rWRU) in the receiving Region — same $0.625/million (On-Demand) or $0.00065/WCU-hour (Provisioned) rate as a native write there, not a discounted “replication” rate
• Cross-Region data transfer for the replicated write payload: $0.09/GB
• Storage is billed per replica — a 2-Region table pays storage twice, a 3-Region table three times
• Adding a new replica to an existing table triggers a one-time backfill, billed at the table-restore rate of $0.15/GB for the data copied into the new replica

Reads are cheap in Global Tables — each Region only bills for the reads it actually serves locally, with no replication surcharge. The cost driver is write volume: a table with heavy write traffic gets meaningfully more expensive with every additional Region, while a read-heavy, write-light table (the workload Global Tables is usually built for — low-latency reads close to users) scales far more gently.

DynamoDB Accelerator (DAX) is an in-memory read cache that sits in front of a table to cut read latency from milliseconds to microseconds — but unlike DynamoDB itself, DAX is not serverless. It runs as a cluster of provisioned nodes, billed per node-hour regardless of whether it's actually caching anything:

A minimally viable production DAX cluster for high availability uses 3 nodes across availability zones. Three dax.r5.large nodes run ~$589/month (3 × $0.269/hr × 730 hrs) whether the cache hit rate is 95% or 5% — a fixed cost that only makes sense once the read volume and latency requirement genuinely justify it. For spiky or low-read-volume tables, that fixed monthly floor is usually larger than what caching would ever save on RRU spend.

DynamoDB can bulk-export a table to S3 for analytics (Parquet or DynamoDB JSON) or bulk-import from S3 into a new table, both billed per GB rather than per request:

• Export to S3: $0.10 per GB of table data exported — billed on the size of the table data (items plus local secondary indexes) at the export point in time, not the size of the output files in S3. Applies to both full and incremental exports.
• Import from S3: $0.15 per GB of uncompressed source data, including any failed records — no consumed WCU on the destination table during the bulk load.

Both are meaningfully cheaper than reading or writing the equivalent volume through normal RRU/WRU billing at any real scale, which is why export-to-S3 (rather than scanning the table) is the standard pattern for feeding DynamoDB data into Athena, Redshift, or a data lake.

Low, spiky traffic (5M eventually consistent reads + 1M writes/month, 5 GB storage): On-Demand: 2.5M RRU (5M reads × 0.5) at $0.125/million = $0.31, plus 1M WRU at $0.625/million = $0.63, plus $1.25 storage — ~$2.19/month total. Provisioning enough WCU/RCU to safely absorb unpredictable spikes at this volume would cost more in reserved-but-unused capacity than the entire On-Demand bill. On-Demand wins decisively at low or unpredictable volume.

Steady, predictable high volume (sustained 100 WCU + 300 RCU, 24/7, 50 GB storage): Provisioned: 100 WCU × $0.00065 × 730 hrs = $47.45, plus 300 RCU × $0.00013 × 730 hrs = $28.47, plus $12.50 storage — ~$88.42/month. The On-Demand equivalent of that same sustained throughput (100 writes/sec and 300 strongly-consistent reads/sec, continuously, all month) works out to 262.8M WRU ($164.25) plus 788.4M RRU ($98.55) plus $12.50 storage — ~$275.30/month. Provisioned is roughly 3.1x cheaper for identical, steady throughput. Add a 1-year Reserved Capacity commitment (up to 54% off the WCU/RCU portion) and the Provisioned total drops to roughly ~$47/month before amortizing the upfront fee — about 5.8x cheaper than On-Demand for this specific workload.

Multi-Region Global Table (10M writes + 40M eventually consistent reads/month, 20 GB storage, adding one replica Region): Single-Region On-Demand baseline: $6.25 (WRU) + $2.50 (RRU) + $5.00 (storage) = $13.75/month. Adding a second replica Region that serves an equivalent local read volume: +$6.25 in replicated WRU, +$2.50 in local reads, +$5.00 in duplicated storage, plus roughly $0.90 in cross-Region transfer for the replicated write payload (10 GB × $0.09/GB) — ~$28.40/month total, roughly doubling the bill for two Regions. A third replica would add a comparable increment again.

• Leaving a steady, high-volume table on On-Demand. The worked example above shows roughly 3x higher cost versus Provisioned + Auto Scaling for identical, predictable throughput — On-Demand's convenience has a real, quantifiable price at steady volume.
• Over-provisioning WCU/RCU “to be safe” instead of using Auto Scaling. Reserved-but-unused capacity bills the same as used capacity every hour; a wide Auto Scaling min/max range with a reasonable target utilization captures most of the same safety margin without paying for permanent headroom.
• Adding Global Table replicas without pricing the write multiplication. Every additional Region roughly re-bills the full write volume again, plus cross-Region transfer — fine for read-heavy tables, expensive for write-heavy ones.
• Running a DAX cluster sized for peak load 24/7 for a table with spiky or low read traffic. DAX has no serverless tier — a 3-node HA cluster runs a fixed ~$589–plus monthly floor (at dax.r5.large) regardless of cache hit rate.
• Letting DynamoDB traffic route through a NAT Gateway in a private subnet. A free VPC Gateway Endpoint eliminates NAT processing charges for DynamoDB traffic entirely — see our NAT Gateway pricing guide.
• Forgetting PITR and on-demand backups are billed separately from table storage. Continuous PITR runs $0.20/GB-month on top of the $0.25/GB-month Standard storage rate — easy to omit from a cost estimate built only around WCU/RCU or WRU/RRU.