Amazon MSK has two pricing models. MSK Provisioned bills per broker-hour plus storage and is cheapest for steady, high-throughput streams where brokers stay busy. MSK Serverless bills per cluster-hour, partition-hour, and gigabyte of throughput, and is cheapest for bursty, low-average, or unpredictable workloads with no idle brokers to pay for. The decision is utilization: if your throughput is high and flat, provision and keep the brokers full; if it is spiky or new and uncertain, go serverless. Either way, the largest hidden costs are over-long retention, over-sized brokers, and idle capacity.
Last updated: June 2026 · Written by Fredrik Filipsson and reviewed by Morten Andersen · See, Cut, Lock, Run method
Amazon MSK (Managed Streaming for Apache Kafka) runs Kafka without the operational burden, and it offers two pricing models that suit very different traffic. MSK Provisioned is the classic model: you choose broker instance types and counts and pay for them by the hour, busy or not. MSK Serverless removes broker sizing entirely and bills by what flows through the cluster. Neither is universally cheaper; the cost-efficient choice depends on how steadily your streams run, which is the question this comparison answers. It is part of our complete guide to AWS cost optimization, the cluster pillar it links up to.
What is the difference between MSK Provisioned and MSK Serverless?
The difference is whether you buy and pay for broker capacity or pay only for throughput. MSK Provisioned bills per broker-hour by instance type plus provisioned storage and data transfer, so the meter runs at a fixed rate regardless of how busy the cluster is. MSK Serverless bills a per-cluster-hour charge plus per-partition-hour plus the gigabytes written and read plus storage, so the meter tracks usage and there are no brokers to size or leave idle. The practical consequence is that provisioned rewards high, steady utilization and punishes idleness, while serverless rewards variability and punishes sustained heavy throughput. Verify the current per-broker, per-partition, and per-gigabyte rates on the Amazon MSK pricing page before modeling either, since both schedules change.
| Criterion | MSK Provisioned | MSK Serverless |
|---|---|---|
| Billing basis | Per broker-hour + storage + transfer | Per cluster-hour + partition-hour + throughput + storage |
| Idle cost | Full broker cost even when quiet | Minimal beyond the cluster-hour and partitions |
| Cheapest for | Steady, high, predictable throughput | Bursty, low-average, unpredictable, or new workloads |
| Capacity planning | You size brokers and partitions | Scales automatically |
| Commitment discounts | Eligible via the EC2-backed model where applicable | None; pay-as-you-go |
The verdict: start a new or variable stream on MSK Serverless to avoid paying for brokers you cannot yet size, and move it to provisioned once throughput is steady and high enough that wholesale broker capacity is cheaper than per-gigabyte pricing. Many estates run both, matching each stream to its shape.
When is MSK Provisioned the cheaper choice?
Provisioned wins when throughput is high and flat enough to keep brokers consistently busy. Because you pay a fixed broker rate, the cost per gigabyte falls as utilization rises, so a stream that runs near capacity around the clock pays less per unit on provisioned than it would on serverless throughput pricing. The discipline that makes provisioned cheap is the same as any reserved capacity: size the broker count and type to real throughput rather than a comfortable maximum, set topic retention to what consumers actually replay rather than a default many days long, and use tiered storage where supported to push older log segments to a cheaper tier. This is the streaming counterpart to the right-sizing in our sibling guide on cutting Amazon Kinesis and Data Streams costs.
Kafka retention defaults are generous, and storage bills accrue every hour data sits on the cluster. Set retention to the longest replay any consumer genuinely needs, and tier older data, before you decide the cluster is too small.
When is MSK Serverless the cheaper choice?
Serverless wins when traffic is bursty, low on average, or too new to size confidently. Provisioned brokers bill continuously, so a stream that is busy for a few hours a day and quiet the rest pays for capacity it does not use; serverless charges track the actual partitions and gigabytes, so the quiet hours cost little. It is also the right default for a new workload whose throughput you cannot yet predict, because guessing broker counts wrong means either throttling or paying for idle capacity. The trade is that at sustained high throughput the per-gigabyte serverless pricing can exceed wholesale broker cost, so revisit the choice once the workload matures. This mirrors the provisioned-versus-serverless logic in our work on EMR cost optimization.
Paying for Kafka brokers that sit half idle?
Our AWS cost audit maps each stream to the cheaper MSK model, right-sizes brokers and retention on the provisioned clusters, and proves the saving against a clean baseline. On the performance model, you pay only from realized savings. No savings, no fee.
Book a cloud cost audit →The AWS Cost Optimization Field Guide includes the streaming cost model we use to decide provisioned versus serverless per stream and to right-size retention.
Frequently asked questions
Is MSK Serverless cheaper than MSK Provisioned?
It depends on the traffic shape. MSK Serverless is cheaper for bursty, low-average, or unpredictable workloads because it bills per partition-hour and per gigabyte of throughput with no idle brokers. MSK Provisioned is cheaper for steady, high-throughput streams because you buy broker capacity wholesale and keep it busy.
How is MSK Provisioned priced?
MSK Provisioned bills per broker-hour by instance type, plus provisioned storage per gigabyte-month, plus data transfer. You pay for the brokers whether or not they are busy, so the cost is predictable and lowest per unit when utilization is high.
How is MSK Serverless priced?
MSK Serverless bills a per-cluster-hour charge plus per-partition-hour, plus throughput for data written and read in gigabytes, plus storage. There are no brokers to size, so cost tracks actual usage, which suits variable or spiky workloads but can exceed provisioned cost at sustained high throughput.
What is the biggest hidden MSK cost?
Storage and over-provisioning. On provisioned clusters, long retention and over-sized brokers inflate the bill, and idle brokers bill continuously. Set retention to what consumers actually need, right-size broker count and type to real throughput, and use tiered storage where supported to move older data to a cheaper tier.
The short version
MSK Provisioned bills for brokers around the clock and is cheapest for steady, high throughput; MSK Serverless bills for partitions and gigabytes and is cheapest for bursty or unpredictable streams. Match each stream to its shape, right-size brokers and retention on the provisioned side, and revisit serverless streams as they mature toward steady load. When you want every stream mapped to the cheaper model and the provisioned clusters right-sized, that is part of what our AWS cost optimization service delivers.
Cloud pricing and service behavior change frequently. Verify the specifics in this guide against the providers’ own current documentation and the FinOps Foundation: AWS pricing ↗, AWS documentation ↗ and FinOps Foundation Framework ↗. This article also reflects Cloud Cost Room’s hands-on, vendor-neutral engagement experience.