Kafka pricing in 2026: what Confluent Cloud, MSK and self-managed Kafka cost

The table below prices three sustained write rates from the list prices above. It is illustrative, not a quote. The assumptions are fixed so you can swap in your own: one consumer group reads everything (egress equals ingress), seven days of retention on the logical data, a 730-hour month, US East pricing for AWS, one MB/s counted as 1,000 MB. It leaves out connectors, Schema Registry, private networking, cross-zone client traffic, support and any discount.

Confluent Cloud capacity comes from Confluent's documented per-unit ingress (25 MBps per Standard eCKU, 60 MBps per Enterprise eCKU), billed full-time. MSK Express broker counts come from AWS's recommended ingress per express.m7g.large broker (15.6 MBps), with three brokers as the floor, one per availability zone. MSK Serverless assumes 100 partitions. Ranges come from the price ranges Confluent publishes, and from not knowing whether Express storage is billed on one copy of the data or all three.

Self-managed infrastructure is not in the table because it depends on the instance type, disks, region and commitment you pick, and a made-up figure would mislead. The support row is the contract alone and adds to your servers and storage.

Two things stand out. First, capacity is the small part. At 250 MB/s on Confluent Standard the ten eCKUs come to about $5,500 a month, while data in and out is $46,000 to $66,000. On MSK Express the 17 brokers are about $5,100 and the $0.01 per GB ingest fee another $6,600. Second, the platforms meter different things, so the ranking changes with the workload. Run your own numbers on your own throughput and consumer count before you compare.

Sources, all fetched 26 October 2026: Confluent Cloud pricing, Confluent Cloud cluster types, Amazon MSK pricing, MSK Express best practices, Aiven for Apache Kafka pricing, Redpanda Cloud billing, Confluent Platform overview and the Confluent Community License FAQ.

Whichever platform you pick, the same few drivers decide the bill. Throughput and workload shape: Confluent Cloud scales eCKUs with load, while provisioned MSK brokers are sized for the peak, with AWS advising CPU under 60%. Consumer fan-out: every extra consumer group that reads the full stream adds a full copy of egress. At 50 MB/s that is 131,400 GB a month, which is $4,600 to $6,600 on Confluent Standard at $0.035 to $0.050 per GB and about $6,600 on MSK Serverless at $0.05. Partition count: Confluent caps a Standard cluster at 2,500 partitions, MSK Serverless bills $0.0015 per partition-hour (about $1.10 a month each), and AWS guides provisioned brokers to about 1,000 partitions per kafka.m7g.large. Retention: storage grows with every day you keep, and on provisioned brokers with every replica.

Connectors: Confluent bills managed connectors per task-hour and per GB, MSK Connect bills $0.11 per MCU-hour, and Kafka Connect itself ships with Apache Kafka if you run the workers. Data transfer: standard AWS charges apply to traffic in and out of an MSK cluster, so cross-zone and cross-region consumers belong in the estimate. Licensing: the Confluent Community License on components such as Schema Registry allows production use but bars offering a competing hosted service, which matters if you resell Kafka. That is a plain reading of the licence text, not legal advice. Open-source Apache Kafka, with the brokers, Kafka Connect and the clients, carries no such restriction.

AWS publishes an MSK Sizing and Pricing spreadsheet that estimates a provisioned cluster against a similar self-managed EC2 cluster, and its own caveat applies to any calculator: estimates are a starting point to verify with load testing.

Kafka storage costs are the line that grows quietly with scaling. One MB/s of sustained writes is 86.4 GB a day, so the stored volume is that figure multiplied by the days you keep data. The table prices that volume at Confluent's list rate of $0.08 per GB-month, which Confluent states before replication, and at AWS's $0.10 per GB-month for MSK provisioned storage with every byte held on three replicas. It uses the same simplifications as the workload table: logical data, no compression, a steady write rate.

The point is the slope, not the totals. Going from seven to thirty days of retention multiplies the storage line by about four on either platform, and on provisioned brokers it also multiplies disks you size and patch yourself. If a use case needs long replay, price it as its own line, and check whether an archive to object storage or tiered storage covers the old data at a lower rate. Confluent's page lists Freight storage at $0.03 per GB-month, but a different cluster type changes the capacity and networking lines too, so compare the whole bill.

Confluent Cloud bills three things on every cluster: capacity in eCKU-hours (an elastic unit), each GB written and read, and each GB-month stored. Confluent's page lists the capacity rate as $0.75 per eCKU-hour on Standard and $1.75 to $2.25 on Enterprise, with Basic free for the first eCKU. The "about $385" and "about $895" a month starting costs on that page assume the cluster is active about 70% of the time, so a cluster that runs flat out costs more: one Standard eCKU for 730 hours is $547.50.

Capacity limits decide which tier a workload needs. Confluent's documentation puts the cluster ceiling at 250 MBps ingress on Basic and Standard and 1,920 MBps on Enterprise. Each Standard eCKU carries 25 MBps of ingress, each Enterprise eCKU 60 MBps. The page does not print a price for connectors, Schema Registry, private networking or Flink; it says they are billed per task, per environment, per GB or per CFU, and Dedicated clusters say "Contact us". Those are the lines that move a bill after the first quote.

One more line to check when you size: Confluent's page prices data per GB before replication, and storage at $0.08 per GB-month on the same basis. MSK provisioned brokers store every replica on disks you pay for, so the same retention costs more storage there. Compare like with like.

MSK has three pricing models. Standard brokers bill by broker-hour: AWS lists kafka.m7g.large at $0.204 an hour in US East (N. Virginia), about $149 for a 730-hour month, plus $0.10 per GB-month of storage and $0.08 per MB/s-month if you provision storage throughput. Three of those brokers are about $447 a month before storage. Express brokers cost $0.408 an hour for an express.m7g.large, add $0.01 per GB ingested, and come with replication set to three. Serverless bills $0.75 per cluster-hour, $0.0015 per partition-hour, $0.10 per GB in, $0.05 per GB out and $0.10 per GB-month stored.

Other MSK lines: MSK Connect at $0.11 per MCU-hour (one MCU is 1 vCPU and 4 GB), multi-VPC connectivity at $0.0225 per hour plus $0.006 per GB, and standard AWS data transfer for traffic in and out. AWS does not charge for replication between brokers. Sizing has its own cost: AWS recommends keeping broker CPU under 60% so the cluster survives a patch or a failed broker, and recommends about 1,000 partitions per kafka.m7g.large, so a partition-heavy estate buys bigger brokers sooner than its throughput suggests.

An idle Serverless cluster still costs $0.75 an hour, about $547.50 a month, before a single message. That is the number to check against a development environment that nobody turned off.

Aiven publishes plan prices: Free at $0 a month for up to 250 KiB/s, Developer at $35 a month (or $0.04 an hour) for up to 1 MB/s with a 99% SLA, and Professional from $180 a month (or $3.84 an hour) with unlimited throughput and a 99.99% SLA. Business and Premium plans are named but not priced on the Kafka page. Its pricing page also shows workload estimates, from $276 a month at 0.1 MB/s to $4,226 at 10 MB/s, without saying which plan or assumptions sit behind them; treat them as Aiven's ballpark, not a quote.

Redpanda does not publish numeric rates. Its billing documentation names the meters (Serverless: uptime, GB in, GB out, partitions and storage; Dedicated: uptime, GB in and out, storage; BYOC: compute units, where one unit is 2 vCPU and 8 GB of memory, plus GB in and out and storage) and sends you to a price estimator or to billing@redpanda.com for rates. The pricing page says BYOC needs an annual commitment, and the self-managed Enterprise Edition comes with a 30-day trial. Any claim that Redpanda is cheaper is therefore a vendor estimate until you hold a quote. The licence terms are a separate question, answered in the FAQ below.

Confluent Cloud is the metered, fully managed cloud service. Confluent Platform is software you download and run yourself, and Confluent's documentation describes it as a distribution of Kafka that adds proprietary components to Kafka's open-source features. Confluent's pricing page offers a free download under the Confluent Community License and prints no per-core or per-node price for a Platform subscription, so a Platform deployment follows a different pricing model from the cloud service: a quote conversation, plus the infrastructure costs and the engineers who run it. Confluent Kafka and Apache Kafka share the Kafka protocol and the client libraries; what differs is the surrounding platform and its terms.

Confluent's Community License FAQ says the Community License is source-available rather than open source, because the Open Source Initiative has not approved it. You may run the software in production and embed it in software you distribute. The Excluded Purpose bars offering a SaaS, PaaS or similar online service that competes with Confluent's own, which is the clause that matters to a managed-service provider and rarely to a team running Kafka for its own applications. The FAQ also says support for Community software is community-only, with patches for one year per version, and that enterprise features are not included. Apache Kafka itself stays under Apache License 2.0, with no such restriction, and some connectors in Confluent's catalogue carry their own terms, so check each connector you rely on. This is a reading of the published FAQ, not legal advice.

Cost follows from that: running Confluent Platform on Kafka clusters in a public cloud account or in your own data centre removes the metered per-GB lines, and adds the deployment and operations work that Confluent Cloud was absorbing. Whether you deploy on the major cloud providers or in your own data centre, the question for a team weighing a hosted service against the option to run Kafka itself is which of the two bills you would rather manage, and who will manage Kafka day to day.

Apache Kafka is licensed under Apache 2.0, so the platform has no licence fee. The cost is the cluster plus the people. The cluster is brokers, disks for every replica, cross-zone traffic, and monitoring. The people are on-call, upgrades, partition rebalancing and the 3am incident. AceMQ puts a part-time engineer at $50,000 or more a year fully loaded, and independent 24/7 Kafka support runs about $24,000 to $72,000 a year, priced on cluster count and response tier.

The arithmetic from the table: at 5 MB/s a managed service lists at roughly $1,300 to $3,200 a month, so a support contract alone can cost as much as the whole platform bill, and a hosted service is the sensible choice unless data residency rules it out. As volume rises the per-GB lines overtake the contract. At 50 MB/s the managed lists above run $5,500 to $23,400 a month against a support contract of $2,000 to $6,000 plus your infrastructure. Where the crossover falls depends on your infrastructure rates, so get both numbers before deciding.

Support also covers the managed services. The 15-minute first response on a production emergency applies to self-managed, Amazon MSK and Confluent Platform clusters, which matters because a platform's SLA covers its service, not your consumers or topics. Ask for a Kafka cost review on your own throughput and consumer count through Kafka consulting.

This page prices the platform. It does not give a quote: committed-use discounts, marketplace purchases and negotiated enterprise agreements change every figure above. It does not cover support plan prices from Confluent or AWS, Kafka on Azure Event Hubs or Google Cloud, or the engineering cost of a migration. If you are weighing a move, use the cost view here first and then plan the work with an engineer.

• A workload near 5 MB/s with no residency rule: a managed service. At that rate the worked example lists a managed service at roughly $1,300 to $3,200 a month, so a support contract alone can cost as much as the platform.
• A workload near 50 MB/s: price Amazon MSK Express ($5,500 to $11,600 a month), Confluent Cloud Enterprise ($9,000 to $17,200) and Standard ($12,700 to $16,700) against self-managed Kafka under a support contract ($2,000 to $6,000 a month plus your infrastructure). The crossover depends on your infrastructure rates.
• A workload near 250 MB/s, or many consumer groups: per-GB charges dominate the bill. On Confluent Cloud Standard, data in and out is $46,000 to $66,000 of a month listed at $63,600 to $83,300, and each extra consumer group at 50 MB/s adds 131,400 GB of egress a month.
• A small or early workload: Aiven's Developer plan lists at $35 a month for up to 1 MB/s with a 99% SLA, and Professional from $180 a month.
• Data residency or control of the cluster: self-managed Apache Kafka, which carries no licence fee, with a support contract. You pay for the cluster and the people who run it.
• Redpanda: no numeric rates are published, so you need a quote before you can compare it.