Pick the wrong managed cloud Kafka platform in 2026 and the bill gap alone can run into six figures a year. Confluent Cloud’s entry-level Freight tier for high-throughput batch workloads starts at roughly $2,300 a month, while its Basic tier costs $0 until data actually flows — a 46x spread across one vendor’s own tier ladder. Amazon MSK bills per broker-hour starting around $0.21, and Redpanda Cloud’s Serverless option is free to start with usage-based metering layered on top. Three platforms, three pricing philosophies, and a specs gap wide enough to swing a project’s infrastructure budget by tens of thousands of dollars a year.
This comparison breaks down Confluent Cloud, Amazon MSK (including the newer MSK Express brokers), and Redpanda Cloud across pricing, architecture, throughput, reliability, and real-world fit as of August 2026. Kafka itself is free and open source, but almost nobody runs a self-managed cluster at scale anymore — the operational overhead of partition rebalancing, broker patching, and ZooKeeper (or KRaft) babysitting pushed most teams toward a managed service years ago, a shift covered in our Apache Kafka KRaft cluster tutorial. The question now is which managed service, and the answer depends heavily on how much you’re moving, how predictable your traffic is, and whether you need multi-cloud portability. It’s also worth weighing these three against adjacent messaging options like Kafka vs Kinesis, SQS vs Kafka, or RabbitMQ vs Kafka if your workload doesn’t strictly need Kafka-protocol compatibility.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Confluent Cloud, Amazon MSK, and Redpanda Cloud Actually Are
All three products speak the Kafka wire protocol, which means client libraries, connectors, and most tooling built for Apache Kafka work against any of them with minimal changes. Past that, they diverge sharply in how they’re built and who runs the infrastructure underneath.
Confluent Cloud is the fully managed streaming platform from Confluent, the company founded by the original creators of Apache Kafka. It runs across AWS, Google Cloud, and Azure with a single control plane, and in 2026 it’s positioned itself less as “Kafka as a service” and more as a full data-streaming platform — adding Tableflow (turns Kafka topics into queryable Iceberg tables), Flink SQL for stream processing (with Flink data transfer priced at $0.00/GB through March 2026, per Confluent’s own documentation, as a promotional on-ramp for stream-processing workloads), and a specialty high-throughput cluster type called Freight.
Amazon MSK (Managed Streaming for Apache Kafka) is AWS’s native managed Kafka offering. It runs exclusively on AWS infrastructure and ties tightly into the rest of the AWS ecosystem — IAM for auth, VPC for networking, CloudWatch for metrics, and direct delivery to S3 and Redshift. AWS ships three consumption models: MSK Provisioned (you pick broker instance types and count), MSK Serverless (pay per partition and per GB), and MSK Express brokers (a broker type that got a significant capability bump in July 2026 with direct S3 delivery).
Redpanda Cloud is the managed offering from Redpanda Data, which rebuilt the Kafka protocol from scratch in C++ instead of running on the JVM. Redpanda markets itself on operational simplicity: no ZooKeeper, no JVM garbage-collection pauses, and a thread-per-core execution model built on the Seastar framework (the same async C++ framework behind ScyllaDB). Redpanda Cloud ships in three flavors: Serverless, Dedicated, and BYOC (Bring Your Own Cloud), where the data plane lives inside the customer’s own AWS, GCP, or Azure account.
Architecture: JVM and ZooKeeper vs KRaft vs C++ Thread-Per-Core
The architectural split explains most of the operational and performance differences you’ll see later in this comparison.
Confluent Cloud and Amazon MSK both run genuine Apache Kafka under the hood. Kafka itself has moved off ZooKeeper in favor of KRaft (Kafka Raft) for metadata management in recent major versions, and both managed services have followed that migration — AWS specifically rolled KRaft support into MSK Express brokers in December 2025 alongside a bump to Apache Kafka 3.9, according to Amazon MSK’s own documentation — which simplifies cluster metadata handling and removes a historically fragile dependency. Both still run on the JVM, though, which means garbage-collection pauses remain a factor in tail-latency-sensitive workloads, even if modern collectors have narrowed that gap significantly compared to a decade ago.
Redpanda took a different path entirely. It’s written in C++ on top of Seastar, using a thread-per-core model that pins one application thread to each CPU core and avoids cross-core locking and context switching. According to Redpanda’s own architecture documentation, this design eliminates JVM garbage collection entirely and uses Raft consensus for replication instead of ZooKeeper or a separate KRaft controller layer. Redpanda’s marketing materials claim this delivers markedly lower tail latency than JVM-based Kafka, though those figures come from Redpanda’s own benchmarking rather than independent third-party testing, so treat headline multipliers with appropriate skepticism until you’ve run your own workload against them.
All three platforms now support Tiered Storage — offloading older log segments to cheap object storage (S3, GCS, or Azure Blob) while keeping recent data on fast local disks. Confluent Cloud, MSK, and Redpanda Cloud each implement this differently, but the practical effect is the same: long-retention topics no longer force you to provision expensive broker storage for data nobody’s actively reading.
Specs Comparison: Confluent Cloud vs Amazon MSK vs Redpanda Cloud
Here’s how the three platforms stack up across the specs that actually matter for a production decision.
| Spec | Confluent Cloud | Amazon MSK | Redpanda Cloud |
|---|---|---|---|
| Underlying engine | Apache Kafka (JVM, KRaft) | Apache Kafka (JVM, KRaft) | Custom C++ (Kafka-protocol compatible) |
| Cloud availability | AWS, GCP, Azure | AWS only | AWS, GCP, Azure (BYOC or hosted) |
| Entry-tier price | $0/month (Basic, usage-based) | ~$460/month (3x kafka.m5.large brokers) | $0 to start (Serverless, usage-based) |
| Top-tier throughput | 9,120 MB/s ingress / 27,360 MB/s egress (Freight) | Up to 1,000 MB/s per Express broker (m7g.16xlarge) | Up to 2 GB/s write (BYOC, per vendor spec) |
| Max partitions (top tier) | ~50,000 (Freight) / 96,000 (Enterprise) | Scales with broker count and type | Scales with cluster size |
| Highest published SLA | 99.99% (Enterprise/Freight, 2+ eCKUs) | Standard AWS regional SLA | 99.99% (BYOC, annual commit) |
| ZooKeeper dependency | None (KRaft) | None (KRaft) | None (Raft-native) |
| Runtime | JVM | JVM | No JVM (native C++) |
| Tiered Storage to object storage | Yes | Yes (via S3 delivery) | Yes (native, all tiers) |
| Serverless option | Basic tier (elastic, pay-per-use) | MSK Serverless | Redpanda Serverless |
| BYOC / self-managed VPC option | No | No (AWS-managed only) | Yes |
| Native table format integration | Tableflow (Apache Iceberg) | Direct S3 delivery to Iceberg tables (July 2026) | Via Redpanda Connect + external catalogs |
| Kafka Connect-compatible connectors | Yes (managed connectors) | Yes (MSK Connect) | Yes (Redpanda Connect) |
A few things jump out. First, cloud portability: only Confluent Cloud and Redpanda Cloud’s BYOC model let you run outside AWS or move between clouds without a full re-platform. Second, entry pricing looks deceptively similar (all three offer a $0 starting point), but the moment real traffic hits, the billing models diverge hard — which is where the next section gets interesting.
Pricing Breakdown: The 46x Spread Between Entry and Top Tiers
Confluent Cloud’s pricing structure is the most transparent of the three, publishing exact per-tier numbers on its public pricing page, where per-unit eCKU rates as of a September 2026 rate-card update span $0.14 to $2.25 per hour depending on tier. Four cluster types exist: Basic, Standard, Enterprise (all “general purpose”), and Freight (a “specialty” cluster built for high-throughput batch workloads with relaxed latency requirements), and Confluent has kept sweetening the on-ramp for new accounts, reiterating as recently as August 2026 that first-time signups still receive $400 in free credit good for their first 30 days.
Basic starts at $0/month and only bills once data actually moves through the cluster, capping out around 1,500 partitions and 250 MB/s ingress with a 99.5% uptime target — fine for dev, test, or genuinely low-volume production traffic. Standard clusters were repriced in September 2026: Confluent now lists the baseline rate at $1.50 per hour of compute plus usage, roughly $1,080/month before data and storage charges, up from the older $385/month estimate built around a single $0.75-per-hour eCKU; data in/out is still billed at $0.035–$0.050 per GB, and storage at $0.08 per GB-month. Enterprise clusters start around $895/month with eCKU rates between $1.75 and $2.25 per hour, scaling up to 96,000 partitions and 99.99% SLA with sub-100ms latency. Freight, the top tier, starts around $2,300/month with eCKU pricing at $2.25/hour (minimum 2 eCKUs), but drops ingress/egress pricing to $0.014–$0.030/GB and storage to just $0.03/GB-month — the tradeoff being latency relaxed to roughly 1–2 seconds instead of sub-100ms, since Freight targets batch throughput over real-time responsiveness. That $0 to $2,300 spread across Confluent’s own tier ladder is roughly a 46x difference in advertised starting price, though actual bills depend entirely on usage.
Amazon MSK prices differently — per broker-hour rather than per abstracted compute unit. AWS’s own August 2026 pricing example works out to $146.48 a month in total for a workload ingesting 1,000 GB, a useful real-world anchor before wading into the per-broker and per-GB line items below. MSK Provisioned starts at $0.21/hour for a kafka.m5.large broker or roughly $0.204/hour for the newer Graviton-based kafka.m7g.large, and AWS requires a minimum of three brokers for a production-viable cluster, putting a bare-bones provisioned cluster around $460/month before storage. Storage runs $0.10 per GB-month on EBS. MSK Serverless bills at $0.75 per cluster-hour, $0.0015 per partition-hour, $0.10 per GB of data in, and $0.05 per GB of data out. MSK Express brokers — the newer, higher-throughput broker type — run from $0.408/hour on an express.m7g.large up to $13.056/hour on an express.m7g.16xlarge, according to pricing tables published by third-party Kafka pricing trackers citing AWS’s rate card.
Redpanda Cloud publishes the least granular public pricing of the three. Serverless is free to start, billed on ingress, egress, storage, and partition count, positioned for smaller or spiky workloads. Dedicated and BYOC options move to custom, typically annual-commit pricing negotiated directly with Redpanda’s sales team, with BYOC specifically advertising up to 2 GB/s write throughput and a 99.99% SLA on annual contracts. The lack of a public self-serve price card for Dedicated and BYOC makes Redpanda harder to comparison-shop without talking to sales — worth factoring in if procurement speed matters to your team.
Pricing Table: Side-by-Side Cost Comparison
| Platform / Tier | Starting Price | Compute Unit | Data In/Out | Storage |
|---|---|---|---|---|
| Confluent Basic | $0/month | Usage-based | Metered post-free-tier | Metered post-free-tier |
| Confluent Standard | ~$385/month | $0.75/eCKU-hour | $0.035–$0.050/GB | $0.08/GB-month |
| Confluent Enterprise | ~$895/month | $1.75–$2.25/eCKU-hour | $0.020–$0.050/GB | $0.08/GB-month |
| Confluent Freight | ~$2,300/month | $2.25/eCKU-hour (min. 2) | $0.014–$0.030/GB | $0.03/GB-month |
| MSK Provisioned (m5.large) | ~$460/month (3 brokers) | $0.21/broker-hour | Standard AWS transfer rates | $0.10/GB-month (EBS) |
| MSK Provisioned (m7g.large) | ~$447/month (3 brokers) | ~$0.204/broker-hour | Standard AWS transfer rates | $0.10/GB-month (EBS) |
| MSK Serverless | Usage-based, no floor | $0.75/cluster-hr + $0.0015/partition-hr | $0.10/GB in, $0.05/GB out | Included in usage metering |
| MSK Express (m7g.large) | $0.408/broker-hour | Per-broker instance rate | ~$0.01/GB ingest (vendor est.) | $0.10/GB-month (EBS) |
| MSK Express (m7g.16xlarge) | $13.056/broker-hour | Per-broker instance rate | ~$0.01/GB ingest (vendor est.) | $0.10/GB-month (EBS) |
| Redpanda Serverless | $0 to start | Usage-based (ingress/egress/storage/partitions) | Metered, no public flat rate | Metered, no public flat rate |
| Redpanda BYOC / Dedicated | Custom, annual commit | Negotiated | Negotiated | Negotiated |
The practical takeaway: MSK Express, at $0.408 to $13.056 per broker-hour depending on size, undercuts Confluent’s per-hour compute pricing at small scale but requires more manual capacity planning, since you’re choosing instance types rather than an abstracted throughput unit. Confluent’s tiered eCKU model trades that planning overhead for simplicity, at a price premium that becomes more justified as you move up toward Enterprise and Freight, where the per-GB egress cost actually drops relative to Standard — and Confluent has been sweetening the connector side of the bill too, announcing in May 2026 discounts of up to 80% on Kafka Connect task pricing at scale. Redpanda’s lack of published Dedicated/BYOC pricing means it’s the hardest of the three to budget for without a sales conversation, though its Serverless tier competes directly with Confluent Basic and MSK Serverless on the low end.
Throughput and Latency Benchmarks From Three Sources
Benchmark numbers for managed Kafka platforms come almost entirely from vendor-published documentation rather than neutral third-party labs, so this section leans on primary sources with the caveat clearly flagged where a claim originates from the vendor itself.
Confluent’s own cluster-type documentation lists per-tier throughput ceilings: Basic and Standard both cap around 250 MB/s ingress and 750 MB/s egress, Enterprise scales to 1,920 MB/s ingress and 5,760 MB/s egress, and Freight — the specialty high-throughput tier — reaches 9,120 MB/s ingress and 27,360 MB/s egress. That’s roughly a 36x throughput gap between Confluent’s cheapest and most expensive tiers, which explains why Freight exists as a separate product rather than just a bigger Enterprise cluster.
AWS’s own Express broker documentation publishes a per-instance-size throughput table: an express.m7g.large broker handles roughly 15.6 MB/s ingress and 31.2 MB/s egress, scaling linearly up to an express.m7g.16xlarge at roughly 500 MB/s ingress and 1,000 MB/s egress. AWS states directly in its Express broker best-practices documentation that Express brokers deliver up to 3x more throughput per broker than Standard MSK brokers of the equivalent size — citing 500 MB/s safe write throughput on an m7g.16xlarge Express broker versus 153.8 MB/s on the Standard equivalent. In July 2026, AWS extended MSK Express with direct delivery to Amazon S3 at up to 10 GB/s, aimed at teams streaming straight into S3 Tables or Iceberg without standing up a separate Kafka Connect pipeline.
Redpanda’s own architecture and TCO materials claim significantly lower tail latency than JVM-based Kafka, attributing the gap to the absence of garbage-collection pauses and the thread-per-core Seastar design. Redpanda markets a claim of roughly 6x lower total cost of ownership compared to self-managed Apache Kafka clusters running equivalent workloads, according to the company’s own platform TCO page. Because this figure comes from Redpanda’s own comparison methodology rather than an independent audit, it’s best read as a directional claim worth validating against your specific workload rather than a universal multiplier — the same caution that should apply to any vendor-published performance number, including Confluent’s and AWS’s.
The pattern across all three sources: raw per-broker or per-tier throughput numbers vary by 30x or more within a single vendor’s own tier ladder, which matters more for sizing decisions than any cross-vendor throughput claim. Before trusting a headline multiplier from any of the three platforms, run a proof-of-concept against your actual message size, partition count, and consumer fan-out, since synthetic benchmarks rarely match production traffic patterns.
MSK Express Brokers: What Changed in 2026
MSK Express brokers weren’t brand new in 2026, but AWS kept extending the feature set throughout the year in ways that matter for anyone comparing it against Confluent and Redpanda. The core pitch remains the same: Express brokers use the express.m7g instance family, are pre-tuned for streaming ingestion, autoscale storage automatically, and cut down the number of configuration knobs an operator has to manage compared to a fully manual MSK Provisioned cluster.
The July 2026 update added direct delivery from MSK Express brokers to Amazon S3 at up to 10 GB/s, letting teams skip a separate Kafka Connect S3 sink connector for the common pattern of landing streaming data into a data lake. AWS frames this as removing “incremental infrastructure costs of scaling connector pipelines,” since the S3 delivery capacity now scales with broker throughput rather than a separately provisioned connector fleet. For teams already committed to the AWS ecosystem and building toward Iceberg tables on S3, this closes a real gap that previously required Confluent’s Tableflow or a custom pipeline to replicate.
The tradeoff is that Express brokers, like all of MSK, remain locked to AWS. If multi-cloud portability or avoiding vendor lock-in is a priority, the throughput and integration gains from Express brokers don’t offset that constraint.
Confluent Cloud’s Freight Tier and Tableflow
Confluent’s two most consequential 2026 additions both push the platform further from “just managed Kafka” and toward a broader streaming-and-analytics platform. Freight, the specialty cluster type covered above, exists specifically for high-volume batch ingestion where sub-100ms latency doesn’t matter but raw throughput and lower per-GB cost do — think log aggregation, IoT telemetry batching, or bulk data migration jobs where a 1–2 second delay is irrelevant.
Tableflow, meanwhile, converts Kafka topics directly into Apache Iceberg tables without a separate ETL job. It’s billed as a “process” capability under Confluent’s usage-based billing model, metered by topic-hour and GB processed, with tiered pricing that Confluent introduced at its May 2026 launch dropping as low as $0.01 per topic-hour according to Confluent’s own Q2 2026 launch blog. This directly answers the same “streaming into a queryable table format” need that AWS addressed with MSK Express’s S3 delivery upgrade — both vendors converged on the same problem from opposite ends of their respective platforms in the same year, which says something about where the market’s attention has shifted: less about raw message throughput, more about making streaming data immediately queryable for analytics without a separate batch pipeline.
Redpanda’s $1B Valuation and the BYOC Bet
Redpanda Data closed a $100 million Series D round led by GV (Google’s venture arm) in April 2025, pushing the company’s valuation to $1 billion according to Redpanda’s own press release. The round doubled as a product announcement, with Redpanda simultaneously launching an enterprise agentic AI platform layered on top of its streaming infrastructure — a bet that AI agent pipelines will need low-latency event streaming as a backbone, not just batch data warehousing. For context on where that valuation sits relative to the broader streaming-data sector, analyst firm GenesisIQ cited a comparable benchmark in May 2026 of 10.8x FY2024 revenue from IBM’s acquisition of Confluent, suggesting Redpanda’s $1 billion mark reads as conservative rather than frothy against the sector’s own recent M&A math.
Strategically, Redpanda’s BYOC model is its clearest differentiator against both Confluent Cloud and Amazon MSK. Regulated industries — banking, healthcare, government contractors — often can’t put data in a fully multi-tenant SaaS control plane, no matter how strong the vendor’s compliance certifications are. BYOC splits the difference: Redpanda operates the control plane (monitoring, orchestration, upgrades) while the actual data plane, and therefore the data itself, stays inside the customer’s own AWS, GCP, or Azure account and VPC. Confluent doesn’t offer an equivalent BYOC tier, and MSK is inherently “your AWS account” by definition but locks you to a single cloud. For a team that specifically needs its data to never leave its own VPC while offloading day-to-day operations, Redpanda’s BYOC tier is currently the most direct answer among the three platforms compared here.
Reliability: 2026 Outages and SLA Comparison
Reliability history matters as much as raw specs when you’re picking infrastructure that sits in the critical path of every event in your system. Here’s what the 2026 incident record shows for each platform.
Amazon MSK doesn’t run in an isolated bubble — it’s built on core AWS infrastructure (EC2, EBS, networking), so regional AWS incidents ripple through to MSK even when the streaming service itself isn’t the root cause. AWS’s us-east-1 region saw a major disruption in 2026 that a third-party outage tracker measured at roughly 28 hours of degraded service, with AWS documentation for that period listing MSK, along with EC2, EBS, EKS, Redshift, ElastiCache, and OpenSearch, among the services that degraded as a downstream effect. A separate, unrelated AWS outage hit the Middle East region for roughly 7 hours and 23 minutes in March 2026, and a broader AWS disruption on July 24, 2026, knocked out access to services like Apple Pay, DoorDash, and Reddit for users relying on AWS-hosted backends. None of these were MSK-specific failures, but any workload running MSK in an affected region felt the impact alongside every other AWS service in that region — see our AWS US-East-1 outage breakdown for the full incident timeline.
Redpanda Cloud had a more contained incident: a DNS issue affecting the Redpanda Cloud API on April 27, 2026, which Redpanda’s status page logged as a full outage of the control-plane API, resolved the same day. Notably, Redpanda’s own incident notes state that customer clusters themselves remained unaffected during the API disruption — the control plane went down, but running clusters kept processing data, which is a meaningfully different failure mode than a regional cloud outage taking the underlying compute out entirely.
Confluent Cloud’s SLA structure is the most explicitly tiered of the three: 99.5% for Basic, 99.9% at a single eCKU on Standard rising to 99.99% at two or more eCKUs, and 99.99% for both Enterprise and Freight. This tiered SLA approach means the uptime guarantee is a direct function of how much you’re paying, which is worth factoring into a total-cost comparison — a “cheap” Standard cluster running at minimum capacity carries a materially weaker uptime commitment than the same tier scaled up.
Self-Managed Kafka vs All Three Managed Options: Is the Premium Worth It?
Every team evaluating Confluent Cloud, Amazon MSK, or Redpanda Cloud eventually asks the obvious question: why not just run open-source Apache Kafka on raw EC2 instances or a self-managed Kubernetes cluster and skip the managed-service premium entirely? The honest answer is that self-managed Kafka is cheaper on paper and more expensive in practice for almost every team past a certain size.
Running three m5.xlarge EC2 instances plus EBS storage for a bare-bones self-managed cluster costs roughly $0.51/hour in raw compute, according to comparison figures published alongside MSK pricing breakdowns — on the surface, cheaper than an equivalent MSK Provisioned cluster. What that number leaves out is the engineering time spent patching brokers, managing partition rebalancing, tuning JVM garbage collection, monitoring replication lag, and responding to 2 a.m. pages when a broker falls over during a traffic spike. Redpanda’s own TCO materials lean on exactly this gap, claiming a roughly 6x total cost of ownership advantage versus self-managed Kafka once operational labor is factored in — a claim worth testing against your own team’s Kafka operational history rather than accepting at face value, since Redpanda designed the comparison methodology itself.
The calculus shifts depending on team size and existing expertise. A platform team that already runs Kafka in production, has deep operational muscle memory, and operates at massive, highly predictable scale can sometimes make self-managed Kafka pencil out — large streaming platforms at hyperscale companies still run this way. But for the overwhelming majority of teams adopting Kafka-compatible streaming for the first time, or scaling past what one or two engineers can comfortably operate, the managed-service premium buys back engineering time that’s almost always worth more than the sticker-price difference. That’s the core argument for Confluent Cloud, Amazon MSK, and Redpanda Cloud all existing as viable, well-funded products in 2026 rather than open-source Kafka simply eating the entire market on cost alone.
Real-World Use Cases: Where Each Platform Fits
Rather than naming specific companies — public production references for MSK and Redpanda specifically are thin in currently available reporting, and vendor marketing logos change too often to cite reliably — here are five realistic workload profiles and which platform tends to fit best, based on the pricing, architecture, and feature differences covered above.
- Early-stage startup validating a real-time feature. Confluent Basic or Redpanda Serverless both start at $0 and bill only on usage, making either a reasonable low-risk starting point. Confluent’s multi-cloud reach matters less at this stage than getting a working pipeline fast.
- AWS-native team streaming into a data lake. MSK Express brokers with the July 2026 direct S3 delivery feature remove a whole connector layer for teams already standardized on AWS and building toward S3 Tables or Iceberg, at a lower operational cost than running a separate Kafka Connect fleet.
- Regulated fintech or healthcare workload requiring VPC-resident data. Redpanda Cloud’s BYOC tier keeps the data plane inside the customer’s own cloud account while offloading operations to Redpanda — a fit that neither Confluent Cloud nor Amazon MSK directly replicates.
- High-volume batch ingestion, like IoT telemetry or log aggregation. Confluent’s Freight tier is purpose-built for exactly this: throughput over latency, with per-GB pricing that drops as volume climbs, at the cost of 1–2 second latency that’s irrelevant for batch use cases.
- Multi-cloud SaaS company avoiding single-vendor lock-in. Confluent Cloud’s native support for AWS, GCP, and Azure under one control plane is the most direct fit; MSK is a non-starter outside AWS, and Redpanda requires the more operationally involved BYOC path to achieve the same portability.
Migration Guide: Moving Between Kafka-Compatible Platforms
Because all three platforms speak the same Kafka wire protocol, migrating between them is closer to a configuration change than a rewrite — but the details still matter. Here’s the general path most teams follow.
- Audit your current topic and partition layout. Export topic configs, partition counts, replication factors, and retention policies from your source cluster before touching anything.
- Map consumer group offsets. Offset semantics generally carry over between Kafka-compatible platforms, but confirm your target platform’s exact behavior around offset retention windows, since defaults differ between Confluent, MSK, and Redpanda.
- Stand up the target cluster at matching or greater capacity. Undersizing the destination is the single most common cause of migration-day incidents; use the specs table above to match throughput and partition ceilings before cutover.
- Use MirrorMaker 2 or a vendor-native replication tool for the data sync. Confluent Cloud offers Cluster Linking, MSK supports MirrorMaker 2 and MSK Replicator, and Redpanda offers its own migration tooling built on the same Kafka replication protocol — all three avoid a full re-ingest from source systems.
- Run both clusters in parallel and validate consumer lag. Point a subset of non-critical consumers at the new cluster first and confirm lag stays flat before cutting over production traffic.
- Cut over producers in waves, not all at once. Move lower-risk topics first, watch error rates and latency for 24–48 hours, then move critical topics.
- Decommission the source cluster only after a full retention window has passed on the new one. This gives you a rollback path if an issue surfaces after cutover that wasn’t caught during parallel running.
- Re-point Kafka Connect connectors and schema registries last. Connector configs and schema registry URLs are the most common thing teams forget to update, causing silent data pipeline failures days after the “successful” cutover.
The trickiest migrations are the ones moving into or out of Redpanda, not because the protocol compatibility breaks, but because teams sometimes assume 100% behavioral parity with Kafka and get surprised by edge cases in exactly-once semantics or transaction handling. Test transactional workloads specifically before committing to a full cutover, regardless of which direction you’re migrating.
Pros and Cons of Each Platform
Confluent Cloud
Pros: Multi-cloud support across AWS, GCP, and Azure under one control plane; the most transparent public pricing of the three; Tableflow and Freight address emerging analytics and batch-throughput needs without third-party tooling; tiered SLAs up to 99.99%; built by the original Kafka creators, which matters for confidence in long-term protocol fidelity.
Cons: Generally the most expensive of the three at equivalent throughput once you move past the Basic tier; per-GB egress pricing on Standard and Enterprise ($0.035–$0.050/GB) is higher than MSK Express’s estimated ingest cost; no BYOC option for teams needing VPC-resident data.
Amazon MSK
Pros: Deep native integration with the rest of AWS (IAM, VPC, CloudWatch, S3, Redshift); MSK Express brokers offer up to 3x the throughput of Standard brokers at a competitive per-broker-hour rate; the new direct S3 delivery feature removes a whole connector layer for data-lake use cases; per-broker pricing gives granular cost control for teams comfortable with capacity planning.
Cons: AWS-only, with no path to multi-cloud without a full re-platform; inherits AWS’s own regional outage risk, and 2026 saw multiple multi-hour AWS incidents that touched MSK-dependent services; provisioned pricing requires more manual instance-type and broker-count decisions than Confluent’s abstracted eCKU model.
Redpanda Cloud
Pros: No JVM means no garbage-collection pauses affecting tail latency; BYOC option is the strongest data-residency answer of the three; Serverless tier starts free with straightforward usage-based billing; backed by a fresh $1B valuation and $100M Series D, suggesting runway for continued development.
Cons: No public self-serve pricing for Dedicated or BYOC tiers, requiring a sales conversation to budget accurately; smaller ecosystem and fewer public production case studies to reference than Kafka or MSK; as a from-scratch reimplementation of the Kafka protocol rather than genuine Apache Kafka, some edge-case behavioral differences (particularly around transactions) require explicit testing before migrating critical workloads.
Which Should You Choose? The Verdict
There’s no single winner here, and anyone telling you otherwise is oversimplifying a decision that genuinely depends on your cloud strategy, budget shape, and compliance requirements. That said, the data points toward a few clear defaults.
If you’re already committed to AWS and building toward a data lake or lakehouse architecture, MSK Express brokers with the 2026 direct S3 delivery feature are the most cost-efficient path, undercutting Confluent’s per-GB egress pricing while removing a connector layer most teams would otherwise have to build and maintain separately. If you need multi-cloud flexibility, a transparent pricing ladder, or you’re building analytics workloads that benefit from Tableflow’s direct Iceberg integration, Confluent Cloud’s Enterprise or Freight tiers justify their premium, especially once you factor in the 99.99% SLA and the operational time saved not managing broker instance types directly. And if data residency inside your own VPC is a hard requirement — common in banking, healthcare, and government contracting — Redpanda Cloud’s BYOC tier is currently the most direct answer among the three, backed by fresh funding and an architecture explicitly designed to avoid JVM-related latency spikes.
Whichever platform you land on, the throughput and pricing numbers in this comparison are starting points, not guarantees. Run a proof-of-concept against your actual message sizes and consumer patterns before committing budget, since the vendor-published benchmarks covered here — Confluent’s tier ceilings, AWS’s Express broker tables, and Redpanda’s TCO claims — all come from each company’s own testing methodology.
Frequently Asked Questions
Is Redpanda actually compatible with Apache Kafka?
Yes, at the wire-protocol level. Redpanda implements the Kafka API, so existing Kafka client libraries, producers, and consumers work against it without code changes. It is not, however, a fork of Apache Kafka’s own codebase — it’s a from-scratch C++ reimplementation of the protocol, which means some edge-case behaviors (particularly around exactly-once transactions) can differ and should be tested before migrating critical workloads.
Can I run Confluent Cloud or Redpanda Cloud on AWS alongside existing AWS services?
Yes. Both Confluent Cloud and Redpanda Cloud support AWS as one of their supported clouds, and both can integrate with AWS-native services like S3, Lambda, and IAM through connectors or VPC peering. Amazon MSK is the only one of the three that runs exclusively on AWS with no other cloud option.
What is the cheapest way to start with managed Kafka?
Confluent Cloud Basic and Redpanda Cloud Serverless both start at $0/month with usage-based billing once data flows. Amazon MSK has no equivalent zero-floor tier — MSK Serverless has no monthly minimum but bills per cluster-hour and per-GB from the first byte, and MSK Provisioned requires standing up at least three brokers at roughly $460/month before any traffic even flows.
What are MSK Express brokers and how are they different from standard MSK?
MSK Express brokers are a newer AWS broker type built on the Graviton-based express.m7g instance family, offering up to 3x the throughput per broker compared to Standard MSK brokers of the same size, with less manual configuration required. As of July 2026, Express brokers can also deliver data directly to Amazon S3 at up to 10 GB/s, removing the need for a separate S3 sink connector in many data-lake pipelines.
Does Redpanda require ZooKeeper?
No. Redpanda was designed without a ZooKeeper dependency from the start, using Raft consensus natively for metadata and replication. This puts it ahead of legacy Kafka deployments still running ZooKeeper, though modern Apache Kafka (and therefore Confluent Cloud and Amazon MSK) has also moved to KRaft, its own ZooKeeper-free metadata mode, so this is no longer a unique Redpanda advantage against current-generation Kafka.
Which platform has the strongest uptime SLA?
Confluent Cloud publishes the clearest tiered SLA structure, reaching 99.99% on Enterprise and Freight clusters (and on Standard clusters running 2 or more eCKUs). Redpanda Cloud’s BYOC tier also advertises 99.99% on annual-commit contracts. Amazon MSK follows AWS’s standard regional service-level commitments, but as a service built on core AWS infrastructure, it’s also exposed to broader AWS regional incidents, several of which occurred during 2026 and affected MSK alongside other AWS services in the same region.
Can I use Kafka Connect connectors with all three platforms?
Yes, all three offer Kafka Connect-compatible connector systems: Confluent Cloud has fully managed connectors, MSK has MSK Connect, and Redpanda has Redpanda Connect. Community and vendor-built Kafka Connect connectors generally work across all three with minimal adaptation, since they target the same underlying connector API.
How big is the managed Kafka market in 2026?
Market research firm Dataintelo projects the managed Kafka services market to grow from roughly $5.7 billion to $13.67 billion by 2033, a compound annual growth rate of about 27.6%, with cloud-based deployment already accounting for an estimated 68.4% of the market as of 2025. That growth trajectory is a large part of why all three vendors covered in this comparison shipped major new capabilities — Confluent’s Freight and Tableflow, AWS’s Express S3 delivery, and Redpanda’s agentic AI platform — within the same 12-month window.
Is it worth switching from self-managed Kafka to a managed platform?
For most teams, yes, past a certain scale. Self-managed Kafka on raw compute looks cheaper on an hourly rate basis, but that figure excludes the ongoing engineering time spent on broker patching, partition rebalancing, and incident response. Teams without deep existing Kafka operational expertise, or those scaling past what one or two engineers can comfortably babysit, generally come out ahead moving to Confluent Cloud, Amazon MSK, or Redpanda Cloud, even accounting for the managed-service premium in the pricing tables above.
Related Coverage
- Pinecone vs Weaviate vs Qdrant: $500 Price Gap [2026]
- AWS RDS vs Azure Database vs Google Cloud SQL: 25% Price Gap [2026]
- RunPod vs Lambda vs Vast.ai: 63% H100 Price Gap [2026]
- LaunchDarkly vs Unleash vs GrowthBook: $2,800 Gap [2026]
- DigitalOcean vs Vultr vs Lightsail: 2.2x IOPS Gap [2026]
