Two months after IBM closed its $11B Confluent acquisition, every platform team I talk to is asking the same question: do we stay, or is this the year we finally rip Kafka out? It’s a fair question. The streaming landscape went through more change in the first half of 2026 than it did in the previous three years combined, and most of the comparison content on the open web is either vendor marketing or predates the acquisition entirely.
So here’s a vendor-neutral take. I’ve run Kafka in production, lost weekends to MSK upgrades, and migrated one workload from Confluent Cloud to AutoMQ this year. The recommendations below assume you have an actual production workload to feed, not a greenfield Hello World.
The 2026 streaming landscape in one paragraph
IBM closed the Confluent deal on March 17, 2026. WarpStream is now an IBM product, Confluent Cloud and Kora are inside the same org as watsonx, and the Q1 renewal cycle has been bumpy. AutoMQ stayed independent and published its 2026 benchmark. Bufstream went GA. Redpanda shipped 25.x with mature tiered storage. Apache Kafka 4.0 has been KRaft-only for over a year — Zookeeper is finally dead. And the diskless category — stateless brokers writing directly to S3 — went from “interesting experiment” to “the default for new high-volume workloads” in about eighteen months.
If you’re picking a streaming engine in 2026 without considering at least one diskless option, you’re probably overpaying.
What IBM owning Confluent actually means
IBM’s pitch is integration: Confluent Cloud, WarpStream, and Kora consolidated, plumbed into watsonx, sold into IBM’s enterprise account base. The reality customers are reporting in Q2 renewals is messier. List prices haven’t moved much yet, but the discounting pattern is different — multi-year commitments with watsonx attach are being pushed hard, and the standalone Confluent Cloud renewal without IBM cross-sell is suddenly less flexible than it was a year ago.
The WarpStream product specifically is in a strange spot. It was acquired by Confluent in late 2024, then rolled into IBM eighteen months later. The team is still mostly intact and the product is still shipping, but if you signed a WarpStream BYOC contract because you didn’t want to depend on Confluent, you now depend on IBM. That’s a different vendor risk profile than the one you bought.
For most existing Confluent customers, the right move this quarter isn’t a panicked migration — it’s getting a real cost model for what staying actually costs over the next three years, and a credible alternative quote to put next to it.
Diskless streaming, explained briefly
The old model: Kafka brokers run on EBS or local disks, replicate three ways across availability zones, and you pay for storage three times plus a non-trivial inter-AZ replication bill. The new model: brokers are stateless, every produce writes straight to S3 (or whatever your object store is), and a separate metadata service handles topic/partition state.
The economics are absurd if your workload tolerates it. Inter-AZ replication traffic goes to zero. Storage goes from 3x EBS to 1x S3. At 100 MiB/s with three-day retention, I’ve seen real bills drop from $14k/month on MSK to under $2k/month on AutoMQ self-hosted. AutoMQ’s own published benchmarks show ~90% cost reduction at high throughput, and they’re not wildly off.
The catch: latency. Diskless brokers buffer writes for ~250ms to amortize the cost of S3 PUTs, so your p99 produce latency is in the 300–500ms range instead of Kafka’s sub-10ms. For analytics ingestion, CDC, log pipelines, and AI-agent fan-out, that’s fine. For low-latency trading signals or real-time inventory updates, it’s not.
AutoMQ
Apache 2.0 licensed, single-binary deployment, and — this is the part that matters — it reuses Kafka’s actual computation layer instead of reimplementing the protocol. The team calls it “Native Kafka” and claims 73 of Kafka’s APIs are supported, which in practice means most Kafka Connect connectors, Kafka Streams jobs, and KSQL queries just work.
This is the one I’d pick today for a new high-throughput workload. It’s the most credible open-source diskless option after Confluent bought WarpStream, the team is shipping fast, and the licensing makes it easy to run anywhere. The trade-off is operational: you’re running it yourself, and the AutoMQ Cloud product is newer than WarpStream Cloud or Confluent Cloud. If you don’t want to operate a streaming platform, that matters.
WarpStream
Now an IBM product. The original pitch — a Go reimplementation of the Kafka protocol on top of S3, deployed BYOC in your own AWS account so your data never leaves — is still intact. WarpStream supports about 26 of Kafka’s APIs, which is enough for produce/consume but not enough for the full Kafka Connect ecosystem.
I’d still consider WarpStream if you’re already an IBM/Confluent customer and the BYOC isolation matters more than independence. If you originally picked WarpStream specifically to avoid Confluent dependency, the acquisition undermined the thesis and AutoMQ is the natural place to look.
Bufstream
Buf Technologies’ play, GA in 2026. S3-backed, Kafka-compatible, and the interesting bit: it’s Iceberg-native. Every topic can be queried as an Iceberg table without a separate sink job. If your end-state is “this data needs to land in our lakehouse anyway,” Bufstream collapses two pipelines into one.
Schema enforcement is Protobuf-first. That’s either a feature or a deal-breaker depending on your stack. For teams already on Buf’s schema registry it’s an obvious fit. For teams on Avro or JSON Schema, the migration tax is real.
Redpanda
The C++ thread-per-core Kafka clone. Single binary, no Zookeeper, no KRaft, and the lowest latency in the comparison. Redpanda 25.x shipped mature tiered storage, which lets you offload cold segments to S3 without going fully diskless.
This is the pick when you can’t tolerate diskless’s p99 floor. Real-time fraud, exchange order books, anything where the streaming layer’s latency directly affects user-visible behavior. The pricing is the gotcha — Redpanda has been steadily raising prices since their 2025 funding round, and the community edition has feature gaps (no tiered storage in CE, for one). Budget accordingly.
Apache Kafka, self-managed
Sometimes the right answer is still “just run Kafka.” KRaft is stable, Strimzi on Kubernetes is genuinely good now, and MSK Serverless solves the broker-sizing problem for teams who don’t want to think about partitions.
Where self-managed Kafka still wins: when you have an operations team that already knows Kafka inside-out, when your throughput is moderate enough that the diskless cost savings don’t matter, or when you have a regulatory constraint that forbids any vendor cloud touching your data. For everyone else, the operational cost of Kafka — broker sizing, partition rebalancing, ISR babysitting, the joy of running ZooKeeper — was always real, and diskless eliminates most of it.
RabbitMQ, NATS, and Pulsar — when Kafka is the wrong tool
I see teams reach for Kafka when they should be reaching for a message broker, and the bills get ugly fast.
RabbitMQ 4.x is still the right answer for complex routing, per-message acknowledgment, and workloads where each message is precious and you need head-of-line blocking semantics. Quorum queues and streams in 4.x close a lot of the gap with Kafka for moderate-throughput workloads, and the operational model is dramatically simpler than running brokers.
NATS / JetStream has become my default recommendation for AI agent fan-out. Thousands of agent processes subscribing to events, sub-millisecond pub/sub, lightweight enough to embed. JetStream gives you durable streams without Kafka’s operational weight. It is not a Kafka replacement for analytics pipelines, but if your “streaming” workload is actually pub/sub to a lot of consumers, this is what you want.
Apache Pulsar has the strongest geo-replication story and the cleanest multi-tenancy model. StreamNative keeps shipping. And yet most teams still don’t pick it. The operational complexity (BookKeeper + brokers + ZooKeeper-or-Oxia) is real, and the diskless wave has eaten Pulsar’s tiered-storage advantage. I’d consider it for genuinely global, multi-region streaming with strict tenant isolation. Otherwise, no.
Managed cloud options
AWS MSK is fine. MSK Serverless is better than its early reviews suggested, but the per-partition-hour pricing punishes high-partition-count workloads. GCP Pub/Sub is the most opinionated managed option and the easiest to start with — its lack of Kafka compatibility is a hard wall when you outgrow it. Azure Event Hubs exposes a Kafka API and works fine if your stack is already on Azure. Upstash Kafka wound down in late 2025, so don’t pick it. Aiven for Apache Kafka is a credible managed-Kafka option for teams who want vanilla Apache, not Confluent.
Real cost math, three scales
These numbers are AWS us-east-1, three-AZ deployment, 7-day retention, list pricing (no enterprise discount), as of May 2026. Treat them as order-of-magnitude, not invoice-ready.
10 MiB/s steady-state side project, single AZ tolerable.
- Self-hosted Kafka on Hetzner: ~$60/mo total
- AWS MSK Serverless: ~$220/mo
- Confluent Cloud Basic: ~$280/mo
- AutoMQ on a single EC2 + S3: ~$110/mo
At this scale, the diskless savings barely matter and operational simplicity wins. Pick whatever your team already operates, or pay for Confluent Basic if you have nobody to operate it.
100 MiB/s mid-size SaaS, three-AZ required.
- Self-managed MSK on m7g.large × 6: ~$4,800/mo with storage
- Confluent Cloud Standard: ~$11,000/mo
- WarpStream BYOC: ~$2,400/mo
- AutoMQ self-hosted: ~$1,900/mo
- Redpanda BYOC: ~$5,200/mo
- Bufstream self-hosted: ~$2,100/mo
This is the band where diskless changes the conversation. The gap between Confluent Cloud and AutoMQ is roughly 5x, and at $9k/month delta you can fund the engineering to operate AutoMQ yourself with money left over.
1 GiB/s with 3x fanout, three-AZ, enterprise.
- Self-managed Kafka: $90k+/mo (and a full-time platform team)
- Confluent Cloud Dedicated: $180k+/mo
- AutoMQ self-hosted: $24k–30k/mo
- WarpStream BYOC: $28k–34k/mo
At this scale, every team I’ve seen evaluate honestly has either moved to diskless or is in the middle of a migration. The math is too lopsided.
Decision matrix
- Low-latency event-driven microservices: Redpanda or Kafka. Diskless is wrong here.
- High-throughput analytics ingestion: AutoMQ. Best price/performance.
- AI agent fan-out, IoT, real-time pub/sub: NATS JetStream.
- CDC into Iceberg lakehouse: Bufstream. The integration savings are real.
- Complex routing, transactional messaging: RabbitMQ. Stop forcing Kafka into this shape.
- Regulated, air-gapped, must run on-prem: Self-managed Kafka with Strimzi, or AutoMQ self-hosted on MinIO.
- Multi-tenant SaaS streaming: Pulsar if you need strict isolation, otherwise AutoMQ with tenant-prefixed topics.
- Already on Confluent, contract renews this quarter: Get an AutoMQ quote, then negotiate.
Migration notes
The cleanest migrations I’ve seen this year share three properties. They run both systems in parallel for at least a sprint with production traffic mirrored to the new system. They cut over consumer groups one at a time, not all at once. And they keep the old cluster running for two weeks post-cutover with no traffic, just in case.
Confluent Cloud → AutoMQ is the migration of the year. The protocol compatibility is high enough that most Kafka clients work without code changes; the main work is operational, not application. Plan for two to four weeks of engineering for a mid-size workload.
WarpStream → AutoMQ is the post-acquisition route I’m seeing requested most. The motivation is usually “we picked WarpStream specifically to avoid Confluent.” Mechanically similar to a Confluent migration but with less consumer-API surface to worry about.
Self-managed Kafka → diskless is the highest-ROI migration most teams have available right now, and the one that gets postponed forever because nobody wants to touch a working streaming pipeline. The forcing function is usually an EBS bill audit.
RabbitMQ → NATS JetStream comes up when teams realize their RabbitMQ deployment is actually doing pub/sub at scale, not RPC. The semantic mismatch (no per-message ack, different routing model) makes this less of a drop-in than the Kafka migrations.
What I’d do this quarter
If you’re on Confluent Cloud and renewing this quarter, get an AutoMQ Cloud quote and a WarpStream IBM quote, side by side. Use them as leverage, or use them for real.
If you’re running Kafka on EBS and your bill is over $5k/month, pilot AutoMQ on a non-critical topic. The cost math is too obvious to ignore.
If you’re picking from scratch, default to AutoMQ for analytics-shaped workloads, Redpanda when latency matters, NATS for agent fan-out, RabbitMQ when each message is precious. Reach for vanilla Kafka only when you have a specific reason — and “we already know Kafka” is a legitimate reason, but be honest about whether it’ll still be one in two years.
The interesting thing about 2026 isn’t that any of these tools won. It’s that “Kafka or not” stopped being the question. Pick the engine that fits the workload, and keep an eye on what IBM does with WarpStream — that’s the variable that’ll move next.