The insight: Most Kafka cost isn't compute—it's the architectural overhead of disk-based replication. Eliminate the disks, eliminate the overhead.
Before optimizing, understand where the money goes. Here's how a typical Kafka deployment burns through your budget.
Based on 300 MB/s throughput, 50TB retention
+ engineering time not included
Three sources of cross-AZ traffic, all charged at $0.01-0.02/GB:
EBS GP3 costs $0.08/GiB vs S3 at $0.023/GB. With 3× replication, 50TB becomes 150TB. Add 100% over-provisioning for headroom, and you're paying for 304TB to store 50TB of data.
Partition rebalancing, broker scaling, capacity planning, incident response, on-call rotations. A typical team dedicates 0.5-1 full-time equivalent (FTE) to Kafka operations. What could they build instead? That's the cost you can't put on a spreadsheet.
Cross-AZ traffic—not storage—is often the largest infrastructure cost. Replication, producer writes, consumer reads—they all cross AZ boundaries. And it's the cost most teams don't see coming until the bill arrives.
Traditional Kafka's cost problems aren't bugs—they're architectural. Diskless fixes them at the root.
S3 provides 11 nines durability natively. No broker-level replication needed. Store 50TB as 50TB, not 304TB.
No cross-AZ replication traffic. Writes go to S3 in a single AZ. S3 handles cross-AZ durability internally—at no extra charge.
Stateless brokers scale to actual demand. No over-provisioning for peak. Scale down during off-hours, scale up when needed.
Scenario: 300 MB/s average throughput (614 MB/s peak), 1x fanout, 2-day retention, 50TB storage, Multi-AZ deployment
| Cost Component | Apache Kafka (Self-Hosted) | AWS MSK | Confluent Cloud | Redpanda (Tier 6) | AutoMQ |
|---|---|---|---|---|---|
| Compute | $4,415/mo (24× r5.xlarge) | $19,622/mo (32× m5.2xlarge) | Included in CKU | $28,827/mo (6× im4gn.8xl + ...) | $2,144/mo (9× m7g.2xlarge) |
| Storage | $24,300/mo (304TB EBS GP3) | $30,375/mo (304TB EBS) | $12,150/mo (50TB @ $0.24/GB) | $1,150/mo (50TB NVMe) | $2,926/mo (50TB S3) |
| Cross-AZ TrafficKey | $51,328/mo (replication + produce + consume) | $20,531/mo (produce + consume) | Included | $42,671/mo (network egress) | $0 |
| Platform Fee | $0 | $0 | $16,808/mo (CKU uptime) | $20,417/mo (licensing) | $300/mo |
| Ingress/Egress Fee | N/A | N/A | $47,735/mo (770TB × 2) | N/A | $15,696/mo (770TB × 2) |
| Total | $80,043/mo | $70,529/mo | $94,282/mo | $93,065/mo | $21,513/mo |
| vs AutoMQ | 3.7×more | 3.3×more | 4.4×more | 4.3×more | Baseline |
Different names, same disk-based design. Every platform below inherits the same architectural overhead.
The DIY tax. You provision for peak, pay for idle. 3× replication means 3× storage and 2× cross-AZ traffic. Ops burden not included in the bill—but you'll feel it.
Managed, but still disk-bound. EBS attached to every broker. AWS absorbs replication cost, but you're still over-provisioned. No elasticity—brokers run 24/7.
CKU pricing bundles compute and hides the math. Looks simple until you scale. Ingress/egress fees add up. Enterprise features require higher tiers.
Faster than Kafka, but same architecture. Local NVMe means attached storage on every node. Replication still generates cross-AZ traffic. Better performance, similar cost structure.
Diskless. Stateless. S3-native. The architecture eliminates the root causes of Kafka cost: no local disks, no replication traffic, no over-provisioning. Storage costs what S3 costs. Compute scales with demand.
The bill you see is only part of the story. Here's what doesn't show up on your cloud invoice.
Our team includes engineers who've managed Kafka at scale—and felt the pain firsthand. We'd love to hear about your challenges and share how we've solved similar problems.
No sales pitch. Just an honest conversation about Kafka economics.
Available on Cloud Marketplaces
Subscribe to AutoMQ directly from your preferred cloud platform
