Manu Shukla

Posted on Aug 3 • Originally published at ecorpit.com

$1 per million events: what event-driven architecture really costs in 2026

#aws #eventbridge #serverless #architecture

$1 per million events: what event-driven architecture really costs in 2026

Summary. Amazon EventBridge bills custom events at $1.00 per million, AWS management events at $0.00 per million, EventBridge Pipes at $0.40 per million requests and API destinations at $0.20 per million events. EventBridge Scheduler includes 14,000,000 free invocations a month before charging $1.00 per million. Amazon SQS gives every account 1 million free requests a month. At those rates, a service moving 50 million business events a month pays roughly $50 for the bus. That number is why the cost conversation about event-driven architecture is usually held in the wrong place. The expensive parts are the ones with no line item: debugging a flow with no single stack trace, the absence of ordering guarantees, and an operating model that assumes someone is watching a dead-letter queue at 2am. AWS's own decision guide for choosing between SQS, SNS and EventBridge was last updated on 31 July 2024 and contradicts itself on delivery semantics. This guide sets out the published rates, the guarantees you actually get, and the five signals that say a system is ready to move.

The rate card, in one place

Every figure below appears on Amazon's own pricing and FAQ pages as of August 2026.

Component	Billing unit	Published rate
Custom events on an event bus	Each 64 KB chunk of payload is billed as one event	$1.00 per million
AWS management events on an event bus	Same 64 KB chunking	$0.00 per million
EventBridge Pipes	Each 64 KB chunk billed as one request, charged after filtering	$0.40 per million requests
API destinations	Each 64 KB chunk billed as one event, same rate for public and private endpoints	$0.20 per million events
EventBridge Scheduler	Invocations, after a free tier of 14,000,000 per month	$1.00 per million
Archive and replay	Archive processing plus storage, then replayed events at custom-event rates	$0.10 per GB processed, $0.023 per GB stored, $1.00 per million replayed
Cross-Region event replication	Per event replicated	$1.00 per million
Schema Discovery	Each 8 KB chunk billed as one ingested event, after 5 million free per month	$0.10 per million
Amazon SQS	Every API action is a request; each 64 KB chunk of payload is one request	1 million requests free per month

Two details in that table catch teams out more than the headline rate.

The first is chunking. EventBridge bills each 64 KB chunk of payload as a separate event, so a 256 KB event is billed as four. SQS does the same: an API action with a 1 MiB payload is billed as 16 requests, and a single SQS request can carry 1 to 10 messages up to a total of 1 MiB. Fat events are not four times more expensive in some abstract sense; they are four times more expensive on the invoice. Sending an identifier and letting the consumer fetch the record is a pricing decision as much as a design one.

The second is that the Schema Registry is free but Schema Discovery is not, and Discovery chunks at 8 KB rather than 64 KB. Leaving discovery running across a busy bus after the schemas have stabilised is a small, silent bill.

Run the arithmetic on a realistic pipeline. A retail order service publishing 50 million custom events a month, each comfortably under 64 KB, pays $50 for ingestion. Archive the same 50 million events at roughly 6 KB each and you are storing about 286 GB: $28.61 in archive processing and $6.58 a month in storage. A full replay of that archive costs another $50. Replicating the same stream to a second Region for resilience adds $50. The whole event layer for that service lands under $200 a month. One senior engineer spending two days a month chasing a message that vanished costs more than the entire bus.

The real cost is the operating model, not the messaging bill.

What you actually get, and what you do not

AWS publishes the guarantees. They are worth reading before the design review, not after the incident.

Property	Amazon SQS	Amazon SNS	Amazon EventBridge
Communication model	"Pull-based (consumers poll messages from the queue)"	"Push-based (subscribers receive messages when published)"	"Event-driven (rules match events and route to targets)"
Persistence	"Messages persist until consumed or expired"	"Messages do not persist; delivered in real-time to subscribers"	"Events do not persist; processed in real-time"
Ordering	"FIFO (First-In-First-Out) queues ensure strict ordering"	"Amazon SNS FIFO topics guarantee order"	"No ordering guarantees"
Filtering	Basic controls through visibility timeouts and dead-letter queues	"Message filtering using subscription filter policies based on message metadata"	"Complex event pattern matching and content-based filtering"
Typical latency	Polling interval dependent	"Amazon SNS typical latency is under 30 milliseconds"	"Typical latency is about half a second"
Typical use	"Decoupling microservices, buffering requests, processing tasks asynchronously"	"Fanout notifications, pub/sub messaging, mobile push notifications"	"Event-driven architectures, real-time stream processing, cross-account event sharing"

Three consequences follow directly.

EventBridge gives no ordering guarantee. AWS states it plainly: "EventBridge does not provide message ordering guarantees, instead delivering events to targets in an arbitrary order." If your consumer assumes that an order.updated event arrives after order.created, that assumption is yours to enforce with a version number or a sequence check, not the bus's to keep. EventBridge Pipes is the exception; AWS says Pipes "will maintain the order of events received from an event source when sending those events to a destination service."

Latency differs by an order of magnitude between the two push services. Under 30 milliseconds for SNS against about half a second for EventBridge is not a rounding difference in a user-facing path. Use SNS or a direct call where a person is waiting; use EventBridge where a system is.

And AWS's own guidance is stale and internally inconsistent on this exact point. The decision guide for choosing between the three services carries a "Last updated" date of 31 July 2024. Its summary table lists EventBridge under "At-least-once delivery", while the prose section on the same page claims it "ensures exactly-once processing, guaranteeing that each event is processed only once by the target services." Those two statements cannot both hold, and the exactly-once claim also conflicts with the EventBridge Scheduler FAQ, which states that Scheduler "provides at-least-once event delivery to targets." Build for at-least-once and make your consumers idempotent. Treating a vendor's marketing prose as a delivery contract is how duplicate charges reach customers.

For scale, EventBridge carries a 99.99% monthly uptime commitment and AWS says default quotas "can be increased to process hundreds of thousands of events per second," with over 90 AWS services available as event sources, over 15 as targets and over 45 SaaS integrations.

The five signals that say move

Event-driven architecture is not a maturity level. It is a trade: you exchange a call stack you can read for a coupling you can survive. Make the trade when the coupling is actually hurting.

Signal you can measure	What it means	Move to events?
A synchronous chain of three or more internal service calls sits in a user-facing request	Every downstream failure is a user-facing failure, and latency compounds	Yes, for the steps after the write that the user does not wait on
One slow consumer forces you to scale an unrelated upstream service	Throughput is coupled where the business logic is not	Yes, a queue between them is the cheapest fix available
Adding a new consumer of an existing business fact requires changing the producer	The producer has become a distribution hub it was never designed to be	Yes, this is the case EventBridge exists for
The workflow is a linear sequence with retries and human approval	You need durable execution and visibility, not routing	No, use a workflow service; events add debugging cost with no payoff
Consumers depend on strict ordering of related records	EventBridge has "no ordering guarantees"	Not on a plain bus; use a FIFO queue or Pipes, or carry a version
The team has no on-call rotation and no dead-letter queue alarm	Failures will be silent, and silent failures in async systems surface as customer complaints	Not yet, fix operations first

The last row is the one most often skipped. In a request/response system, a failure is a 500 that someone sees. In an event-driven system, a failure is a message sitting in a dead-letter queue that nobody has alarmed on. The architecture does not create the outage; it changes who finds out first.

How to move without a rewrite

The migration pattern that works is boring and incremental.

Pick one write path that already has downstream side effects, typically order placement, user signup or payment confirmation.
Keep the synchronous write. Publish an event after it commits, and change nothing else. At $1.00 per million custom events, running the bus in shadow mode for a month costs less than the meeting to approve it.
Move exactly one side effect, usually a notification, onto the event. Keep the old code path behind a flag until the new one has run clean for a full billing cycle.
Add a dead-letter queue and an alarm on it before the second consumer, not after the fifth.
Make every consumer idempotent against a business key, because delivery is at-least-once whatever a documentation page claims.
Only then retire the synchronous call.
Add archive and replay once you have more than two consumers. At $0.10 per GB processed and $0.023 per GB stored, replay is cheap insurance against a consumer bug that silently dropped a day of events.

Testing is where most of these migrations stall, because the flow no longer fits in a single test. The current tooling picture is covered in our guide to AWS serverless integration testing with LocalStack and the Step Functions TestState API; the short version is that AWS has marked its old Step Functions Local emulator unsupported and now points teams at the TestState API, which costs nothing extra on top of Step Functions.

Who operates it afterwards

An event-driven system moves work from build time to run time. Somebody has to own the dead-letter queues, the replay runbook, the schema changes and the on-call rotation, and that ownership question decides more migrations than any architecture diagram.

Dr. Werner Vogels, VP and CTO at Amazon.com, framed the direction of travel for engineering teams in his 2026 predictions with a line that lands here: "Professional developers will soon become renaissance developers." In practice that means the person who writes the consumer is increasingly the person who gets paged when it stops consuming. Teams that split those roles across a build squad and a separate operations group find that event-driven systems expose the seam quickly.

Budget for it honestly. The messaging bill for a mid-sized service is under $200 a month at published rates. The operating cost is one engineer's attention, indefinitely.

India-specific considerations

Two points matter specifically for teams building and running from India.

Event payloads are a data-residency decision with a price attached. Under the Digital Personal Data Protection Act 2023, a business event that carries a customer name, phone number or address is personal data, and replicating that event to a second AWS Region moves it. EventBridge charges $1.00 per million events for cross-Region replication, so the residency choice is visible on the invoice as well as in the compliance register. The design answer is usually the same as the cost answer: put an identifier in the event and keep the personal data in the Region where it was collected, so the replicated stream carries no personal data at all. That single decision removes a class of cross-border argument and cuts the replicated payload size at the same time.

Second, the free tiers matter more at Indian price points than they do in a US budget. SQS gives 1 million free requests a month, EventBridge Scheduler gives 14,000,000 free invocations a month, Schema Discovery gives 5 million free ingested events a month, and AWS management events on the bus are free. A team running a small production workload and three non-production environments can keep most of the event layer inside those allowances if the environments are sized deliberately rather than cloned from production. Teams already running this discipline on compute will recognise the pattern from cutting cloud spend for Indian teams.

What eCorpIT builds here

We design and run event-driven and serverless architectures on AWS for teams that have hit the coupling wall in a request/response system. A typical engagement runs in three phases. First, an assessment of the current call graph to find the synchronous chains that are actually causing incidents, with the published rate card applied to your event volumes so the cost is a number rather than a worry. Second, an incremental migration of one write path at a time, with dead-letter queues, alarms and idempotency built in before the second consumer rather than after the fifth. Third, handover: runbooks for replay, a schema change process, and the observability needed to answer "where did that event go" in minutes.

eCorpIT was founded in 2021 and is based in Gurugram. We are an ISO 27001:2022 certified, CMMI Level 5 assessed, MSME registered organisation, and an AWS partner. Our teams are senior-led and multi-disciplinary, which matters here because event-driven work sits across application code, infrastructure and operations rather than inside any one of them. We design applications aligned with DPDP Act requirements, including the residency decisions described above.

This work pairs naturally with API integration and modernisation when the event layer has to coexist with existing synchronous contracts, with cloud migration and modernisation when the move is part of a larger platform change, and with multi-Region resilience and disaster recovery when cross-Region replication is on the table.

FAQ

How much does Amazon EventBridge cost per million events?

Custom events published to an event bus are billed at $1.00 per million, and AWS management events at $0.00 per million. Each 64 KB chunk of payload counts as one event, so a 256 KB event bills as four. EventBridge Pipes is $0.40 per million requests and API destinations are $0.20 per million events.

Does EventBridge guarantee event ordering?

No. AWS states that EventBridge "does not provide message ordering guarantees, instead delivering events to targets in an arbitrary order." EventBridge Pipes is different: AWS says Pipes maintain the order of events received from a source. If you need strict ordering elsewhere, use an SQS FIFO queue or carry a version number in the event.

Is EventBridge delivery exactly-once or at-least-once?

AWS's decision guide is inconsistent, listing at-least-once in its comparison table and exactly-once in the prose on the same page, which was last updated on 31 July 2024. The EventBridge Scheduler FAQ states at-least-once delivery. Build consumers to be idempotent against a business key and the ambiguity stops mattering.

What latency should we expect from EventBridge?

AWS states that "typical latency is about half a second" and notes this can vary. Amazon SNS is much faster, with typical latency under 30 milliseconds. For a user-facing path where someone is waiting on the result, that difference matters; for system-to-system work it usually does not.

When should we not move to event-driven architecture?

When the workflow is a linear sequence with retries and approvals, a workflow service fits better than a bus. When consumers depend on strict ordering, a plain event bus is the wrong tool. And when the team has no on-call rotation or dead-letter queue alarm, fix operations before adding asynchronous failure modes.

What does archive and replay cost?

Archive processing is billed at $0.10 per GB and archive storage at $0.023 per GB per month, with replayed events charged at the custom-event rate of $1.00 per million. Archiving 50 million events of roughly 6 KB each works out to about 286 GB, or $28.61 in processing and $6.58 a month in storage.

How does DPDP affect event payloads?

A business event carrying a customer name, phone number or address is personal data under the Digital Personal Data Protection Act 2023, so replicating it across Regions moves personal data. Publishing an identifier instead, and keeping the record in the Region where it was collected, removes the cross-border question and reduces the billed payload size.

Which free tiers apply to the event layer?

Amazon SQS includes 1 million requests free each month across all Regions except GovCloud. EventBridge Scheduler includes 14,000,000 invocations a month at no cost. Schema Discovery includes 5 million ingested events a month before charging $0.10 per million, and the Schema Registry itself is free.

How eCorpIT can help

If a synchronous chain in your product is causing incidents, we can put a number on what moving it would cost and what it would fix, using your own event volumes against the published AWS rate card rather than a generic estimate. Our senior engineering teams run the migration one write path at a time, with dead-letter queues, alarms and idempotency in place before the second consumer, and hand over the replay runbooks and schema process at the end. eCorpIT is an ISO 27001:2022 certified, CMMI Level 5 organisation based in Gurugram and an AWS partner. Talk to us about an architecture review.