Apache Kafka & the Mainframe (IBM Confluent Acquisition)

why IBM acquired Confluent

Author

Stanislav Kozlovski
April 02, 2026

The news of the month (March) in the Kafka space has been that IBM finished their Confluent acquisition. Three months earlier than expected. There was no shortage of takes on the matter.1

Today I’m here to bring you a take you won’t read anywhere else.

Confluent and Kafka were frequently used as a way to liberate data from the Mainframe and reduce lock-in. IBM wanted to control that chokepoint.

I know what you’re thinking - “Wtf Stan?”

I never thought I’d talk about Mainframes in my life either, but hear me out.

Mainframes 101

In case you were born in the 90s (like me), you probably don’t really know what mainframes are (like me). They’re basically these huge server boxes that come packaged with hardware and software. Nowadays it’s a monopoly run by IBM as the other vendors went bankrupt.

Nobody builds modern infra on mainframes, but a ton of old infra is running on them. Everything that was built in the 80s (banks, government systems, insurance, retail stores inventory) has a lot of its critical systems written in COBOL code running on top of these legacy systems.

New boxes are still being built, mainly to run the old code (through emulation) and support some newer workloads.3

Mainframes are awfully expensive to run. IBM sells software licenses for the proprietary software that runs on these machines.

They have usage-based billing which depends on how much you use the general processor of the mainframe (called CP).

The machine tracks the CPU cycles called MSUs (formerly MIPS) per 4hr rolling average. It then bills you for the month based on the LARGEST reading from that whole month.8

We are talking up to tens of millions $ per year for some customers. When moving to AWS can yield 50-90% cost savings, you know you’re doing something wrong!

For real, this is a massive business for IBM:

  • they enjoy 80-90% profit margins

  • they are making ~$8-$9B a year from mainframe software licensing

  • 70% of clients are GROWING their usage

  • 2025 was their stronger mainframe sales year in the last 20 years. ~50% growth

It’s, quite frankly, their cash cow. They admit so themselves.

But moving off isn’t easy.

  • Australia once spent $120M trying to migrate its social security system to a more modern Linux-based solution and failed. 120 million USD down the drain 🫢

  • TSB, a UK bank, was fined $62M for experiencing major downtime while trying to migrate off of the mainframe.4 Get it wrong and you lose a fortune when you’re running a tightly regulated business.

There’s also a massive talent drought, as COBOL programmers are pretty old and retiring.2 If you wanna see what COBOL is like, here’s what a Kafka producer looks like:

IBM actually released an SDK for Kafka that lets you run Kafka clients on z/OS (mainframes). This lets you use Kafka APIs through COBOL.
https://www.ibm.com/products/open-enterprise-sdk-apache-kafka

I think you can now see why customers are preferring to limit changes to their mainframes...

👉 where Kafka comes in

Kafka is perfect for liberating data out of a silo and into the hands of many downstream consumers.
Kafka is perfect for offloading work off of the mainframe.

It can act as a forward cache.

You CDC data out of the mainframe into Kafka (through a Sink Connector), then feed downstream apps that data.

Read-only workloads don’t load the mainframe and therefore don’t inflate your MSU costs.

Similarly, you can think about adding new applications outside, then feeding their results into the mainframe through a Source Connector.

One interesting note is the zIIP processor here where Connect runs. IBM Mainframes give people separate cores that let them run more modern stacks (Linux, Java, etc.) without paying through the nose for MSUs. I guess this is their way of softly locking people in by front-running mainframe migrations - give the customer something a tad cheaper to ever-so-slightly disincentivize their migration. This still forces users to expand mainframe footprint and pay for upgrading the hardware, so it’s a win-win.

This is why Confluent built and released Premium (i.e, proprietary, paid) IBM MQ Sink and Source connectors in 2022.

This approach saw Confluent customers like Fidelity save $3M and others reduce CPU usage per message by 50%. White papers were written on the matter, claiming a 50% cost reduction from mainframe bills.5

The Acquisition Timeline

  • Oct 8, 2025: Rumor drops Confluent is in talks of being acquired

  • Nov 3, 2025: I post “Event Streaming is Topping Out”, talking about how the industry is bound to start consolidating very soon6

  • Dec 8, 2025: IBM announces that it intends to acquire Confluent for $11B. “expected to close by the middle of 2026.”

  • March 17, 2025: IBM completes merger a few months earlier than planned.

    • Confluent experiences layoffs across HR, Finance, Sales, Marketing. ~800 people affected. Others claiming ~37% of the company.

A low-conviction prediction I’m willing to make here is that IBM won’t take the RedHat/HashiCorp route with Confluent. I foresee a much deeper integration with IBM, and I can imagine the Confluent brand not surviving after 2 years. The reason I’m saying this is:

  • Their merger-completion announcement has the strongest integration language (including day one integration) out of all its other acquisitions.

    • It explicitly frames the combo as a single operating stack (“a single, governed platform”)

  • Kafka integrates with IBM much more naturally than products like Terraform (HashiCorp) or RHEL (Red Hat)

  • Their merger completed significantly faster than the others.7

Credit to the mainframe idea goes to Justin Manchester, who introduced me to this idea from a quick informal chat over LinkedIn. We found a lot of interesting topics to discuss and recorded a podcast going over them.

We went deep on:

• why mainframes are huge cash cows people don't appreciate
• the real value of Kafka as the TCP to your data layer
• why no vendor won in killing mainframes
• AI-washing, layoffs and the ZIRP hangover in today's industry
• why is IBM interested in Confluent?
• Kafka as a forward cache that helps you migrate off mainframes
• the future of Flink, WarpStream & other Confluent products
• IBM’s continual move toward "soft" lock-in and bundling
• Databricks and Snowflake's potential moves following this deal
• what other acquisitions may follow

If you enjoyed this letter, you will love the podcast:

You can also find the transcript here, in case you want to interactively ask your favorite AI about the conversation.

You can also listen to it on:

Apache®, Apache Kafka®, Kafka, and the Kafka logo are either registered trademarks or trademarks of the Apache Software Foundation in the United States and/or other countries. No endorsement by The Apache Software Foundation is implied by the use of these marks.

1  Conduktor, Gravitee, Tinybird, Lenses, Aiven, RedPanda, FactorHouse plus a lot of individual takes. More are coming out as I write this newsletter too.

2  In 2019, the average age of a COBOL programmer was 58. 10% were retiring each year. Today it must be 65 I guess. Hopefully AI fixes this. I bet their investors hope it does (need a lot of cash to repay a ~1T valuation).

3  it’s wild to think about, but these things are generally compatible with software written for mainframes that launched way back in 1964…

4  if interested, there is a 150-page RCA here: https://www.fca.org.uk/publication/final-notices/tsb-bank-plc-2022.pdf. You know it’s bad when the government forces you to write an RCA… 😥

5  how long do u bet until this gets deleted off the website? feel free to reply to this email and we make a bet

6  I think a fair amount of people (perhaps just Confluent employees) were in disbelief when they saw this. Some mocked it. It’s a real echo chamber in there tbh.

7  for comparison, HashiCorp took 10 months (Apr 24, 2024 to Feb 27, 2025) despite expecting it to take 6 months (worded as until the end of 2024). RedHat took 9 months (Oct 28, 2018 to July 9, 2019). Confluent took 4 months (ahead of schedule). Either IBM got much better at acquisitions, or it’s in a rush to integrate. (or both)

8  This means a single spike can skyrocket your bill. Hence, there is real incentive to migrate workloads off.