Since our first visit in 2018, the conversation has shifted, and clearly so: from “why streaming?” to “how do we run this efficiently?” Add AI to the picture and you can feel where the next few years are heading. But this year, in London on the 19th and 20th of May, the focus was still maturity and operational efficiency.
Four of us flew in from Stockholm with a simple plan. Three hand-picking the best talks across the two days, and one lonely warrior working the expo floor, shaking hands and listening in on the gossip.
Here is what we saw and heard.
Kafka and the ecosystem
Kafka is still the beating heart of the event, but the ecosystem around it is taking more of the spotlight. That tracks with where streaming and event-driven architecture have landed. Most enterprises now run some flavor of it, so the question has moved from “should we?” to “how do we make it work efficiently across the whole stack?” The sessions followed: governance, multitenancy, security, and running clusters reliably at scale. Change Data Capture even got its own dedicated track this year, a sign the community has plenty left to say about it.
Best talk? Hard to pick. Most of the strongest content was not about shiny new features but about the unglamorous reality of operating streaming infrastructure day to day. One that stood out was Marcos Prado’s Life as a Kafka Admin: Lessons from Running 30+ Clusters in Production. Great energy on stage, and a sharp, honest take on what it actually takes to keep a fleet of Kafka clusters healthy in production.
Marcos Prado, Kafka Specialist, Data Streaming
The state of Apache Flink
The Kafka and Flink pairing came up again and again. The mood was consistent: powerful, but nothing comes for free. Sessions ranged from “getting started” basics all the way to running Flink across multiple clusters in production, which is a fair snapshot of where the community sits right now. Still maturing, but clearly going somewhere.
Our own read: for stateful stream processing, joins, windowing, enrichment, Flink is becoming the natural default next to Kafka, and for good reason. But “default” is not the same as “free.” It adds another distributed system to operate, another set of failure modes to learn, and a real learning curve for teams still getting comfortable with Kafka itself. For plenty of use cases, Kafka Streams or plain consumers are still the lower-risk choice. The gap is operational maturity, not capability.
AI and streaming
AI was woven through much of the programming, as you would expect in 2026. The central theme: real-time data and streaming pipelines are becoming critical infrastructure for AI and agentic workloads. It came up across keynotes and breakouts alike. Worth keeping a critical eye, though. Some of this is real emerging practice and some is positioning. The interesting question is how much of the “streaming for AI” story is actually running in production versus living on slide decks.
The most grounded conversation came from a town hall on AI and how we work with it day to day. Multi-agent systems were the dominant theme, with practitioners moving away from single-model prompting toward orchestrated pipelines where specialised agents handle planning, implementation, and review in sequence. Local models came up as a serious alternative to cloud APIs, especially where internet connectivity simply is not an option. The example that stuck with us: transaction monitoring on sensitive financial data, where no external connection can be permitted at all. That is exactly the kind of constraint our customers live with.
Private cloud and the IBM factor
Private cloud got serious airtime in the keynote, and that is no accident. IBM’s acquisition of Confluent closed on the 17th of March this year, an all-cash deal worth roughly 11 billion dollars, and the strategic centre of gravity is visibly shifting. For years, Confluent pushed hard toward its managed cloud service. This year the tone was different: self-managed and private deployments were front and centre.
The reason is not hard to spot. Geopolitical instability, tightening data sovereignty rules, and growing unease about where sensitive workloads actually run have companies, public sector bodies, and governments all rethinking the cloud-first default. IBM’s hybrid and on-prem heritage slots neatly into that mood. The real question is what the combined roadmap looks like twelve months from now.
Expo conversations
Many of the booth conversations centered on Kafka control planes and removing operational complexity. That is what you would expect from a room full of Kafka developers who would rather not babysit clusters. But for a consultancy like Irori, working with enterprise and government customers in Sweden, control planes are not the tech we touch firsthand.
So we usually asked two questions:
- Is it open source?
- Can you run it on-prem?
The on-prem question is not academic. When you build for regulated, high-security environments, on-prem is often the only option on the table. The open source question? We just like it. 🙂
So even though Confluent is not looking to sell much in the way of open-source products, it was good to hear the renewed focus on private cloud and on-prem. Hoping to see more of that through the year.
Final thoughts
If we had to compress two days into a single line: the streaming world has stopped asking whether to do this and started asking how to run it well, on its own terms and, increasingly, on its own infrastructure. The novelty has worn off, in the best possible way. What is left is the engineering, and that is the part we enjoy.
Would we recommend the trip? Yes, especially if your work sits closer to running systems than to greenfield experiments. The operational talks and the expo floor were worth the flights on their own.
If you are working through any of this, streaming at scale, the pull toward private cloud and on-prem, or simply making Kafka boring and reliable in production, we are happy to compare notes. Irori Integration does this for a living.
Read our takeaways and top 3 sessions from Current 2025.
