KubeCon EU 2025: 10 Talks That Illustrate Current Kubernetes Trends

KubeCon EU 2025 is over, and the CNCF has made all the talks available on YouTube. This is a good chance to go back and see what was shared, especially the main cloud-native - kubernetes trends from this year. The FikaWorks team watched many of the sessions and picked 10 talks that we thought were especially interesting and useful. These talks are not ranked in any particular order—instead, we grouped them by theme. From platform engineering to AI, this list shows some of the talks we think are most worth your time.

Platform engineering:

Is 2025 the year when Jeff Bezos’ API mandate finally becomes a reality? Judging by the number of talks focused on multi-cluster setups and Kubernetes API extensions, it certainly seems that way. Projects like KCP are gaining traction, sending a strong signal that Kubernetes is no longer just about pods and containers. Instead, it’s increasingly seen as a powerful and flexible API platform that can serve as the backbone of modern enterprise infrastructure.

One great talk that explores this is:

• Extending Kubernetes Resource Model (KRM) Beyond Kubernetes Workloads - Mangirdas Judeikis, Cast AI & Nabarun Pal, Independent

One of the key advantages of using a Kubernetes-style API is the ability to build operators, helping organizations move closer to infrastructure hyper-automation. In this talk, Nick Young shares a lot of practical insight into the common issues you might face when writing controllers and operators. A surprising (and encouraging) moment came when he asked the audience how many had already written an operator—about 60–70% raised their hands. It’s a strong sign that the shift toward automation using operators is well underway. At FikaWorks, we were especially glad to see this, as we’ve been advocating for the operator pattern since the early days of Kubernetes.

• Don’t Write Controllers Like Charlie Don’t Does: Avoiding Common Kubernetes Controller Mistakes - Nick Young, Isovalent at Cisco

One of the long-time goals for many platform teams is having the ability to manage everything from a single, centralized place. In the following talk, Marvin Beckers and Stefan Schimanski present a proof of concept for adding multi-cluster support to the well-known controller-runtime library. If this idea gains traction in the community, it could be a real game changer for platform engineering, making it much easier to build organization-wide control planes.

• Dynamic Multi-Cluster Controllers With Controller-runtime - Marvin Beckers & Stefan Schimanski

If you’re wondering whether organization-wide operators can actually handle that scale, Tim Ebert’s talk is a must-watch. He dives into how to implement resource sharding for Kubernetes operators, which allows them to scale horizontally and manage large amounts of resources more effectively.

• Beyond the Limits: Scaling Kubernetes Controllers Horizontally - Tim Ebert, STACKIT

If both of these ideas continue to mature and move beyond the proof-of-concept stage, they could significantly expand what’s possible with Kubernetes and how it’s used at scale.

Another important talk to catch is the one about Helm 4, which is planned for release later this year. This new version brings several improvements, including better support for using Helm as a library—not just as a command-line tool. For more details on what’s coming and how it can improve your workflows, we recommend watching the presentation by Matt Farina and Andrew Block.

• Helm 4 You - Matt Farina, SUSE & Andrew Block, Red Hat

The final talk we’d like to highlight in the platform engineering section comes from engineers at Bloomberg, who shared their experience running Trino—a distributed SQL query engine—on Kubernetes. It’s always encouraging to see Java-heavy workloads like Trino successfully integrated into the Kubernetes ecosystem. This was made possible through the use of custom resources, an operator, and established cloud-native tools like OPA and Envoy. It’s a great example of the operator pattern being applied in the data analytics space. Talks like this are a good reminder that with the right approach, even complex software can be made Kubernetes-native.

• Trino and Data Governance on Kubernetes - Sung Yun & Aki Sukegawa, Bloomberg

AI:

We start the AI section with a talk from Justin Santa Barbara and Walter Fender of Google, who showed how large language models can go beyond code suggestions to generate production-grade Kubernetes controllers. By breaking the work into smaller steps—such as generating KRM types and reconcilers—and using fine-tuned models with custom tooling, their team built controllers for over a thousand Google Cloud resources, something not easily achievable with tools like Terraform. This talk offers a practical look at how AI can support real-world infrastructure automation code.

• AI Beyond Autocomplete: Using LLMs To Create 1000 Kubernetes Controllers
  • Justin Santa Barbara & Walter Fender, Google

AI depends heavily on high-performance computing, but connecting HPC systems with Kubernetes isn’t easy. In this talk, Dennis Marttinen from Aalto University introduces Supernetes, a tool that maps Slurm-managed HPC jobs directly to Kubernetes using thousands of virtual kubelets. Tested on LUMI, one of the world’s top supercomputers, it shows that Kubernetes is ready for HPC—and that bridging the two is key for future AI workloads.

• Thousands of Virtual Kubelets: 1-to-1 Mapping a Supercomputer To Kubernetes With Supernetes - Dennis Marttinen, Aalto University

GPUs are powerful—but expensive—so making the most of them is critical for running AI workloads at scale. In this talk, Yuan Chen (NVIDIA) and Chen Wang (IBM Research) offer a hands-on guide to benchmarking GPU performance in Kubernetes. They walk through practical tools and benchmarks for training, inference, and stress testing, using frameworks like NVIDIA Triton, MLPerf, and fmperf. This session provides valuable insights for teams looking to optimize GPU utilization and improve the efficiency of AI workloads in Kubernetes environments.

• A Practical Guide To Benchmarking AI and GPU Workloads in Kubernetes - Yuan Chen, NVIDIA & Chen Wang, IBM Research

Last but not least, don’t miss the talk from FikaWorks collective member Andrea Giardini, who shared how Kubernetes is used to power AI systems for wildfire prevention. The session covers processing large volumes of satellite and environmental data, using GPUs effectively, and building a reliable, scalable platform. It’s a strong example of how Kubernetes can support real-world, high-impact use cases beyond traditional IT.

• Kubernetes and AI To Protect Our Forests: A Cloud Native Infrastructure for Wildfire Prevention - Andrea Giardini, Crossover Engineering BV

KubeCon EU 2025 made it clear:

Kubernetes is no longer just the foundation of cloud infrastructure — it’s becoming the platform where organizations build everything from internal tooling to AI-powered systems. Whether it was pushing the limits of multi-cluster controllers, automating infrastructure with operators, or bridging HPC and AI, this year’s talks showed just how much innovation is happening across the ecosystem.

At FikaWorks, we’re always looking for practical ideas that teams can build on,and we hope this list helps you catch up on the most relevant trends.  Let us know if any of these talks sparked new ideas for you!