Kubernetes has been called a lot in the past decade: the cloud operating system, the universal control plan, the large layer of abstraction. And that won this reputation. Kubernetes tamed chaos of containers, gave us a common language for infrastructure and has become the backbone of the native movement of the cloud.
But today we are looking at a new border: AI-NATIVE APPLICATIONS. Traine of massive models through GPU clusters. Execution of distributed inference pipelines. Serving responses to low latency at the edge. The management of data pipelines is as critical as the calculation itself. And suddenly, our faithful Kubernetes hammer looks a little less suitable for the AI nail.
Which raises the question: AWe try to settle down square kubernetes in the round hole of the AI-Native applications?
Kubernetes: the control plan that won
Let’s give Kubernetes its due. Born in Google and Open in 2014, Kubernetes was designed to plan and orchestrate stateless microservices. It sums up the disorderly details of the place where a container works, how it evolves and how it connects to other services. With its extensibility – personalized resources, operators, controllers – Kubernetes has grown up to orchestrate not only workloads, but the entire ecosystem that surrounds them.
This flexibility is the reason why Kubernetes won. This is now the de facto standard for native platforms in the cloud. If you build microservices, you almost certainly perform them on Kubernetes, whether on site, in the cloud or via a managed service.
But the aitic workloads are not microservices. And that’s where tension begins.
Why does AI do not correspond perfectly
The workloads of the AI underline the Kubernetes in a way for which it has never been designed.
First, there is material planning. Kubernetes was designed for processor and memory as primary resources. GPUS? Tpus? Other accelerators? These are bolted, awkwardly represented as extensive resources. The planning of GPU work effectively – and equitably – is a completely different match.
Second, Types of jobs. Kubernetes thrives on stateless services and short -term work. AI workloads are often long, with state and spread over hundreds or thousands of nodes. The formation of an LLM is not the same as the service of a web API.
Third, data severity. AI workloads are not only a question of calculation. They are counting on massive data sets that must be mixed, staged and disseminated. Kubernetes does not natively manage this complexity.
Finally, Latence sensitivity. The inference workloads can be brutally sensitive to milliseconds. Abstractions that make Kubernetes so powerful can also introduce friction that AI teams cannot afford.
Bypass
Of course, the industry did not stop. Many projects work to make Kubernetes more suited to AI.
- Kubeflow has become the reference frame for automatic learning pipelines on Kubernetes.
- Ray on K8s And Kuberay Bring workloads AI distributed in the cluster.
- Volcano focuses on the planning of computer work by lots and high performance.
- Cloud suppliers are building all their own AI-On-Kubernetes offers with personalized operators and GPU planners.
These solutions work. But too often, they feel like complementary modules – bolted adapters on Kubernetes rather than on the capacities for which it was designed. It’s like putting a new transmission in a sedan and calling it a racing car. Will it get around the track, but was it really built for that?
What a native needs
So what would a control plan designed for native AI applications look like?
It would start with GPU- and accelerator-premier planning. Not a reflection afterwards, but at the heart of the system.
He would integrate Data pipelines as first class concern. Not just pods and volumes, but streaming, fragment and high speed chatting.
He would manage Distributed training jobs Native, understand how to orchestrate thousands of GPUs on several clusters with resilience.
It would optimize Large -scale inference – Autoscanic set not for the use of the processor but for the competition, the latency and the load of the model.
And it would be Politics and costBecause cloud bills for AI are already shocking. A real a-native control plan would apply the railings against the flight GPU work before your CFO strikes.
Can Kubernetes bend without breaking?
Some maintain that Kubernetes can evolve and evolve. After all, it was not designed to execute databases either, and the operators and the CRDs made it possible to do so. With enough extensions, Kubernetes could also become the AI control plan.
The counter argument? Kubernetes is downright optimized for microservices. Renovating it for AI can always be unnatural – more like adhesive tape than design. Ai-native workloads can be better served by systems specially designed as rays, mosaic or even orchestrators owned by cloud sellers.
My intuition? We will see a hybrid future. Kubernetes will remain the control plan for corporate infrastructure – the place where compliance, networking and security policies live. But the orchestrators specific to AI will be seated alongside its side, optimized for training and inference. The challenge will be to integrate both without creating even more complexity.
The platform engineering angle
For platform teams, the debate is not academic. Their work is to hide all this complexity from developers. Whether Kubernetes evolves to manage the workloads of the AI or that we adopt new orchestrators, the key is to provide golden paths where the developers do not care what is under the hood.
This means building IDPS (internal developer platforms) Who manage GPUs, data and pipelines transparently. Developers must request training work or an inference termination point without having to understand if Kubernetes, Ray or something else is the heaviness.
In this sense, the platform engineering movement can be the bridge, which makes Kubernetes “good enough” for AI in the company by wrapping it in abstractions.
Taking Shimmy
I have been in this industry for a long time to see this model before. When Kubernetes was released for the first time, it was not an excellent choice for applications with condition, for databases or for the service jersey. But the ecosystem has adapted. The operators, the CRDs and the sidecars have folded Kubernetes in the directions for which it has never been designed.
So, can Kubernetes look again for AI? Maybe. But here is the difference: the The AI wave moves faster that everything we have seen before. We may not have the luxury of waiting for incremental ecosystems. The rhythm of model training, the demand for GPU clusters, the need for inference at the edge – all of this goes beyond the ability of Kubernetes to evolve.
My catch: Kubernetes will play a role. It’s too rooted not to. But it may never be the perfect cut. Instead, we will see a new generation of natives control planes have increased – and the challenge will be to sew them in the world of Kubernetes in which we already live.
Closing reflections
Kubernetes deserves his crown as a universal native Cloud control plan. But the aitic workloads are a different beast. They do not integrate perfectly into the Kubernetes model, regardless of the number of extensions we are launching.
The future may not be to force square Kubernetes in the round hole of the AI-Native applications. Instead, it is a question of determining where the Kubernetes belong to the AI era – and where we need new abstractions.
Because the truth is that Kubernetes does not have to do everything. He just has to do his job well. And if the AI-Native applications require something new, then maybe the most native thing in the cloud that we can do is to adopt this evolution.
Related
