Selectel added high-end servers for model training and expanded its image catalog

Selectel Adds Flagship Servers for Model Training and Expands Image Catalog

Selectel wrapped up February and demonstrated how its infrastructure is evolving to meet growing demand for AI workloads. The main updates include new dedicated servers for training large models, fresh images in the AI marketplace, and several notable changes to the cloud Kubernetes platform.

Servers for Training

The flagship innovation is the dedicated GL8-B300-HGX-25GE server based on the NVIDIA HGX B300 platform. The company calls it the most powerful configuration in its current lineup: the solution is designed for AI training, fine-tuning, and industrial inference of large language models. Selectel specifically emphasizes the total video memory volume of 2.3 TB — a level needed not for pilots, but for heavy pipelines, long context, and large batches, where the bottleneck becomes not only GPU but also the stability of the entire infrastructure.

"NVIDIA HGX B300 is a flagship platform for large-scale AI workloads."

For teams that don't need top-tier cluster-level capability, the company has also added more practical options. The AR45G-NVMe-H200 configuration is positioned as a compact and more affordable server for inference, prototypes, and small-scale training. Another option — AR45G-NVMe-RTX6000 — is designed for scenarios where 96 GB of VRAM is sufficient: LLM pilots, computer vision tasks, graphics, and applied workloads. In parallel, Selectel expanded its storage lineup by adding 28 TB HDDs for archives, backups, and logs.

Marketplace and Cloud

New ready-to-use images have appeared in Selectel's AI marketplace, covering different stages of the ML cycle — from experiments to quality control and data labeling. This is not simply expanding the catalog for quantity's sake: the company is adding tools that help build a more complete ecosystem around a model, rather than just running inference in a container. For engineering teams, this is also a way to quickly assemble a working environment without extra manual integration.

• Aim — for experiment tracking and comparing model tuning results.
• Lobe Chat — an interface for working with LLM with emphasis on plugins and ready community solutions.
• Evidently AI — model quality monitoring and metrics on which model updates can be automated.
• Xtreme1 — data labeling service, including scenarios with lidar and object classification.

Separately, Selectel has deployed GPU L4 with 24 GB of memory to the cloud. The company calls it a universal card for AI/ML tasks, video processing, streaming, and VDI, as well as for applied scenarios like audio-to-text transcription. This launch is important not only for AI teams: L4 is often chosen where a balance is needed between cost, energy efficiency, and sufficient performance without switching to expensive training-grade accelerators. For teams with limited budgets, this is a particularly practical option.

Kubernetes and System Layer

A significant infrastructure update affected Managed Kubernetes. Selectel added full support for the ephemeralStorage resource in Cluster Autoscaler and Karpenter. The practical value here is straightforward: when scaling the cluster, the system now more accurately understands how much local temporary storage new Pods actually need. Previously, in scenarios where the cluster had no nodes with explicitly defined ephemeralStorage, autoscaling could calculate requirements inaccurately. For teams with data-processing and AI workloads, this eliminates an unpleasant class of errors at the scale-up stage.

Several other changes relate to control and the basic system layer. Managed Kubernetes now features audit logs for key cluster and node group operations — this simplifies change analysis, security, and compliance. The interface also opened up viewing private DNS configuration, and for node groups updated the display of User Data. At the OS level, Selectel released SELECTOS 1.3 with package updates to Debian 12.13, closing 177 vulnerabilities, and adding a container image. Additionally, the company launched Astra Linux for A-DC servers — this is a move toward customers who need a certified secure environment.

What This Means

The February update package shows that Selectel wants to be more than just hardware rental, but a full-fledged AI infrastructure platform. At the top, the company expands its catalog of ready-made ML tools and cloud GPUs; at the bottom, it strengthens Kubernetes, base operating systems, and the compliance layer. For teams building their own AI services, this reduces manual assembly and accelerates the path from pilot to production deployment.