Six NVIDIA GPUs in Switzerland. Datasets and secrets in-country.

VM for an isolated data scientist or a long job. Kubernetes for sharing, autoscaling and addons. Same card catalogue, same jurisdiction. On the managed cluster: GPU Operator (optional addon) installs NVIDIA drivers and the device plugin. HAMi, also optional, virtualises one card across pods and requires GPU Operator. Hikube spreads workers across Geneva, Gland and Lucerne: you do not place nodes DC by DC. Platform detail: the Kubernetes platform.

S3 object storage in Switzerland. Access secrets in Vault. Images in Registry as a Service, not GHCR or ECR by default. A us-east-1 checkpoint breaks job sovereignty. Products: S3 object storage, the secrets vault, the image registry.

Hidora SA operates the DCs, GPU attachment, the Kubernetes control plane if you use it, and FR/EN support. You keep images, checkpoints, training code, pod topology spread and the VM vs node-group choice. Terraform integration in preparation: today, Hikube API and console. Not CNCF AI Conformance.

Inference in Switzerland on data that must not leave. Fine-tuning on internal datasets. Batch jobs on VMs. Card sharing via HAMi on Kubernetes. This is not a worldwide catalogue of thousands of GPUs, and not an opaque ML PaaS.

Six published SKUs, not a farm. L4 24G (165 CHF/month): inference and vision on small models. L40S 48G (760): larger inference, rendering. A100-80 (1495): fine-tuning and training. RTX 6000 Pro 96G (1550): more VRAM, workstation profile. H100 80G (1999): heavier training. H200 141G (2499): more memory for long contexts. We do not publish tokens/s or IOPS. The engineer sizes during the trial. Catalogue: the GPU catalog.

Checkpoints, weights and cold sets: S3 object storage in Switzerland. Access secrets in Vault, not a US KMS. For scratch disk on the VM or pod: network NVMe volumes, always geo-redundant across Geneva, Gland and Lucerne, synchronous (real time) or asynchronous (improved performance) replication. Sync and async do not share the same RPO; we publish neither a number nor IOPS. No unreplicated local block volume. Products: S3 object storage, block storage, the secrets vault.

No. Hikube automatically distributes workers, including GPU node groups, across Geneva, Gland and Lucerne. You do not choose the DC. Topology spread and PodDisruptionBudget for pods stay yours. HA spare-capacity restart in normal mode excludes GPUs: a card is not an interchangeable CPU worker. Platform detail: the Kubernetes platform.

No. Terraform and Cluster API integration in preparation. Today: Hikube API and console. kubectl, Helm and S3 clients stay. A tender that requires Terraform in production today does not yet match the catalogue. An engineer says so when framing the trial.

If the model or dataset touches Swiss health data, the frame is nFADP, not HDS. Hikube is not an HDS host and is not HDS-certified. Datasets, checkpoints and backups stay in Geneva, Gland and Lucerne. A US bucket breaks the file. Healthcare page: healthcare. Engineer support FR/EN.