Genesis Cloud

Provisions NVIDIA GPU instances (H100, H200, B200, RTX 4090/3090/3080) on-demand with per-GPU hourly billing, supporting single-GPU to 8-GPU node configurations. Instances are allocated from Genesis Cloud's renewable-energy data centers across Europe and North America, with no minimum commitment for single-GPU SKUs but full-node (8x GPU) minimum for HGX multi-GPU configurations. Billing is metered hourly with no setup fees or egress charges.

Offers reserved instance pricing for committed capacity over longer periods (details not fully documented), allowing users to lock in lower per-hour rates compared to on-demand pricing. Reserved instances are allocated from the same infrastructure as on-demand but with upfront or monthly commitment terms. Pricing structure and commitment periods not detailed in available documentation.

Offers inference endpoint capability (mentioned but not detailed) for deploying trained models for real-time or batch inference. Endpoints are deployed on GPU instances and are accessible via HTTP/REST API. Specific features (auto-scaling, load balancing, model versioning, A/B testing) not documented; unclear if endpoints are managed service or manual instance management.

Offers MLOps platform (mentioned as solution but not detailed) for orchestrating training pipelines, managing experiments, and tracking model artifacts. Platform capabilities, integration with Genesis Cloud infrastructure, and supported frameworks not documented. Unclear if this is a proprietary platform or integration with third-party tools (Kubeflow, MLflow, Weights & Biases).

Offers data management platform (mentioned as solution but not detailed) for versioning datasets, tracking data lineage, and managing data pipelines. Platform capabilities, integration with Genesis Cloud storage, and supported data formats not documented. Unclear if this is a proprietary platform or integration with third-party tools (DVC, Pachyderm, Lakehouse platforms).

Enables users to select and deploy GPU instances across Genesis Cloud's data centers in Europe (Norway, France, Spain, Finland), North America (USA, Canada), and UK (Great Britain). Each region has different GPU availability (e.g., B200 only in Norway, RTX 3090 only in Norway/Netherlands), and instances are deployed to Tier-3 ISO 27001-certified data centers with 99.9% uptime SLA and 100% renewable energy. Users select region at provisioning time; no automatic multi-region failover or load balancing documented.

Provides 400 Gbps non-blocking RDMA (Remote Direct Memory Access) networking between GPUs within a node and across nodes in the same region, enabling low-latency, high-throughput communication for distributed training. RDMA is included at no additional cost and is optimized for collective communication patterns (all-reduce, all-gather) used in data-parallel and model-parallel training. Network is non-blocking, meaning no bandwidth contention between node pairs; latency and throughput characteristics not specified.

Provides persistent block storage (SSD or HDD) attachable to GPU instances at $0.04/GB/month, enabling durable storage of training datasets, model checkpoints, and application state across instance restarts. Storage is provisioned separately from compute and can be resized or migrated between instances. Storage type (SSD vs. HDD) affects I/O performance but pricing is uniform; IOPS and throughput specifications not documented.

Provides high-speed shared file storage at $0.10/GB/month, integrated with VAST Data platform for parallel I/O optimization across multiple GPU nodes. File storage is accessible from all instances in a region via NFS or proprietary protocol, enabling concurrent reads/writes from 8+ GPU nodes during distributed training. VAST Data integration optimizes metadata operations and parallel I/O patterns; specific throughput and latency not documented.

Provides S3-compatible object storage at $0.03/GB/month for storing training datasets, model artifacts, and inference outputs. Storage is accessed via standard S3 API (boto3, AWS CLI, etc.), enabling seamless integration with existing ML workflows. No ingress/egress fees charged, reducing total cost of ownership vs. AWS S3. Storage is region-specific; cross-region replication not documented.

Enables creation of point-in-time snapshots of block storage volumes at $0.02/GB/month, allowing users to capture training state, datasets, or application data for backup, recovery, or cloning. Snapshots are stored separately from active volumes and can be used to create new volumes or restore to existing volumes. Snapshot creation time and restore time not documented.

Provides public IPv4 addresses (paid) and IPv6 addresses (free) for instances, with multi-Gbps redundant internet connectivity and zero traffic charges for ingress and egress. Instances are assigned public IPs at provisioning time; no NAT or load balancing service documented. Traffic charges are explicitly waived for both public internet and RDMA networking, reducing total cost vs. AWS/Azure/GCP.

Provides web console and API (details not documented) for creating, managing, monitoring, and terminating GPU instances. Users can provision instances, configure storage, manage networking, and monitor resource usage through the console or programmatic API. No CLI tool or SDK documented; API patterns, authentication, and rate limits unknown.