DataCrunch

Provisions bare-metal NVIDIA GPU instances (A100, H100, B200, GB300) hosted exclusively in European datacenters with guaranteed EU data residency and SOC 2 Type II certification. Uses pay-as-you-go pricing model with instant activation via CLI or Terraform IaC, eliminating need for multi-region failover or data transfer compliance audits. Infrastructure ownership by European entity provides contractual GDPR compliance without third-party data processor agreements required by US cloud providers.

Provisions fixed-size GPU clusters (16x, 32x, 64x, 128x GPUs) with NVLink and InfiniBand networking for distributed training workloads. Clusters use bare-metal architecture with direct GPU-to-GPU communication via NVLink (for A100/H100) or RoCE (RDMA over Converged Ethernet) for lower-latency collective operations (all-reduce, all-gather) required by distributed training frameworks like PyTorch DDP, DeepSpeed, and Megatron-LM. Self-service provisioning via CLI or Terraform with fixed cluster sizes (not dynamic scaling) and custom pricing for enterprise deployments.

Manages batch training and inference jobs with automatic resource allocation, job queuing, and execution monitoring. Users submit job specifications (container image, resource requirements, input/output paths) and system schedules execution on available GPU resources. Supports job dependencies, retry policies, and timeout management. Abstracts away resource scheduling complexity and enables efficient resource utilization by batching jobs across multiple instances.

Native integration with NVIDIA software stack (CUDA, cuDNN, NCCL, TensorRT) and optimization for NVIDIA GPU architectures (A100, H100, B200). Instances come pre-configured with NVIDIA drivers and libraries; Verda's infrastructure is NVIDIA Preferred Partner certified, indicating validated performance and support. Enables use of NVIDIA-specific optimization tools (Nsight, NVIDIA Profiler) and frameworks (Megatron-LM, DeepSpeed) without additional configuration. Provides access to latest NVIDIA hardware (B200 Blackwell, GB300) for cutting-edge performance.

RESTful API for programmatic control of all Verda resources (instances, clusters, storage, networking, inference endpoints). Supports resource creation, deletion, status queries, and metric retrieval via HTTP requests with JSON payloads. Enables integration with custom automation tools, CI/CD pipelines, and third-party orchestration platforms. API authentication via tokens; responses include resource metadata and status codes for error handling.

Instances come pre-configured with popular ML frameworks (PyTorch, TensorFlow, JAX) and dependencies (CUDA, cuDNN, NCCL) ready for immediate training without additional setup. Supports distributed training frameworks (PyTorch DDP, DeepSpeed, Megatron-LM, TensorFlow Distributed) with optimized configurations for Verda's NVLink/InfiniBand clusters. Eliminates dependency installation overhead and ensures framework versions are compatible with GPU drivers and NVIDIA libraries.

Deploys containerized inference models as auto-scaling serverless endpoints using pay-per-request pricing. Accepts Docker containers with custom inference code, automatically scales replicas based on request volume, and exposes HTTP API endpoints. Abstracts away container orchestration and infrastructure management — users push container image to Verda registry, define endpoint configuration, and system handles scaling, load balancing, and billing per request. Supports image and audio model inference with managed endpoint templates for common model types.

Provides pre-built HTTP API endpoints for state-of-the-art image and audio models without requiring container deployment or infrastructure management. Users call managed endpoints directly via REST API with model inputs (image URLs, audio files, text prompts) and receive structured outputs. Verda handles model hosting, GPU allocation, scaling, and optimization — users only pay for API calls. Eliminates need to download model weights, manage dependencies, or optimize inference code.

Enables declarative infrastructure provisioning via Terraform and OpenTofu providers, allowing users to define GPU instances, clusters, storage, and networking as code. Verda provider translates HCL (HashiCorp Configuration Language) into API calls to provision resources, manage state, and support infrastructure versioning and reproducibility. Reduces vendor lock-in by using standard IaC tooling and enables GitOps workflows for infrastructure management. Supports state management, variable interpolation, and module composition for complex multi-resource deployments.

Command-line interface for provisioning, managing, and monitoring GPU instances, clusters, storage, and networking resources. Supports resource creation/deletion, SSH access management, storage operations, and real-time monitoring of resource utilization (GPU memory, compute, network). CLI abstracts API complexity and provides shell-friendly commands for scripting and automation. Integrates with standard Unix tools (pipes, grep, jq) for advanced resource queries and monitoring.

Provides persistent block storage volumes and shared network filesystems (SFS) for GPU instances and clusters. Block storage attaches to individual instances as persistent disks; shared filesystems enable multiple instances to access same data simultaneously via NFS-like protocol. Supports volume snapshots, resizing, and backup. Eliminates need for external storage services (AWS EBS, GCP Persistent Disk) and enables data persistence across instance termination.

Provides S3-compatible object storage for storing model weights, training datasets, inference results, and other artifacts. Supports standard S3 API operations (PUT, GET, DELETE, LIST) and integrates with common ML tools (PyTorch, TensorFlow, Hugging Face Transformers) via S3 protocol. Enables cost-effective storage of large files without provisioning dedicated block storage, and supports lifecycle policies for automatic archival or deletion.

In-house container registry for storing and managing Docker images used in serverless inference endpoints. Supports image push/pull via standard Docker CLI, image tagging and versioning, and automatic image scanning for vulnerabilities. Integrates with serverless inference deployment — users push image to registry, reference in endpoint configuration, and system pulls image during deployment. Eliminates need for external registries (Docker Hub, ECR) and keeps container images within EU infrastructure for GDPR compliance.

Provides real-time monitoring of GPU utilization (compute, memory, temperature), CPU usage, network throughput, and storage I/O for provisioned instances and clusters. Exposes metrics via CLI commands, web dashboard (implied), and API endpoints. Enables cost optimization by identifying underutilized resources and performance debugging by correlating metrics with training job progress. Supports alerting and historical metric retention for capacity planning.