> ## Documentation Index > Fetch the complete documentation index at: https://hyperbolic.ai/docs/llms.txt > Use this file to discover all available pages before exploring further. # Platform Comparison > Choose the right Hyperbolic service for your AI workload. ## Quick Comparison Table | Feature | On-Demand GPUs | Reserved | Private Cloud | | - | - | - | - | | **Best for** | Training, fine-tuning, experiments, burst compute | Sustained 24/7 workloads, predictable capacity | Production-scale, lowest pricing, long-term | | **Setup time** | \< 5 minutes | \< 5 minutes | Custom (days); capacity confirmed in 24h | | **Minimum commitment** | None (hourly) | From 1 week | Long-term (multi-month to multi-year) | | **Pricing model** | \$/GPU/hour | Discounted prepaid \$/GPU/hour | Custom contract | | **GPU access** | Full SSH / root (VM or bare metal) | Full SSH / root (VM or bare metal) | Full root, single-tenant; optional managed k8s / Slurm | | **Scaling** | Manual; self-serve multi-node | Self-serve; reserved up front | Pre-provisioned, custom topology | | **Networking** | Ethernet or InfiniBand (multi-node) | Ethernet or InfiniBand (multi-node) | InfiniBand, private subnets, network isolation | | **Support level** | Standard (sub-24h; P0 \< 1h) | Standard | Dedicated, direct-to-engineer (24×7) | | **SLA** | 99.5% uptime | 99.5% uptime | Custom, negotiated per contract | GPU rates vary by region and availability. Always check [app.hyperbolic.ai/gpus](https://app.hyperbolic.ai/gpus) for current rates and real-time availability. ## Detailed Comparison ### On-Demand GPUs Self-serve, pay-as-you-go GPU instances drawn from an aggregated supplier network. Spin up a single GPU or an interconnected multi-node cluster in minutes — no contracts, no upfront commitment, no sales calls. **When to use:** * Training and fine-tuning custom models * Running experiments, notebooks, and evaluations * Burst and batch processing jobs * Short training runs where you need full control of the environment **Key features:** * Full root access via SSH * Virtual Machine or Bare Metal configurations * Deploy interconnected H100, H200, B200, or B300 clusters (8, 16, 32, 64, 128+ GPUs) * Hourly billing with no hidden or egress fees — failed instances are never charged ### Reserved The same GPU capacity you would use on-demand, reserved for a fixed term at a discounted rate. Same platform, same hardware, same setup — you simply commit up front for a lower price. **When to use:** * 24/7 production inference and LLM tooling * Sustained or scheduled training runs * High-volume, predictable usage * Teams that want a discount on capacity they already run on-demand **Key features:** * Same on-demand GPUs, networking, and environment — reserved for your term * Self-serve in the app; larger commitments can also be arranged with our team * Discounted rate in exchange for committing to a term * Reserved instances run for the full term; extend before it ends, and Reservation Protection keeps them running at the on-demand rate afterwards **Pricing structure:** * Discounted prepaid \$/GPU/hour — longer terms unlock lower rates * Paid up front for the reservation period * See [Pricing](/docs/on-demand/pricing) for details ### Private Cloud Dedicated, single-tenant GPU infrastructure for production-scale AI workloads. Get an isolated environment with custom networking, storage, and SLAs — plus the best unit economics on long-term commitments. **When to use:** * Enterprise production workloads at sustained scale * Security-conscious teams that require single-tenant isolation * Large, long-term commitments (typically \$1M+ deployments) * Workloads that need custom SLAs, custom networking, or custom storage **Key features:** * Single-tenant, customer-isolated GPU clusters (H100, H200, B200, B300) * Network isolation with private subnets and firewall rules * High-bandwidth InfiniBand interconnect for distributed training * Choice of operating model: full platform visibility, [Managed Kubernetes](/docs/on-demand/managed-kubernetes) (beta), or a fully isolated environment where only your team holds credentials. Managed Slurm is on the roadmap. * Custom storage architecture (node-local NVMe and shared/parallel filesystems) * Dedicated, direct-to-engineer support with 24×7 coverage for critical issues **Pricing structure:** * Custom contract based on cluster size, GPU type, term length, and configuration * Long-term commitments with the best unit economics * [Talk to sales](mailto:sales@hyperbolic.ai) for a tailored quote ## Available GPUs Hyperbolic offers H100, H200, B200, and B300 capacity across On-Demand GPUs, Reserved, and Private Cloud. Availability may vary by product, region, supply, and configuration. If you need a specific GPU type, cluster configuration, or capacity profile, contact [sales@hyperbolic.ai](mailto:sales@hyperbolic.ai) and our team can help evaluate what is available or what can be sourced for your workload. | GPU | Architecture | Memory | Memory Bandwidth | NVLink | TDP | Best for | | - | - | - | - | - | - | - | | **B300** | Blackwell Ultra | 288 GB HBM3e | 8 TB/s | 1.8 TB/s | up to 1400 W | Long-context inference and frontier-scale training | | **B200** | Blackwell | 192 GB HBM3e | 8 TB/s | 1.8 TB/s | \~1000 W | Frontier-scale training and highest-throughput inference | | **H200** | Hopper | 141 GB HBM3e | 4.8 TB/s | 900 GB/s | 700 W | Large-model training and memory-bound inference | | **H100** | Hopper | 80 GB HBM3 | 3.35 TB/s | 900 GB/s | 700 W | Mainstream training, fine-tuning, and inference | **Node configuration (H100 / H200):** 8 GPUs per node with up to 160 vCPUs and up to 2 TB RAM per node, and up to 24 TB of local NVMe storage. Available as Virtual Machines (fast, flexible, ideal for development and small-to-medium training) or Bare Metal (full root access and system-level control, ideal for large-scale distributed training and custom CUDA configurations). **Node configuration (B200):** 8× NVIDIA B200 per node with dual Intel Xeon 6960P CPUs, \~3 TB of RAM, NVMe storage (960 GB OS drive plus 4× 3.84 TB in RAID5), and 2× 100 GbE bonded networking. **Interconnect:** Single-node instances use standard networking; multi-node clusters can use InfiniBand (up to 3.2 Tb/s, NDR / ConnectX-7) for low-latency GPU-to-GPU communication — essential for distributed training at 32+ GPUs. Choosing between H100, H200, B200, and B300? Use H100 for mainstream training, fine-tuning, and inference. Choose H200 for larger models, longer context, or memory-bound inference. Choose B200 for maximum per-GPU throughput on frontier-scale training and inference. Choose B300 when memory per GPU is the constraint — largest models and longest contexts. ## Need Help Deciding? Get personalized recommendations and custom quotes for Reserved and Private Cloud Get help from our support team Price a reserved cluster in-app in minutes Launch your first GPU instance in minutes This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.