Compute topology
GPU profile, node configuration, interconnect and system architecture aligned with the workload.
Capacity
Dedicated compute, matched to the workload.
We structure single-tenant GPU reservations for teams that need reliable access, predictable performance and a defined operating model across global deployment locations.
The service
A committed infrastructure envelope.
The result is not a metered cloud account. It is capacity reserved for an agreed term and built around the actual production requirement.
Storage and network
Data plane, storage performance, ingress, egress and connectivity specified as part of the deployment.
Reservation term
Capacity held under a defined commercial horizon, with ramp and headroom considered upfront.
Sourced capacity · coming soon
512 H100s and 1,024 H200s entering allocation.
We have sourced the capacity below. Full-node and multi-node allocation registration is open now; deployment dates, topology and final commercial terms are confirmed before contract.
512× NVIDIA H100 SXM
80 GB HBM3 per GPU · HGX form factor
Node unit8 GPUs
Available capacity512 GPUs
Aggregate HBM640 GB per node
Reservation3–36 months
LocationConfirmed per allocation
TenancyDedicated
1,024× NVIDIA H200 SXM
141 GB HBM3e per GPU · HGX form factor
Node unit8 GPUs
Available capacity1,024 GPUs
Aggregate HBM1.128 TB per node
Reservation3–36 months
LocationConfirmed per allocation
TenancyDedicated
Market snapshot
Published on-demand pricing
USD per GPU-hour, observed 20 August 2026. Public cloud benchmarks—not a Belcastra offer.
Sources: Runpod, Lambda, DataCrunch. Taxes, storage, network and commitment discounts vary.
H100 80 GB
$2.99–$4.29 / GPU-hour
8-GPU node equivalent: $23.92–$34.32/hour
H200 141 GB
$4.13–$4.59 / GPU-hour
8-GPU node equivalent: $33.04–$36.72/hour
Technical delivery model
Remote control, telemetry and billing—defined upfront.
Each allocation is delivered as isolated dedicated capacity with an agreed access model, runtime stack, monitoring boundary and billable unit.
Secure remote administration
Tenant access is provided through an IP-allowlisted SSH/HTTPS endpoint, WireGuard or IPsec VPN, or private connectivity where available. MFA, role-based accounts and customer-owned SSH keys control the management plane. Out-of-band BMC access remains restricted to infrastructure operations.
Bare metal, Slurm or Kubernetes
Capacity can be delivered as an accepted Linux image on dedicated nodes, as a Slurm partition, or as isolated Kubernetes GPU worker pools. NVIDIA drivers, CUDA and container runtime versions are pinned during acceptance and changed under an agreed maintenance process.
GPU and node-level observability
NVIDIA DCGM/NVML telemetry covers utilization, HBM use, temperature, power, clocks, ECC events and Xid errors. Node, network and storage health can be exposed through Prometheus/Grafana dashboards, alerts or a customer monitoring integration.
Controlled lifecycle
Provisioning includes image validation, network segmentation, account handover and burn-in testing. Incidents, node swaps, driver changes and maintenance are handled through named support contacts and the SLA agreed for the allocation.
How consumption is calculated
Reserved GPU-hours, not job runtime.
A dedicated reservation is billed for capacity held available to the customer, whether GPU utilization is 0% or 100%. Contracted maintenance treatment and service credits are defined in the SLA.
GPU-hours = reserved GPUs × billable hours
Monthly compute = GPU-hours × contracted GPU-hour rate
Example using a 730-hour month: 8 GPUs = 5,840 GPU-hours; 512 H100s = 373,760 GPU-hours; 1,024 H200s = 747,520 GPU-hours.
Persistent storage is measured in GB-month, outbound traffic in GB or TB, and optional managed services as a fixed recurring line item. These are quoted separately from compute.
Designed for
Workloads where shared availability is a business constraint.
AI teams and integrators
Training, fine-tuning, inference and customer deployments that need a stable production footprint rather than opportunistic capacity.
Quantitative and algorithmic firms
Research and production workloads where repeatability, isolation and operational responsiveness matter.
High-assurance digital operators
Teams with defined data location, access, security and contractual requirements.
Quoted against real supply.
The allocation bands above are the configurations we are actively sourcing and quoting. A binding offer follows confirmation of the location, topology, start date, contract duration and selected supply partner.
A useful first brief
Workload. Location. Start date. Term.
Those four inputs are enough to begin a qualified sourcing process.