Castellan · Services · HPC Hardware Procurement
Service 04 — HPC Hardware Procurement

GPUs at scale. Procured, delivered, de-risked.

Castellan procures GPU and HPC hardware the way it builds campuses: as an operating discipline. Blackwell-class GPU systems, networking fabric, and storage — sourced through established OEM and distribution relationships, benchmarked on price, and delivered to your dock or into a Castellan facility, commissioned and ready. From single-GPU trades to complete, integrated racks.

Tier-1OEM
Established OEM & distribution relationships
Blackwell-class
GB300 · B200 · H200 system generations
End-to-end
Sourcing → logistics → customs → commissioning
Vendor-neutral
Specified to the workload, not the sales quota
01
What we procure

The full bill of materials for an AI deployment.

Procurement is where AI projects quietly lose months and margin: allocation queues, opaque pricing, freight and customs friction, and hardware that arrives before the site is ready — or long after. Castellan runs this layer as a professional supply-chain operation, from specification to energised racks.

  • 01
    GPU & accelerator systems
    NVIDIA HGX/DGX-class platforms — Blackwell and Hopper generations — specified, quoted, and secured against real delivery slots.
  • 02
    Networking fabric
    InfiniBand and RoCE-class Ethernet fabrics, optics, and cabling — sized for the cluster topology, not bought piecemeal.
  • 03
    Storage & data infrastructure
    High-throughput parallel storage and NVMe tiers matched to training and inference I/O profiles.
  • 04
    Power & cooling equipment
    PDUs, busway, CDUs, and direct-liquid-cooling components — the long-lead items that gate every deployment schedule.
02
Platforms & trading

H100 to GB300: the platforms we source and trade.

Castellan trades GPUs outright — single cards and servers, volume lots, and vetted secondary-market units — and delivers complete, integrated racks: HGX baseboards, NVL72 systems, networking and storage, cabled and burned in. Every current NVIDIA data-centre generation, specified to your workload:

Hopper
NVIDIA H100
80 GB HBM3, 3.35 TB/s memory bandwidth, 900 GB/s NVLink. The deployed workhorse of the AI build-out — deep trading liquidity and strong price-performance for training and inference.
Hopper
NVIDIA H200
141 GB HBM3e, 4.8 TB/s memory bandwidth. Drop-in HGX upgrade on the Hopper platform — the memory play for long-context and memory-bound inference workloads.
Blackwell
NVIDIA B200
192 GB HBM3e, 8 TB/s memory bandwidth, fifth-generation NVLink at 1.8 TB/s. The current training flagship, delivered in HGX B200 8-GPU systems.
Blackwell
NVIDIA GB200 NVL72
72 Blackwell GPUs and 36 Grace CPUs in one liquid-cooled, rack-scale system with unified NVLink domain — delivered as complete racks, ready to energise.
Blackwell Ultra
NVIDIA GB300 NVL72
288 GB HBM3e per GPU on the rack-scale NVL72 platform — built for reasoning-era inference at maximum throughput per rack.
Trading
GPU trading desk
Buy- and sell-side: we place surplus units for owners and source allocation for buyers — inspected, warranted where available, with clean documentation and logistics.
03
How we de-risk

Why buyers procure through Castellan.

01
Allocation access
A decade of buying energy-intensive compute at scale means standing relationships where it counts — and real delivery slots instead of waitlists.
02
Price discipline
Multi-vendor benchmarking on every bill of materials. You see the comparison, not just the quote.
03
Vendor-neutral specification
Systems specified from your workload backwards — training, fine-tuning, or inference — not from a vendor's roadmap forwards.
04
Logistics & customs
EXW to DDP: freight, insurance, customs, and staging across US, EU, and Swiss jurisdictions — handled as one accountable scope.
05
Deployment & commissioning
Rack integration, burn-in, and acceptance testing — delivered into your facility or straight into a Castellan powered shell.
06
Lifecycle & warranty
Warranty management, spares strategy, and refresh planning so the fleet stays productive across its depreciation life.
04
Engagement

One accountable counterparty, from PO to power-on.

We work for enterprises building their first cluster, AI companies scaling an existing fleet, and institutions that need auditable, compliant procurement under Swiss or European contract law. Engagements run standalone or combined with a Castellan lease — hardware and facility delivered as one programme.

Scope
Specification · sourcing · logistics · commissioning · lifecycle
Hardware
GPU systems · networking · storage · power & cooling equipment
Commercial models
Agency (fee-based) or principal (purchase & resale) structures
Jurisdictions
Switzerland · EU · United States
Contract law
Swiss or EU governing law, institutional documentation
Delivery
To client site, or integrated into a Castellan facility

Need GPUs on a real timeline?

Tell us the workload, scale, and target date. We'll come back with a specified bill of materials, benchmarked pricing, and a delivery schedule you can plan against.