Home / Case Studies / Sovereign AI GPU Cluster
Case Study · Sovereign AI · GPU Cluster · Gulf

64× H100 SXM5 cluster for Gulf sovereign AI — 5 petaFLOPS, 10 weeks

A government-linked technology entity in the Gulf required a sovereign AI training cluster — on-premise, under national control, no dependency on hyperscaler GPU cloud. Haink supplied 8× NVIDIA DGX H100 systems with InfiniBand fabric under applicable export licences, delivered and commissioned in ten weeks.

64H100 SXM5 GPUs total
5 PFFP8 compute capacity
5.12 TBHBM3 GPU memory total
10 wkQuote to commissioned cluster

Full cluster specification

Per DGX H100 node × 8 nodes total
ComponentSpecification
GPU8× NVIDIA H100 SXM5 80 GB HBM3 — NVLink 4.0 all-to-all within node
NVLink bandwidth900 GB/s bisection — all 8 GPUs fully connected within chassis
FP8 performance (per node)~640 TFLOPS FP8 sparse · 320 TFLOPS FP16
CPU2× Intel Xeon Platinum 8480C, 60-core / 120-thread
System RAM2 TB DDR5-4800 ECC RDIMM
NVMe storage4× 3.84 TB NVMe SSD (OS + checkpoint cache)
Inter-node fabricNVIDIA QM9700 InfiniBand HDR100 — 8× 200 Gb/s per node
Cluster bisection bandwidthNon-blocking fat-tree, 1.6 Tb/s all-to-all across 8 nodes
Shared storage100 TB NVMe all-flash, GPFS/Lustre target via 200GbE
Power per node10.2 kW typical, 11.2 kW peak (IEC 60309 60A, 3-phase)

Export compliance — what we handled

NVIDIA H100 is subject to US Export Administration Regulations (ECCN 3A090). Haink completed the full compliance process prior to shipment.

Compliance steps completed

BIS Denied Parties List

End-user entity screened against the Bureau of Industry and Security Denied Parties List prior to order acceptance. Cleared.

OFAC SDN screening

End-user and all beneficial owners screened against OFAC Specially Designated Nationals and Blocked Persons list. Cleared.

End-user statement

Signed end-user statement confirming: intended use (AI model training), no re-export to restricted destinations, no military end-use.

Export documentation

Electronic Export Information (EEI) filed via AES. Commercial invoice, packing list, and ECCN classification provided to freight forwarder for customs. Records retained 7 years per EAR 15 CFR § 762.

All hardware supplied under applicable export regulations in effect at date of shipment. Haink retains all export documentation on file per EAR record-keeping requirements.

Sovereign AI in the Gulf — context

Gulf states have made national AI sovereignty a strategic priority. The UAE's AI Strategy 2031 and Saudi Arabia's National Data and AI Authority (NDAIA) have both committed to on-premise sovereign AI compute. These programmes require hardware under national control — not cloud API access. Haink's supply capability for GPU clusters, combined with established export documentation capability for Gulf government entities, positions us as a natural supplier for this growing category of demand.

What Haink supplied

Complete cluster hardware under one PO — DGX nodes, InfiniBand fabric, shared storage, and management infrastructure.

ItemModel / SpecificationQty
GPU cluster nodesNVIDIA DGX H100 — 8× H100 SXM5 80 GB per node8
InfiniBand spine switchNVIDIA QM9700 — 64-port HDR100 (200 Gb/s per port)2
InfiniBand leaf switchNVIDIA QM9700 — fat-tree leaf tier4
Shared AI storageAll-flash NVMe array, 100 TB usable, NVMe-oF 200GbE2
Management servers1U management node, Xeon, 256 GB RAM — cluster orchestration + IPMI2
Out-of-band switch1G management fabric for BMC/IPMI plane1

Related resources

Case Study H200 Inference Cluster — 9-Day Delivery Series B AI startup, 3× throughput vs cloud Case Study Air-Gapped AI for Pharma R&D DGX Spark + RTX 6000 Ada workstations Brand Page NVIDIA supply at Haink DGX H100/H200, Quantum InfiniBand, B200 Comparison H100 vs H200 vs B200 GPU comparison Specs, pricing, lead times Q2 2026

Planning sovereign AI compute in the Gulf?

Haink supplies DGX H100, H200, and Blackwell systems with complete export licensing documentation for Gulf government and government-linked entities.

sales@haink.org