Home / Brands / NVIDIA

NVIDIA GPU servers — H100 to B300, in stock and export-screened

The complete NVIDIA AI infrastructure portfolio: SXM training clusters, inference servers, DGX systems, InfiniBand networking, and RTX Ada workstations — supplied with OEM warranty to Hong Kong, Dubai, and Mainland China.

Maintained by Haink's AI infrastructure team · Updated July 2026 · export-screened, serial-verified

Request a GPU quote Full NVIDIA guide →
H100 SXM5 H200 SXM5 B200 SXM B300 SXM DGX H100 / H200 / B200 / B300 L40S InfiniBand NDR

GPU generations at a glance

Four generations — different memory, compute, and cooling requirements. Pick the one that fits the workload, budget, and facility.

H100 SXM5
H200 SXM5
B200 SXM
B300 SXM
Architecture
Hopper
Hopper
Blackwell
Blackwell Ultra
GPU Memory
80 GB HBM3
141 GB HBM3e
192 GB HBM3e
288 GB HBM3e
FP8 TFLOPS
3,958
3,958
9,000
9,000+
NVLink BW
900 GB/s
900 GB/s
1,800 GB/s
1,800 GB/s
Cooling
Air or DLC
Air or DLC
DLC required
DLC required
Best for
Training 7B–70B, proven ecosystem
70B+ inference, memory-bound training
New training clusters, maximum throughput
1T+ pre-training, frontier inference
Lead times:H100 SXM5: 4–10 weeks  ·  H200 SXM5: 6–12 weeks  ·  B200 SXM: 12–20 weeks  ·  B300 SXM: 14–24 weeks. Contact Haink for current allocation status. Check availability →

Availability & indicative pricing — as of July 2026, updated regularly

PlatformHKDubaiLead timeIndicative price
NVIDIA H200 NVL GPU server (8×)In stockIn stockcards 1–3 days · systems 6–12 wknode from ~$78K; H200 NVL card from ~$31K
NVIDIA H100 SXM5 8-GPU systemTo orderTo order4–10 wkFirm quote
NVIDIA B200 SXM 8-GPU (Blackwell)On allocationOn allocation12–20 wkFirm quote
NVIDIA B300 SXM / GB300 NVL72On allocationOn allocation14–24 wkFirm quote
NVIDIA L40S inference serverIn stockIn stock1–2 wkFirm quote
NVIDIA DGX SparkIn stockIn stock~1 wkFirm quote
RTX PRO 6000 Blackwell Server Ed.To orderTo order2–4 wkFirm quote

Live stock snapshot — 17 July 2026, Indicative prices in USD, confirmed by written quotation. Availability changes daily.

ModelSpecPriceAvailability
NVIDIA H200 141GB900-21010-0040-000, HBM3e$34,50020 pcs · ~1 week · HK
NVIDIA H100 80GBOriginal$29,000~1 week · HK
NVIDIA RTX PRO 6000Workstation / server$15,500~1 week · HK
NVIDIA RTX 5090Workstation$4,350~1 week · HK
NVIDIA DGX Spark128 GB / 4 TB$4,850 + ~$85 shipBeijing / Shenzhen

Data-centre GPUs (H200, H100) to controlled destinations are supplied to order, subject to end-user screening and export licensing.

Prices are indicative and export-screened; request a firm quote with delivered lead time for your configuration — we reply within one business day.

How to buy

ConfigureBTO / CTOBuilt to your workload — GPU count, CPU, memory, InfiniBand/Ethernet fabric and cooling (air or DLC).
ChannelAuthorized-channel sourcingSourced through authorized distribution; serial numbers verifiable before payment — no gray market.
ComplianceExport-screenedEvery GPU order is individually export-screened (end-user, end-use, destination) before shipment.

Deployed in production — see our 8× DGX H100 sovereign AI cluster and H200 inference: POC to production in 9 days.

How NVIDIA allocation actually works

"In stock" and "on allocation" are not the same thing. Hopper (H100/H200), L40S and DGX Spark ship from held stock in days. Blackwell (B200/B300, GB200/GB300) is allocation-constrained — quantities are committed against manufacturer allocation, so the honest lead time depends on the current slot, not a shelf. We tell you which it is up front, hold a slot on deposit where needed, and give you a delivered date rather than a hopeful one.

What we supply

Training H100 / H200 SXM5 GPU servers Supermicro SYS-821GE-TNHR, Dell XE9680, HPE Cray XD670 — 8-GPU SXM platforms for LLM training and fine-tuning. NVLink 4.0 fabric, ConnectX-7 InfiniBand.
Training B200 / B300 SXM GPU servers Supermicro ARS-821GL-NHR — 8-GPU Blackwell platforms. 2.3× H100 FP8 throughput. DLC-mandatory. For new AI training cluster builds and frontier inference.
Inference RTX PRO Blackwell & L40S inference NVIDIA RTX PRO 6000 Blackwell Server Edition (96 GB GDDR7) is the current inference-node GPU; L40S 48 GB Ada (Supermicro SYS-221GE 4×, SYS-111E 2×, no liquid cooling) remains the value option for 7B–70B serving.
DGX NVIDIA DGX H100 / H200 / B200 / B300 Factory-integrated 8-GPU AI appliances with 2 TB DDR5, 30 TB NVMe, 8× ConnectX-7/8 InfiniBand NICs. NVIDIA-validated, zero integration risk.
DGX DGX GB200 NVL72 Rack-scale 72-GPU Blackwell platform with 130 TB/s NVLink 5.0 fabric. Factory-integrated full-rack DLC system. For frontier model pre-training and hyperscale inference.
Workstation DGX Spark + RTX PRO Blackwell / Ada DGX Spark (GB10, 128 GB unified) for personal AI. RTX PRO 6000/5000/4500/4000 Blackwell Workstation Edition for current AI workstations; RTX 4000/5000/6000 Ada as value / from stock.
Networking InfiniBand: NDR + XDR (Quantum-X800) NVIDIA QM9700 / QM9790 (64-port NDR 400G) for H100/H200 fabrics; Quantum-X800 XDR (144× 800G, ConnectX-8/9 SuperNICs, SHARP v4) for GB300 / Rubin clusters. ConnectX-7/8 HCAs.
Networking Spectrum-X Ethernet AI fabric 400G/800G RoCEv2 Ethernet alternative to InfiniBand. Spectrum-X800 (SN5600 800G switch + BlueField-3 SuperNIC) for multi-tenant AI clouds; Spectrum-4 switches with adaptive routing where InfiniBand cost is a constraint.

Why NVIDIA leads enterprise AI infrastructure

The hardware advantages matter — but the real moat is CUDA, which has a 10+ year head start on every competing GPU platform.

CUDA ecosystem PyTorch, TensorFlow, JAX, TensorRT, NCCL — every major AI framework is built and optimized for NVIDIA CUDA first. Competitor GPU platforms run CUDA workloads in compatibility mode, not natively.
NVLink bandwidth NVLink 5.0 at 1,800 GB/s GPU-to-GPU (B200/B300) is 28× the bandwidth of PCIe Gen5. Distributed training all-reduce latency drops proportionally. Multi-node clusters require InfiniBand for between servers.
HBM memory HBM3e provides 8 TB/s memory bandwidth (B200) vs ~2 TB/s for GDDR6-based GPUs. Memory bandwidth, not raw FLOPS, is the throughput bottleneck for transformer inference — making HBM decisive for LLM performance.
Software stack TensorRT-LLM, NIM inference microservices, NCCL, DCGM, Base Command Manager — NVIDIA ships optimized production software alongside every GPU. Competing hardware ships GPUs; NVIDIA ships a platform.

Common questions

Who supplies NVIDIA H100 servers in Hong Kong?
Haink supplies NVIDIA H100 SXM5 and H100 PCIe GPU servers to enterprises in Hong Kong. Hong Kong's free port status means zero import duty on NVIDIA hardware. Standard 8× H100 SXM5 configurations on Supermicro SYS-821GE-TNHR or Dell XE9680 are available with 4–10 week lead times.
Who supplies NVIDIA GPU servers in Dubai and UAE?
Haink supplies NVIDIA H100, H200, B200, and B300 GPU servers to enterprises and data centers in Dubai and the UAE. The UAE is eligible for these GPUs under current US export rules, and every order is individually export-screened (end-user, end-use and destination) before shipment. Hardware is delivered through Dubai free trade zone logistics.
What is the difference between H100, H200, B200, and B300?
H100 SXM5 (80 GB HBM3, 3,958 TFLOPS FP8) — the proven training workhorse with the broadest software support. H200 SXM5 (141 GB HBM3e, same compute) — best for 70B+ inference where memory is the constraint. B200 SXM (192 GB HBM3e, 9,000 TFLOPS FP8, NVLink 5.0) — 2.3× faster than H100, DLC-mandatory. B300 SXM Blackwell Ultra (288 GB HBM3e) — maximum memory and compute for 1T+ frontier model training.
Does B200 require liquid cooling?
Yes — NVIDIA B200 and B300 at full training utilization require direct liquid cooling (DLC). Air cooling is insufficient for sustained B200/B300 thermal output. H100 and H200 can run in air-cooled configurations, though DLC improves thermal headroom. Haink advises on DLC facility requirements for Blackwell-generation GPU cluster orders.
Can NVIDIA GPUs be delivered to Mainland China?
US BIS export regulations restrict export of H100, H200, B200, and B300 to Mainland China. NVIDIA's China-compliant variants — H20, L20, L2 — are available within regulatory thresholds. Haink advises on compliant GPU configurations for Mainland China delivery on a per-inquiry basis, as regulations evolve. Lenovo and Huawei domestic AI server platforms are also available through Haink for China deployments.
What is the minimum hardware to fine-tune a 70B LLM?
For LoRA fine-tuning of a 70B model: 4× H100 SXM5 (320 GB total). For full supervised fine-tuning (SFT) in BF16: 8× H100 SXM5 or 4× H200 SXM5 with gradient checkpointing. A single 8× H100 SXM5 node (DGX H100 or equivalent) is the practical entry point for most enterprise fine-tuning use cases up to 70B parameters.

Need NVIDIA GPU servers?

Send the configuration — GPU generation, node count, and target region. Firm pricing and availability within one business day.

We reply within one business day. Prefer email? sales@haink.org

Go deeper

Technical guides and deployment resources from the Haink knowledge base.

H100 vs H200 vs B200 vs B300 →Detailed GPU spec comparison — architecture, memory, TFLOPS, and when to choose each.
Private AI Infrastructure →Complete guide to on-premise GPU infrastructure — costs, deployment sizes, cloud vs private analysis.
NVLink vs PCIe GPU →When NVLink SXM matters for training and when PCIe is sufficient for inference.
Liquid Cooling for AI Servers →DLC, immersion, rear-door heat exchangers — requirements by GPU generation.
AI Infrastructure Cost Guide →GPU cluster costs, TCO, and ROI vs cloud GPU — with real pricing ranges.
GPU Server Buying Guide →Decision framework: workload → GPU → server platform → network → facility checklist.