- 8x H200 SXM, about 1.1 TB HBM3e
- Over 32 PFLOPS FP8 per server
- Proven Hopper software stack
Image: NVIDIA. Manufacturer render; units ship in manufacturer packaging.
NVIDIA, NVL PCIe, new
NVIDIA H200 NVL 141GB PCIe
The NVIDIA H200 NVL (part number 900-21010-0040-000) is a dual-slot PCIe Hopper GPU with 141 GB of HBM3e memory at 4.8 TB/s and 2-way or 4-way NVLink bridging at 900 GB/s. It brings H200 memory capacity to standard air-cooled servers.
USD 43,220per unit, excl. VAT and duties
10 plus available
| Quantity | Unit price |
|---|---|
| 1 to 3 | USD 43,220 |
| 4 to 7 | USD 42,750 |
| 8 plus | USD 42,350 |
Other in-stock offers for the same part number received by our sourcing desk, as of 16 Oct 2026. Shown for context; offers vary by quantity and terms.
- Exact part number 900-21010-0040-000. No edition substitution without your written approval.
- New, factory sealed. Manufacturer warranty terms apply; coverage and claims route confirmed on your invoice.
- Typically 3 to 7 business days after cleared payment. DAP (Delivered at Place) by default, other terms on request.
- Card (Visa, Mastercard) with 3-D Secure. Bank transfer against pro forma invoice. Card details are entered on Network International's secure payment page, never on our servers.
- Cancel free of charge before dispatch. If we cannot reconfirm the allocation, you receive a full refund.
Offer terms
- Availability
- Verified stock, 10 plus available
- Unit price
- USD 43,220 per unit
- Minimum order
- 1 unit
- Ships from
- Europe, confirmed on order
- Dispatch
- Typically 3 to 7 business days after cleared payment
- Incoterm
- DAP (Delivered at Place) by default, other terms on request
- Tax and duties
- Prices in USD, excluding UAE VAT, duties and import taxes. UAE deliveries add 5% VAT.
- Payment
- Card (Visa, Mastercard) with 3-D Secure; Bank transfer against pro forma invoice
- Condition
- New, factory sealed
Overview
H200 NVL is the PCIe version of NVIDIA H200. Each card carries 141 GB of HBM3e with 4.8 TB/s of bandwidth, enough to serve large language models with long context on fewer GPUs. Cards can be linked in pairs or groups of four with NVLink bridges at 900 GB/s per GPU.
NVIDIA rates it at 3,341 TFLOPS FP8 and 1,671 TFLOPS FP16 and BF16 (with sparsity; dense figures are half). It supports up to seven MIG instances of 16.5 GB, confidential computing, and includes a five-year NVIDIA AI Enterprise subscription. It is passively cooled and configurable up to 600 W.
Typical uses
- LLM inference at long context
- Fine-tuning
- HPC and simulation
- RAG and vector search
Specifications
| Architecture | NVIDIA Hopper |
|---|
| Memory | 141 GB HBM3e |
|---|---|
| Memory bandwidth | 4.8 TB/s |
| FP8 Tensor | 3,341 TFLOPS (with sparsity) |
|---|---|
| FP16 and BF16 Tensor | 1,671 TFLOPS (with sparsity) |
| TF32 Tensor | 835 TFLOPS (with sparsity) |
| FP64 | 30 TFLOPS |
| FP64 Tensor | 60 TFLOPS |
| Bus | PCIe 5.0 x16, 128 GB/s |
|---|---|
| NVLink | 2-way or 4-way bridge, 900 GB/s per GPU |
| MIG | Up to 7 instances at 16.5 GB |
| Max power | Up to 600 W, configurable |
|---|---|
| Cooling | Passive, needs server airflow |
| Form factor | Dual-slot, full height, full length (10.5 in) |
|---|---|
| Designed for | Air-cooled enterprise and MGX servers |
| Included | NVIDIA AI Enterprise, 5-year subscription |
|---|
| Confidential computing | Supported |
|---|
Source: manufacturer documentation, checked 16 October 2026. Tensor figures follow NVIDIA's stated basis; sparse figures are marked.
Edition and part numbers
H200 NVL is the PCIe card. H200 SXM is a different module that only ships on HGX baseboards inside 4-GPU and 8-GPU servers; see HGX H200. OEM versions carry their own part numbers, for example Lenovo 900-21010-2740-030.
Delivery and warranty
Delivery. Typically 3 to 7 business days after cleared payment. DAP (Delivered at Place) by default, other terms on request. Shipments are insured to declared value and tracked. Delivery to every destination is subject to export and sanctions screening; see export compliance.
Warranty. New, factory sealed. Manufacturer warranty terms apply; coverage and claims route confirmed on your invoice. Keep serial numbers with your invoice; our warranty and RMA process explains how claims are handled.
Cancellation. Online orders can be cancelled free of charge before dispatch. If we cannot reconfirm supply after payment, we refund in full. See returns and cancellations.
Questions about the H200 NVL
What is the difference between H200 NVL and H200 SXM?
Are NVLink bridges included?
Is H200 export controlled?
How quickly can you confirm availability?
Related guides
NVIDIA H200, B200 and B300: which to buy
Memory, precision, power and form factor compared across Hopper, Blackwell and Blackwell Ultra.
How to verify enterprise GPU stock before you pay
The evidence to ask for, the warning signs, and how to structure payment when supply is tight.
A buyer checklist for bulk GPU orders
What to settle before ordering tens or hundreds of GPUs: identity, power, delivery, warranty and paperwork.
Alternatives
- 96 GB GDDR7 ECC, 1,597 GB/s
- Passive cooling, up to 600 W
- Rack servers with front-to-back airflow
Often bought with
- 400 Gb/s NDR InfiniBand or 400GbE
- PCIe 5.0 x16, 24.9 W typical
- H100 and H200 era GPU nodes
- 800 Gb/s XDR InfiniBand
- PCIe Gen6 x16
- Quantum-X800 and Spectrum-X fabrics



