NVIDIA H200 141GB
The Hopper variant that bought memory instead of cores: same GH100 silicon as the H100, but 141GB of HBM3e at 4.8TB/s - 1.8x the capacity and 1.4x the bandwidth. For large-model inference and long-context work that is the difference between fitting the model and splitting it across cards. Pricing on request against confirmed stock.
Fully Tested & Documented · Hong Kong Warehouse Stock · Worldwide Shipping 3–7 Days · 1-Year Warranty
Hardware Gallery
The H200, Up Close
Hopper Module Family (Representative)
Representative photo of the Hopper-generation SXM module family - the H200 uses the same GH100 package as the H100 with different HBM stacks. The label on the unit you receive is read and photographed against your order before dispatch; this image is illustrative, not the exact part.
HBM Stacks, Up Close (Representative)
Close representative view of a Hopper package: die in the centre, memory stacks flanking it. H200 carries 141GB of HBM3e at 4.8TB/s - the figure that separates it from the H100's 80GB at 3.35TB/s.
141
GB HBM3e per GPU
4.8
TB/s memory bandwidth
3,958
TFLOPS FP8 tensor (SXM)
7
MIG instances, 18GB each (SXM)
Key Specifications
| Component | Specification |
|---|---|
| Model | NVIDIA H200 - SXM (HGX H200 baseboard) and H200 NVL (PCIe dual-slot) |
| Architecture | Hopper GH100 with HBM3e memory |
| Memory | 141GB HBM3e, 4.8TB/s (both variants) |
| FP8 / INT8 Tensor Core | 3,958 TFLOPS (SXM) / 3,341 TFLOPS (NVL), with sparsity |
| BF16 / FP16 Tensor Core | 1,979 TFLOPS (SXM) / 1,671 TFLOPS (NVL), with sparsity |
| TF32 / FP32 / FP64 | 989 / 67 / 34 TFLOPS (SXM); 835 / 60 / 30 TFLOPS (NVL) |
| TDP | Up to 700W configurable (SXM); up to 600W configurable (NVL) |
| MIG | Up to 7 instances, 18GB each (SXM); 16.5GB each (NVL) |
| Interconnect | NVLink 900GB/s (SXM); PCIe Gen5 128GB/s plus 2- or 4-way NVLink bridge (NVL) |
| Form factor | SXM module for HGX baseboards; PCIe dual-slot air-cooled for NVL |
H200 keeps the H100's compute and changes the memory: 141GB of HBM3e at 4.8TB/s against 80GB at 3.35TB/s. The practical consequence is fewer cards for a given model size - a 70B model in FP8 fits with room for KV cache on one H200 where an H100 forces tensor parallelism. Figures above are from NVIDIA's published H200 specification; the deployment trade-offs are worked through in the AI GPU server buying guide. Full range: data centre GPUs; the H100 sibling: H100 80GB.
Transparent Pricing
2026 Reference Pricing by Configuration
H200 pricing in the secondary market reprices weekly, so we do not print a figure that would be stale before you could act on it. Send the form factor (SXM or NVL), quantity, destination and end use, and we quote against confirmed stock with the date the quote holds until.
Single GPU
On Request
SXM or NVL, one card. Serial, condition and test record disclosed before dispatch.
Multi-GPU / node
On Request
4-8 card nodes or HGX baseboard configurations, with the export documentation prepared per destination.
No price figure is published on this page by design. Quotations state the stock position, the condition line by line, and the date the figure holds until. Range and deployment guidance: AI GPU server buying guide.
Origin
Hong Kong Warehouse
Free port — full export documentation
Transit
3–7 Days Air Freight
Tracked, insured, worldwide
Warranty
1-Year Warranty
Diagnostic report in every box
Quality Process
What We Check Before an H200 Ships
HBM3e Stack and ECC
Full 141GB exercised with ECC telemetry logged; any corrected or uncorrected error history is disclosed with the unit rather than cleared.
Sustained Tensor Load
FP8 and BF16 workload at the configured TDP while clocks, power draw and junction temperatures are logged - a module that throttles below its configuration is rejected.
NVLink / Baseboard Check
On SXM modules, NVLink lane validation on a reference HGX baseboard; on NVL cards, PCIe Gen5 link width and bridge links verified at full rate.
Firmware and VBIOS
Firmware and VBIOS compared against NVIDIA's current data-centre branch; mining-modified or reflashed boards are screened out and never resold as data-centre parts.
Serial and Export Check
Serial and part number verified against board markings. High-performance GPUs are export-controlled: destination, end user and end use are confirmed before a quotation is issued.
Report in the Box
Every unit ships with its per-machine diagnostic report and warranty. Full process: inspection guide.
Frequently Asked Questions
What is the difference between H200 and H100?
Memory, not compute. Both use the GH100 die, so the FP8, BF16 and TF32 throughput figures are identical for a given power budget. The H200 carries 141GB of HBM3e at 4.8TB/s; the H100 carries 80GB of HBM3 at 3.35TB/s. That 1.8x capacity and 1.4x bandwidth is what lets larger models and longer contexts run on fewer cards.
Should I buy H200 SXM or H200 NVL?
SXM for training and tightly-coupled multi-GPU work: it needs an HGX H200 baseboard, runs at up to 700W configurable and gets 900GB/s NVLink. NVL (PCIe dual-slot, up to 600W) drops into standard GPU servers and gives 2- or 4-way NVLink bridging at 900GB/s per GPU. If you are building a new HGX node, SXM; if you are expanding existing PCIe servers, NVL.
How much does an NVIDIA H200 cost?
We quote on request against confirmed stock, with the date the figure holds until stated in writing - secondary-market pricing for H200 moves weekly and a printed number would be stale before you could act on it. Send the quantity, form factor and destination and we confirm availability and export documentation requirements first.
Can HKCHL ship H200 internationally?
High-performance data-centre GPUs are export-controlled items, and the rules depend on destination, end user and end use. We confirm those three points before quoting and handle the documentation. We do not ship where the applicable controls prohibit the transaction.
What condition are the units in?
New-sealed and tested pre-owned units, each with a per-serial test record covering HBM ECC, sustained tensor load and thermals, plus a 1-year written warranty. Engineering samples and reflashed boards are excluded.
Request an H200 Quote
Tell us the form factor (SXM or NVL), quantity, destination and end use. We confirm stock and export requirements, then quote against confirmed supply with the date the figure holds until.
Request a Quote