Buyer's guide · 12 November 2026
NVIDIA H200 or AMD MI350P?
Looking for H200? Read this first. A Supermicro server with ten AMD Instinct MI350P GPUs against an 8-GPU H200 server, scored round by round. If your stack runs on PyTorch, vLLM or SGLang, the challenger wins on memory per server, throughput per GPU and time to deploy.
The short answer
Which one should you buy?
Buy the MI350P if
- Your models run on PyTorch, vLLM or SGLang.
- You want more GPU memory per server: 1.44TB against 1.13TB.
- FP8, BF16 or FP4 inference is your main workload.
- You need the capacity in weeks, not in 25 to 30.
Buy the H200 if
- You depend on TensorRT-LLM or custom CUDA kernels.
- You need FP64 for simulation or HPC: 67 TFLOPS against 36.
- Memory bandwidth limits you: 4.8TB/s against 4.0.
- Your training relies on NVLink between GPUs.
Either way
- Order the network with the servers. Network cards, switches and optics have their own lead times.
- Every GPU system is export-screened before we quote.
- The MI350P lot is sold as complete servers only.
The judges' scorecard
10 x MI350P against 8 x H200, round by round
Per-GPU figures are dense, as published by AMD for the MI350P and by NVIDIA for the H200 (SXM), and as set out in our LinkedIn post of 12 November 2026. Price and lead time are as quoted to us on that date.
| Round | 10 x AMD Instinct MI350P | 8 x NVIDIA H200 | Winner |
|---|---|---|---|
| Price per GPU | About 30% less per GPU, on the quotes we are seeing | Baseline: an 8-GPU H200 SXM server | MI350P |
| GPUs per server | 10, in a 5U chassis | 8 | MI350P |
| GPU memory per server | 1.44TB (144GB HBM3e per GPU) | 1.13TB (141GB HBM3e per GPU) | MI350P |
| FP8, per GPU | 2.3 PFLOPS | 1.98 PFLOPS | MI350P |
| BF16, per GPU | 1.15 PFLOPS | 0.99 PFLOPS | MI350P |
| FP4, per GPU | 4.6 PFLOPS | Not supported | MI350P |
| Lead time | About 3 weeks after compliance | New orders quoted at 25 to 30 weeks | MI350P |
| Memory bandwidth, per GPU | 4.0TB/s | 4.8TB/s | H200 |
| FP64, per GPU | 36 TFLOPS | 67 TFLOPS | H200 |
| Software | PyTorch, vLLM and SGLang run on AMD ROCm today | CUDA still leads for TensorRT-LLM and custom kernels | Draw |
| GPU-to-GPU link | PCIe cards | NVLink | H200 |
Available now
One lot: 100 servers, 1,000 GPUs
Supermicro AS-5126GS-TNRT2, each with ten AMD Instinct MI350P 144GB GPUs, two AMD EPYC 9575F processors and 2TB of DDR5-6400, with three years of onsite next-business-day support. About three weeks after end-user compliance clears. Once the lot is gone, it is back to factory lead times.
Subject to prior sale and end-user compliance. Specification, quantity and price are confirmed on the quotation.
