H200 vs H100
Same compute, 76% more memory. Per GB of HBM, used nodes cost the same.
The difference
The H200 is an H100 with more and faster memory: 141 GB at 4.8 TB/s against 80 GB at 3.35 TB/s, with the same dense BF16 and FP8 throughput and the same 700 W. Used, it costs 1.8 times as much: $43,600 per GPU in an HGX H200 node against $24,400 in an HGX H100 node in October 2026. Per gigabyte of HBM, the two cost the same.
Side by side
| Per GPU | H100 | H200 | H200 / H100 |
|---|---|---|---|
| Architecture, released | Hopper, Late 2022 | Hopper, Mid 2024 | |
| Memory | 80 GB HBM3 | 141 GB HBM3e | 1.76x |
| Memory bandwidth | 3.35 TB/s | 4.8 TB/s | 1.43x |
| BF16, dense TFLOPS | 989 | 989 | 1.00x |
| FP8, dense TFLOPS | 1,979 | 1,979 | 1.00x |
| FP4, dense TFLOPS | Not supported | Not supported | |
| FP64 (Tensor / matrix), TFLOPS | 67 | 67 | 1.00x |
| Max power per GPU | 700 W | 700 W | 1.00x |
| Power per 8 GPUs | About 10 kW | About 10 kW | 1.00x |
| GPU-to-GPU link | NVLink 4, 900 GB/s per GPU | NVLink 4, 900 GB/s per GPU | |
| Scale-up domain | 8 GPUs (HGX board) | 8 GPUs (HGX board) | |
| Form factor, cooling | SXM5 module, 8-GPU HGX board; air | SXM5 module, 8-GPU HGX board; air | |
| Price per GPU, October 2026 | $14,250 (Used, typical realized sale) | $43,600 (Used 8-GPU node, per GPU) | |
| Launch price reported in the press | $25,000 to $40,000 per chip, analyst estimates (2024) | Up to $40,000 per GPU, HSBC (2024) | |
| Residual band, 2027 (conservative) | 11-19% of original price | 34-46% of original price |
Tensor Core figures are dense. Prices as of October 2026.
Sources: American Compute, NVIDIA, Managed GPUs (American Compute), CNBC, Tom's Hardware.
Memory is the whole difference
Same chip, more and faster memory. Both are Hopper GPUs with 989 dense BF16 and 1,979 dense FP8 TFLOPS; the H200 moves from HBM3 to HBM3e, with 76% more capacity and 43% more bandwidth. Inference is usually limited by memory bandwidth, because every generated token reads the model’s weights, and that is where the H200 earns its price.
The market prices Hopper by memory. A used HGX H200 node at $348,500 holds 1,128 GB of HBM, about $309 per GB; a used HGX H100 node at $195,000 holds 640 GB, about $305 per GB.
The H200 is near the end of production
Used H200s sell above new ones. New H200 lead times ran past 40 weeks in September 2026, and used nodes sold above the roughly $340,000 volume price of a new one. Cisco set September 23, 2026 as the last ship date for its HGX H200 server.
Owning one ages fast. An H200 bought new in 2026 is already two years into its technology life, because the generation first shipped in 2024; its conservative band falls from 45-59% of original price in 2026 to 7-15% in 2029.
Sources: NVIDIA, American Compute, Cisco.
Residual value and appraisals
| Year | H100 | H200 |
|---|---|---|
| 2026 | 27-39% | 45-59% |
| 2027 | 11-19% | 34-46% |
| 2028 | 5-7% | 24-35% |
| 2029 | 2-3% | 7-15% |
| 2030 | 1% | 4-6% |
What a used unit is expected to fetch, as a share of its original price: the conservative band from our residual value report.
H100 appraisal · H100 residual value · H200 appraisal · H200 residual value
Discuss a transaction
Send us the equipment list and the loan or lease you have in mind. We will come back with an indication and the questions we would need answered.
Prefer email? hello@amcompute.com