B200 vs H100
A full generation apart. Per FP8 FLOP, a new B200 costs about the same as a used H100.
The difference
The B200 is a generation ahead: 2.3 times the H100’s dense FP8 throughput, FP4 support, 2.25 times the memory and 2.4 times the bandwidth, at 1,000 W instead of 700 W. A new HGX B200 costs $45,000 to $58,750 per GPU; a used HGX H100 node works out to $24,400 per GPU. Per unit of FP8 compute, new Blackwell and used Hopper cost about the same.
Side by side
| Per GPU | H100 | B200 | B200 / H100 |
|---|---|---|---|
| Architecture, released | Hopper, Late 2022 | Blackwell, Early 2025 | |
| Memory | 80 GB HBM3 | 180 GB (1.4 TB per 8-GPU board) | 2.25x |
| Memory bandwidth | 3.35 TB/s | 8 TB/s | 2.39x |
| BF16, dense TFLOPS | 989 | 2,250 | 2.28x |
| FP8, dense TFLOPS | 1,979 | 4,500 | 2.27x |
| FP4, dense TFLOPS | Not supported | 9,000 | |
| FP64 (Tensor / matrix), TFLOPS | 67 | 37 | 0.55x |
| Max power per GPU | 700 W | 1,000 W air-cooled, 1,200 W liquid | 1.43x |
| Power per 8 GPUs | About 10 kW | About 14 kW | 1.40x |
| GPU-to-GPU link | NVLink 4, 900 GB/s per GPU | NVLink 5, 1.8 TB/s per GPU | |
| Scale-up domain | 8 GPUs (HGX board) | 8 GPUs (HGX board) | |
| Form factor, cooling | SXM5 module, 8-GPU HGX board; air | SXM module, 8-GPU HGX board; air or liquid | |
| Price per GPU, October 2026 | $14,250 (Used, typical realized sale) | $45,000-$58,750 (New 8-GPU node, per GPU) | |
| Launch price reported in the press | $25,000 to $40,000 per chip, analyst estimates (2024) | $30,000 to $40,000 per unit, per NVIDIA’s CEO (2024) | |
| Residual band, 2027 (conservative) | 11-19% of original price | 40-54% of original price |
Tensor Core figures are dense. Prices as of October 2026.
Sources: American Compute, NVIDIA, Managed GPUs (American Compute), CNBC.
Compute per dollar
Compute per dollar is about even. A new HGX B200 costs 1.84 to 2.41 times a used HGX H100 per GPU, for 2.27 times the dense FP8. FP4, which Hopper lacks, tips inference toward Blackwell.
Power decides where each one can run
An H100 server fits more data centers. An 8-GPU H100 server draws about 10 kW; a B200 server about 14 kW, air-cooled in 10U or liquid-cooled in 4U. Colocation at $240-325 per kW a month adds $0.41-0.56 per GPU-hour for an H100 and $0.58-0.78 for a B200.
Blackwell holds a higher band
Blackwell holds a higher band. Our conservative band puts the B200 at 40-54% of original price in 2027 and the H100 at 11-19%. Hopper prices are held up by scarce liquid-cooled space; once that space is built, Blackwell will not have the same support.
Sources: NVIDIA, American Compute, Managed GPUs (American Compute).
Residual value and appraisals
| Year | H100 | B200 |
|---|---|---|
| 2026 | 27-39% | 59-67% |
| 2027 | 11-19% | 40-54% |
| 2028 | 5-7% | 30-42% |
| 2029 | 2-3% | 16-26% |
| 2030 | 1% | 6-11% |
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 · B200 appraisal · B200 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