*[DediGPU](https://dedigpu.com/) — Markdown mirror of [https://dedigpu.com/guides/h100-vs-h200-vs-b200-vs-b300](https://dedigpu.com/guides/h100-vs-h200-vs-b200-vs-b300) · updated 2026-09-02 · index for LLMs: [llms.txt](https://dedigpu.com/llms.txt) · everything: [llms-full.txt](https://dedigpu.com/llms-full.txt)*

# H100 vs H200 vs B200 vs B300: which one to rent in 2026

> Memory, bandwidth, NVLink and price per month of the four NVIDIA data-centre generations on the catalogue, and a plain answer for training, fine-tuning and inference.

*Hardware · published 20 August 2026 · updated 2 September 2026 · by DediGPU engineering*

In short:

- H100 for anything already tuned for it; H200 to serve 70B at FP8 with cache to spare; B200 for training; B300 when memory per node decides.
- Per GB the H200 is the cheapest Hopper card on the catalogue; per token trained, the B200 is the cheapest of all.
- All four cost the same in every region from 1× to 8×, minus 10 % on a 12-month term.

Four generations of the same idea, all on the catalogue at cost, all on the same HGX-style baseboards with full NVLink. The differences are memory, bandwidth and price, and they do not move together. Here is the table, then what it means for each job.

## The numbers

| Card | Memory | Bandwidth | NVLink | Power | Per month | Per GB | Per TB/s |
| --- | --- | --- | --- | --- | --- | --- | --- |
| [**A100 80 GB**](https://dedigpu.com/gpu/a100-80) Ampere | 80 GB HBM2e | 2.04 TB/s | NVLink 3 · 600 GB/s | 400 W | $654 | $8.18 | $320.75 |
| [**H100 SXM**](https://dedigpu.com/gpu/h100-sxm) Hopper | 80 GB HBM3 | 3.35 TB/s | NVLink 4 · 900 GB/s | 700 W | $1,134 | $14.18 | $338.51 |
| [**H200**](https://dedigpu.com/gpu/h200) Hopper | 141 GB HBM3e | 4.8 TB/s | NVLink 4 · 900 GB/s | 700 W | $1,252 | $8.88 | $260.83 |
| [**B200**](https://dedigpu.com/gpu/b200) Blackwell | 180 GB HBM3e | 8 TB/s | NVLink 5 · 1.8 TB/s | 1000 W | $1,408 | $7.82 | $176.00 |
| [**B300**](https://dedigpu.com/gpu/b300) Blackwell Ultra | 288 GB HBM3e | 8 TB/s | NVLink 5 · 1.8 TB/s | 1000 W | $1,762 | $6.12 | $220.25 |

Per GB and per TB/s are the monthly price divided by memory and by bandwidth: the two things an LLM workload actually buys.

## H100: the workhorse

80 GB of HBM3 at 3.35 TB/s, Transformer Engine with FP8, 900 GB/s NVLink. Every framework is tuned for it, every quantised checkpoint exists for it, and at $1,134 a month it is the cheapest Hopper card per TB/s on the list. Rent it for fine-tuning up to 70B on 2 to 4 cards, for serving models up to 70B at FP8 on two cards, and for anything where the code was written for an H100 and you want no surprises.

## H200: the same chip with 76 % more memory

Same compute as the H100, 141 GB of HBM3e at 4.8 TB/s. For inference that is the whole story: a 70B model at FP8 fits on one card with 50 GB left for cache, and long-context serving stops being a memory puzzle. The price is 10 % above the H100; the memory is 76 % more and the bandwidth 43 % more, so per GB it is the cheaper card. Rent it for serving and for fine-tuning when you would otherwise need one more H100.

## B200: the training card

Blackwell: 180 GB HBM3e at 8 TB/s, NVLink 5 at 1.8 TB/s, a second-generation Transformer Engine with FP4. Training throughput per card is roughly 2 to 2.5× an H100 at the same precision and more with FP4 inference. At $1,408 it costs 24 % more than an H100 and does more than twice the work: for training and for high-throughput inference it is cheaper per token. Rent it in 8× nodes for pre-training and large fine-tunes, and in 1× or 2× for serving 100B to 250B MoE models.

## B300: memory first

Blackwell Ultra: 288 GB per card, 2.3 TB across an eight-card node, more FP4 throughput than the B200. The reason to rent it is memory per NVLink domain: a 671B MoE model at FP8 fits on four cards, a 405B dense model at BF16 on four. At $1,762 a month it is the flagship, and per GB it costs about the same as an H100.

## And the A100

Still on the catalogue at $654 because it is still the cheapest 80 GB card per month. No FP8, 600 GB/s NVLink, 2.04 TB/s. Fine for fine-tuning with LoRA, for INT8 serving and for anything already tuned for Ampere.

## The plain answer

- **Pre-training or large full fine-tunes:** 8× B200 nodes on InfiniBand. B300 when memory per node is the constraint.
- **Serving up to 70B:** 1× H200 at FP8 (room for cache) or 2× H100 at FP8.
- **Serving 100B to 250B MoE:** 2× H200 or 2× B200 at FP8.
- **Serving 405B and 671B:** 4× B300 or 8× H200 / B200 at FP8.
- **LoRA fine-tunes on a budget:** 1× to 2× A100 80 GB, or 1× H100 if you want FP8 later.

All four are the same price in every region, from 1× to 8×, with the term discount on top: 10 % off on a 12-month term.