Get quotes

Home / Guides / H100 vs H200 vs B200

Buyer guide

H100 vs H200 vs B200 (and A100, L40S): a buyer's guide

Which data-center GPU to rent in 2026: memory, bandwidth, form factor and sourced on-demand prices for H100, H200, B200, A100 and L40S.

Published November 20, 2026 · Prices as of Nov 20, 2026 · Sources at the end

Most GPU buying decisions in 2026 come down to five data-center parts: NVIDIA's A100, H100, H200 and B200, plus the L40S for lighter inference and graphics. This guide compares them on the specs that usually decide the choice, memory and memory bandwidth, and on the published on-demand prices on our price board.

Specs and prices side by side

GPUMemoryBandwidthForm / interconnectMedian $/GPU-hr$/GB-hr
H10080 GB HBM33.35 TB/sSXM, NVLink$3.94$0.049
H200141 GB HBM3e4.8 TB/sSXM, NVLink$5.35$0.038
B200180 GB HBM3e8 TB/sSXM, NVLink 5$8.50$0.047
A10080 GB HBM2e2.0 TB/sSXM, NVLink$2.00$0.025
L40S48 GB GDDR6864 GB/sPCIe, no NVLink$1.09$0.023

Specs from NVIDIA product pages (SXM versions unless noted); B200 memory as listed by providers. Median = median of published on-demand prices on our board as of Nov 20, 2026 (includes "from" prices and range lows). $/GB-hr is our arithmetic: median price divided by GPU memory.

Start with memory, not FLOPs

For large-language-model work the first question is whether the model, its KV cache and your batch fit in GPU memory. A model that doesn't fit has to be split across more GPUs, which adds cost and communication overhead.

  • H100 (80 GB) is the default for training and inference when the workload fits comfortably in 80 GB per GPU. It has the deepest supply: 14 published on-demand prices on our board, from $1.99 to $6.16.
  • H200 (141 GB) uses the same Hopper architecture as H100, so software that runs on H100 runs on H200, with about 76% more memory and roughly 43% more bandwidth. It tends to pay off for long-context inference and large models that would otherwise need extra H100s. Median list on our board: $5.35.
  • B200 (180 GB) is NVIDIA's Blackwell generation, with the most memory and bandwidth here and NVLink 5. It is the choice for new frontier-scale training and high-throughput inference, at the highest prices: median $8.50. Check that your frameworks, kernels and drivers support Blackwell before committing.
  • A100 (80 GB) is the previous generation. Cheaper per hour, and still fine for fine-tuning, smaller training runs and batch inference where time-to-result matters less.
  • L40S (48 GB GDDR6) is a PCIe card without NVLink. It is cost-effective for single-GPU inference of small and mid-size models, image and video generation, and graphics, but a poor fit for multi-GPU training.

Price per hour vs price per result

A higher hourly rate can still be cheaper per result. If a B200 finishes a job in well under half the time of an H100, it can cost less overall even at roughly twice the median hourly list. The only reliable way to know is to benchmark your own workload on a short rental before committing to a long term.

Memory cost points the same way. On median list prices, H200 memory costs less per GB-hour than H100, so memory-bound inference often favors H200 even though its hourly rate is higher.

Same GPU name, different product

Two offers for "H100" can be different products. Check:

  • SXM vs PCIe/NVL. SXM parts connect through NVLink in 8-GPU baseboards; PCIe and NVL variants are cheaper and slower to scale (our board lists RunPod H100 SXM at $4.39 and H100 PCIe at $2.89, for example).
  • Node interconnect. For multi-node training, InfiniBand or RoCE bandwidth per node matters as much as the GPU. Providers list this very differently.
  • CPU, RAM and local storage per GPU, which vary widely between providers at the same GPU price.

Quick picks

  • Fine-tuning or training that fits in 80 GB, cost-sensitive: H100, or A100 if time matters less.
  • Long-context or large-model inference: H200.
  • New large-scale training or maximum throughput: B200, after a compatibility test.
  • Small-model inference, image/video generation: L40S.

Not sure? Send a brief with your model size and workload and ask providers to quote more than one option.

Sources

  1. GPU TRADER price board (Nov 20, 2026)
  2. NVIDIA H100 product page
  3. NVIDIA H200 product page
  4. NVIDIA B200 product page
  5. NVIDIA A100 product page
  6. NVIDIA L40S product page
  7. Lambda pricing (B200 180 GB listing)
  8. RunPod pricing

General information, not professional advice. Prices are published list prices as of Nov 20, 2026 and change often; confirm with the provider.

Tell us what you need.

Send a brief. We collect quotes from independent providers and introduce you to the ones that fit. Free to request.