NVIDIA B200 vs H100: specs, price per hour and when the upgrade pays off

What is the difference between the B200 and the H100?

Both are NVIDIA data-centre GPUs built for AI, and both are sold mostly inside 8-GPU servers that clouds then rent out one card at a time. The H100 belongs to the Hopper generation. The B200 belongs to Blackwell, the generation that follows it.

Three differences matter in practice. The B200 has more than twice the memory, more than twice the memory bandwidth, and a 4-bit number format (FP4) that the H100 spec sheet does not list. Almost everything else in this article follows from those three, plus the price per hour.

NVIDIA B200 specs next to the H100, from NVIDIA’s datasheets

NVIDIA publishes the B200 mainly as part of its 8-GPU systems, DGX B200 and HGX B200, so the per-GPU figures below are the system figures divided by eight. The H100 column is the H100 SXM, the version that goes in the same kind of 8-GPU server.

When a provider lists an “H100”, check which version it means: the PCIe and NVL versions have different memory, bandwidth and compute figures from the SXM one.

* With sparsity, as NVIDIA lists both cards; on its B200 pages NVIDIA states that dense performance is half. B200 figures are the DGX B200 and HGX B200 system figures divided by 8. Checked on 9 November 2026. Sources: www.nvidia.com/en-us/data-center/dgx-b200/, www.nvidia.com/en-us/data-center/hgx/, www.nvidia.com/en-us/data-center/h100/, docs.nvidia.com/datacenter/tesla/mig-user-guide/supported-mig-profiles.html

On paper, that gives the B200 about 2.3x the memory, 2.4x the memory bandwidth and 2.3x the FP8 and FP16 tensor throughput of an H100 SXM. Paper is the key word: these are peak figures, and how much of them a real job uses depends on the job.

What the spec gap means for real workloads

Memory decides what fits on one card. A 70-billion-parameter model in 16-bit precision needs about 140 GB just for its weights (70 billion times 2 bytes), before the cache that grows with every conversation. On H100s that means splitting the model across two cards and paying for the traffic between them; on a B200 it fits on one.

Bandwidth sets the pace of generation. When a model writes its answer token by token, every step reads the weights from memory, so with few users at a time the speed is limited more by memory bandwidth than by raw compute. That is usually where the bandwidth gap shows first.

FP4 halves the memory of FP8. A model quantised to 4 bits takes about a quarter of its 16-bit size, which is how very large models end up on a single card. It only helps if the model has a 4-bit version and your inference software supports it.

FP64 is the exception. For double-precision scientific computing, NVIDIA lists 67 TFLOPS of FP64 Tensor Core on the H100 SXM and 37 TFLOPS on the B200. If your code is simulation in double precision, the newer card is not an upgrade.

And NVIDIA’s headline numbers? Its DGX B200 page claims 3x the training and 15x the inference performance of a DGX H100. Both are labelled “projected performance subject to change” and come from the scenarios in its footnotes: inference with 32,768-token prompts under real-time targets, and training on a cluster of 32,768 GPUs. They describe 8-GPU systems at scale, not one rented card running your model.

How much does a B200 cost per hour compared with an H100?

We checked these prices on 9 November 2026 on each provider’s own pricing page. They are on-demand prices for one GPU, in the currency each provider publishes, without conversion.

Third-party prices checked on 9 November 2026, in each provider’s currency and not converted. GPU Flow prices are live. Sources: modal.com/pricing, nebius.com/prices, docs.nebius.com/overview/regions, verda.com/pricing, docs.verda.com/welcome-to-verda/locations-and-sustainability/, www.ovhcloud.com/en-ie/public-cloud/prices/

At the three providers that list both cards, a B200 hour costs between 1.6x and 1.9x an H100 hour (same provider, same currency, so no exchange rate is involved). Set that against the roughly 2.3x tensor throughput and 2.4x bandwidth above: if your job really uses them, the B200 does more work per euro or dollar on paper. If it does not, you are paying for headroom.

Two things to watch when comparing across rows: some prices exclude VAT and ours include it, and what comes with the GPU (processors, memory, disk) differs from one provider to the next.

When does upgrading from H100 to B200 pay off?

The upgrade tends to pay off when at least one of these is true:

1. 1Your model does not fit in 80 GB. One B200 instead of two H100s means no splitting the model across cards and no traffic between them.
2. 2You serve long contexts or many users at once. The cache grows with every token in flight, and 180 GB leaves room for it next to the weights.
3. 3Your stack runs FP8 or FP4. Most of the B200 advantage is in low-precision tensor maths; a pipeline that stays in FP32 will not see it.
4. 4The card is busy most of the day. A faster GPU only pays when it has work. For a few minutes of use a day, per-second billing or a token-priced API is usually the cheaper way to pay.

It does not pay off for double-precision simulation, for small models that already fit with room to spare, or for experiments that would run just as well on part of a card. For those, a MIG slice is the middle ground: on a B200, NVIDIA’s 1g.23gb profile gets 1/7 of the compute and 23 GB of memory, and the 2g.45gb profile 2/7 and 45 GB, isolated in hardware from whatever runs on the rest of the card.

How this works on GPU Flow

We rent the B200 only, whole or in MIG slices, in a data centre in Spain. Prices are in euros with VAT included; EU businesses with a valid VAT number, and customers in the Canary Islands, Ceuta, Melilla or outside the EU, pay without VAT.

You can pay by the active hour or reserve by the month, which works out as 730 hours at the hourly rate minus 10 %. A reservation includes the GPU, its processors and its memory, and stays yours while paused.

The only thing billed separately is the network disk for your data, at €0.10 per GB a month, for as long as it exists.

Activation is not instant. You choose the configuration, we call you in the slot you pick, confirm availability and switch it on together. From then on you get SSH access with the drivers ready and run your own software.

Related

• GPU rental in Europe: hourly prices for B200, H100 and MIG slices across providers (2026)Hourly prices to rent an NVIDIA B200 or H100 in European data centres, provider by provider, plus MIG slices, reserved vs on-demand and what disk costs.
• GPU Flow vs Modal: GPUs, sandboxes and inferenceWhat a B200, a sandbox for an agent and inference cost on each, with Modal’s prices checked on its own site.