When a B200 is worth asking about
A B200 is worth asking about when your model needs more than 141 GB on one card or more than one H200. Ask about stock, unit and terms before price.
The case for asking
A B200 is worth asking about when your workload needs more memory than an H200's 141 GB on one card, or when it currently spans several older cards that a smaller number of B200s could replace. It has 192 GB and is a Blackwell part. An 80 billion parameter model in 16-bit, about 160 GB of weights, is past an H200 and fits on one B200 with around 32 GB to spare for the cache. Below that size, the card's memory is capacity you may not use, and an older card is the cheaper question.
Whether it is faster for your job is not something this site can say. It holds no benchmarks for the B200 or any other card. What it holds is the memory, the generation and the listed rates, and those are enough to decide whether to start the conversation.
Where a provider quotes a speed figure for the card, ask for the setup behind it (model, precision, batch size and software version) and treat it as a reason to test rather than a result.
Workloads where 192 GB changes the plan
Serving a large model in high precision without splitting it is the cleanest case. One card means no tensor-parallel setup, no traffic between cards and one rate to pay. A 64 billion parameter model kept in 16-bit for quality reasons needs about 128 GB: an H200 holds it with about 13 GB left, a B200 with about 64 GB.
Long-context serving is the second. The key-value cache grows with every token in every active request, and a card with spare memory after the weights can keep more of it. How much more depends on the model and the traffic, so it has to be measured.
The third is larger models in 8-bit. A 120 billion parameter model at 8-bit is about 120 GB, which fits on a B200 with room, while on 80 GB cards it needs to be split across at least 2.
Reading a thin market
Cheapest hourly listing on record for the B200 192GB: $3.74. Sellers with a listed on-demand rate: 22. The current list is on the B200 page. A seller without a public rate may still sell the card on a quote; the missing rate means only that no price was published.
For comparison, the H200 reads $2.09 (sellers with a listed rate: 26), and the B300 with 288 GB reads $4.00. Dividing each rate by the card's memory gives a cost per gigabyte per hour, which is the fairest single comparison when memory is the reason you are looking.
Rates on newer cards also move as more sellers list them, and a rate older than the site's freshness window is hidden rather than shown stale. Check the date beside any figure before quoting it to a finance team.
Questions for the provider
Can you rent a single card, or only a full node? Some providers rent newer SXM parts only by the node, which turns a per-GPU rate into a much larger minimum. Is the rate on-demand, or does it need a commitment? What is the shortest term?
What software image does it ship with, and which driver and CUDA versions? A new architecture needs recent versions of your serving engine and framework, and an old image can stop you before the first test. Ask whether your engine of choice supports the card in the version you plan to run.
When can you start? This site records no stock or availability, and for newer cards it is often the deciding question. A low rate on a card that is free in three months is not an option for a launch next week.
When an older card is the better question
A model needing between 80 GB and 141 GB fits on one H200, which is the first card to price; compare its seller count with the B200's on each card's page. The B200 becomes the question only when 141 GB leaves too little room for the cache your traffic needs.
A model that fits on an 80 GB card in 8-bit, with quality holding at that precision, points to the H100 as the starting point to price, and Hopper supports 8-bit floating point too. A newer card bought for memory the job never fills is capacity paid for and left idle.
Where your stack runs on AMD, the MI300X also carries 192 GB and is worth pricing beside the B200 when memory is the deciding need. The software checks that come with that move are their own question.
A test plan worth the hours
With acceptable answers, rent the smallest unit you can for a day. Load your model in the precision you need, run your real prompts at your real concurrency and record peak memory and tokens per second at your target latency. Do the same on the card you would otherwise use.
Then put both into the rental cost calculator as cost per month at the hours you need. If the B200 holds in one card what took two before, compare it with the two-card bill, not one. The H100 or H200 post works through the same reasoning one generation earlier.
Write down what you learn either way. Should the B200 lose this quarter, the readings from the test tell you what to look for when more sellers publish a rate.
What we could not answer
- ⚑No stock or availability data is held, which is often the deciding factor for newer cards.
- ⚑No benchmarks are held, so whether a B200 finishes your job sooner than an older card must be tested.
What this rests on
infratoapp, “When a B200 is worth asking about”, read 27 October 2026. sellers with a listed B200 rate: 22. https://www.infratoapp.com/blog/when-a-b200-is-worth-asking-about
Questions people ask about this
Is the B200 worth it?
The answer turns on one fact: does your workload need 192 GB on a card? When a model or its cache does not fit on an H200's 141 GB, or currently spans several older cards, a B200 may cost less per job. A model that fits comfortably on an 80 GB card leaves the extra memory unused. Test before committing.
How much memory does a B200 have?
The B200 has 192 GB of memory and is a Blackwell part. That holds an 80 billion parameter model in 16-bit, about 160 GB of weights, with around 32 GB left for the cache and the serving engine. The B300 has 288 GB.
Why is it hard to find a B200 price?
A seller can sell a card on a quote or by the full node without publishing an hourly rate, and this site only records published rates. A missing price means only that none was published. Ask providers directly about the unit, term and start date.
These render as FAQPage structured data · answers are filled from the same live figures
Read next
Pick the smallest memory tier that holds your weights plus measured headroom: 24 GB for small models, 48 GB for mid-size, 80 GB for 70B in 8-bit.
Full fine-tuning needs memory for gradients and optimiser state on top of the weights; inference needs it for the cache. Size the GPU for the job you run.
Weights need the parameter count times the bytes per parameter: a 70B model is about 140 GB in 16-bit, 70 GB in 8-bit and 35 GB in 4-bit, before any cache.