Rent a B200
NVIDIA B200
Specifications
Renting a B200
Rent a B200 by the minute
The B200 is NVIDIA's Blackwell data-center GPU: 180 GB of HBM3e on one card, read at 7.7 TB/s on the spec sheet. It is the card for a model or a context that does not fit anywhere smaller, and for the fastest single-stream answers we have measured.
On ParalonCloud a machine is rented whole and billed by the minute. A rental opens as Jupyter Lab in your browser, with a terminal; the hourly rate you see when you start is locked for the session, billing stops the minute you stop, and there is no contract or minimum spend.
This card is rare on the network. The top of this page shows whether one is free right now, and the list price when none is. The figures below were measured by us on a B200 on 22 October 2026, with the same test every card here gets.
What fits
What fits in 180 GB
On a card this size the question is what is left for context after the model has loaded, and how many long requests it holds at once.
Measured
Measured on a B200
From the card's own verification and from serving a 27B model on it, on 22 October 2026. The spec sheet gives 7.7 TB/s of memory bandwidth and 75 TFLOPS FP32.
Steps
How to rent one
- 01
Check availability
The top of this page says whether a machine with this card is free. If none is, the rent page lists what is, including the H200 NVL and the 96 GB RTX PRO 6000.
- 02
Start the rental
Choose Jupyter Lab, accept the terms and start. You need credits for one hour to open a session and are charged only for the minutes you use.
- 03
Work, then stop
Use the terminal for pip, conda or apt, run your job, download what you want to keep, and stop.
Limits
Know before you rent
- This card is rare on the network and is often not available. The price shown when none is free is the list price, not an offer.
- It is one card, with no NVLink to another. A model whose weights and context need more than about 175 GB does not fit.
- On a 4-bit model the B200 is not ahead at high concurrency: from 32 requests up an H200 serves more tokens per second. Its edge is single-stream speed, long prompts and cache size.
- A rental is an unprivileged container: Docker inside it does not work. Install what you need with pip, conda or apt.
- Access is through Jupyter Lab and its terminal. Other ports are not exposed to the internet.
- Everything on the machine is wiped when the rental ends. Download results before you stop.
FAQ
Questions about the B200
Is a B200 faster than an H200?
For one request, yes: 201 to 231 tokens per second on a 27B model against 163 to 176, and it prefills a 250k-token prompt five seconds sooner. At 32 or more concurrent requests on the same 4-bit model, the H200 served 12 to 14% more in our test. We tested the KV cache format and the attention backend as causes and ruled both out; the full account is linked from this page.
How much of the 180 GB can I use?
vLLM saw 178.35 GiB free on the card. Serving a 27B model it used 19 GiB for weights and gave 139 GiB to the context cache at 92% utilization.
Is the card running at full speed?
The power limit was 1,000 W, the card's default, and the PCIe link at Gen 5 x16. Memory bandwidth measured 5,972 GB/s in verification, 78% of the 7.7 TB/s spec, and FP32 66.6 TFLOPS, 89% of spec.
When is a B200 worth it?
When the answer has to come fast for one user, when prompts run to hundreds of thousands of tokens, or when many long requests must stay in cache at once. For many short requests on a 4-bit model an H200 does as well for less.
Is it available to rent right now?
The badge at the top of this page says so, from the live network. Machines with this card come and go; when none is free the page shows the list price and the rent page shows what is available instead.
Read next
Need B200 servers for months rather than hours? Reserved B200 capacity from data centers.