Searching for an NVIDIA B200 price usually means one of two questions: what does the hardware cost, and what will it cost to run my model on it? NVIDIA does not publish a list price for the B200, and most teams never buy one. They rent B200 time, and the rental price depends as much on how you are billed as on the hourly rate.
This guide covers the B200's specifications, how it compares with the GPUs around it, the three ways B200 capacity is priced, and fal's own published B200 price with worked examples.
What the NVIDIA B200 is
The B200 is the first data-center GPU in NVIDIA's Blackwell generation. It follows the Hopper-based H100 and H200, and it sits one step below the Blackwell Ultra B300.1
Three changes set it apart from Hopper:
- Memory. 192 GB of HBM3e per GPU, up from 141 GB on the H200 and 80 GB on the H100.2
- Bandwidth. 8 TB/s of memory bandwidth, about 1.7 times the H200. Most inference is limited by memory bandwidth, so this is the number that moves tokens per second and frames per second.2
- FP4. Blackwell adds a native 4-bit floating-point format (NVFP4). Hopper has no FP4 path, so FP4 throughput is new rather than faster.3
The B200 also moves to fifth-generation NVLink at 1.8 TB/s per GPU, twice the Hopper NVLink bandwidth, which matters when one model is split across several GPUs.2
B200 specifications compared with the H100, H200 and B300
| Spec | B200 | H200 | H100 | B300 |
|---|---|---|---|---|
| Architecture | Blackwell | Hopper | Hopper | Blackwell Ultra |
| GPU memory | 192 GB HBM3e | 141 GB HBM3e | 80 GB HBM3 | 288 GB HBM3e |
| Memory bandwidth | 8 TB/s | 4.8 TB/s | 3.35 TB/s | 8 TB/s |
| FP4 dense (TFLOPS) | ~9,000 | N/A | N/A | ~13,500 |
| FP8 dense (TFLOPS) | ~4,500 | ~1,979 | ~1,979 | ~4,500 |
| FP16/BF16 dense (TFLOPS) | ~2,250 | ~989 | ~989 | ~2,250 |
| FP64 Tensor Core (TFLOPS) | ~37 | ~67 | ~67 | ~1.25 |
| NVLink per GPU | NVLink 5 (1.8 TB/s) | NVLink 4 (900 GB/s) | NVLink 4 (900 GB/s) | NVLink 5 (1.8 TB/s) |
| 8-GPU system power (max) | ~14.3 kW | ~10.2 kW | ~10.2 kW | ~14 kW |
Blackwell figures are per GPU in NVIDIA's 8-GPU HGX and DGX servers: the board totals divided by eight. Dense throughput is half of the sparse figure, except B300 FP4, where NVIDIA publishes the dense total (108 PFLOPS per 8 GPUs) directly.24 These 8-GPU servers are what cloud platforms rent out.
Two details are easy to miss:
- Usable memory is a little under 192 GB. NVIDIA lists 1,440 GB of total GPU memory for a DGX B200, which is 180 GB per GPU. Plan model and KV-cache sizes against roughly 180 GB.2
- The B200 keeps strong FP64. The B300 traded most of its FP64 throughput for FP4. If a workload mixes AI with scientific computing or simulation, the B200 is the better Blackwell choice.4
When a B200 is the right GPU
A B200 costs more per hour than Hopper GPUs, so it pays off only when the workload uses what it adds.
Large models on one GPU. With about 180 GB usable, many 70B-class models fit on one B200 at FP8 with ample room for KV cache. On an 80 GB GPU the FP8 weights alone take most of the memory, so long contexts or large batches can push the same model past 80 GB and onto several GPUs with tensor parallelism. Fewer GPUs per replica usually means a lower cost per request.
Video generation. Video models are attention-heavy and memory-hungry, and they are often bandwidth-bound. The step from 4.8 TB/s to 8 TB/s shortens each denoising step, and the extra memory fits longer clips and higher resolutions.
FP4-quantized inference. If your model runs well in NVFP4, a B200 roughly doubles the peak math throughput of FP8 on the same GPU. Validate output quality first, because not every model tolerates 4-bit weights.
Where a B200 is overkill. Standard image generation and small models rarely use 192 GB or 8 TB/s. fal's machine-type guidance points image generation to L40 or RTX PRO 6000 GPUs and keeps the B200 for the largest models.5 A smaller GPU at a lower hourly rate is usually cheaper per image.
How B200 pricing works
Owning B200s means buying 8-GPU servers, then paying for power, liquid-ready cooling, high-speed networking and the people who run them. An 8-GPU DGX B200 draws up to about 14.3 kW.2 For most teams, renting is the practical route, and B200 rental time is sold in three ways.
Hourly (on-demand) rates
The headline number is an hourly price per GPU. Two questions decide what you actually pay at that rate:
- What is the smallest unit you can rent? Some offers only rent whole 8-GPU servers, so the real hourly price is eight times the per-GPU rate.
- What is the billing granularity? Hourly billing rounds up. A 10-minute test on an hourly plan costs a full hour.
Per-second (serverless) billing
Serverless platforms bill per second for the time your workload holds a GPU. Per-second billing makes the hourly rate a ceiling rather than a floor. If traffic is bursty and runners scale down between bursts, you pay only for the seconds a runner is alive.
On fal Serverless, a runner is billed while it is in the SETUP, IDLE, RUNNING, DRAINING and TERMINATING states. Time spent waiting for hardware (PENDING) and pulling your container image (DOCKER_PULL) is not billed, and requests that fail with a 5xx error are not charged.6 A deployment that scales to zero runners has no compute charge. Multi-GPU runners are billed as GPU count multiplied by duration.6
Reserved and committed capacity
The lowest B200 rates come from a commitment: reserved runners, a spend commitment or a longer contract. You trade flexibility for a lower per-hour price and guaranteed capacity. Reserved pricing makes the most sense when your baseline traffic keeps GPUs busy around the clock. Many teams combine a reserved baseline with per-second capacity on top for peaks.
fal's B200 price
Prices as of October 15, 2026, taken from fal.ai/pricing.7 Check that page for the current rate before you plan a budget.
| GPU | VRAM | List price | As low as |
|---|---|---|---|
| B200 | 192 GB | $7.99/h | $5.49/h |
The list price is the per-GPU hourly rate, billed per second while a runner is alive. The "as low as" rate depends on your capacity arrangement, such as reserved runners or a usage commitment. Contact sales for those terms.
At the list price, one B200 second costs about $0.00222. Some worked examples:
| Workload | Billed GPU time | Cost at list price | Cost at "as low as" |
|---|---|---|---|
| 10-minute test run on 1 GPU | 600 GPU-seconds | $1.33 | $0.92 |
| 1 hour on an 8-GPU runner | 8 GPU-hours | $63.92 | $43.92 |
| 1 GPU kept warm for a 30-day month | 720 GPU-hours | $5,752.80 | $3,952.80 |
| Bursty traffic: 3 runner-hours a day, 30 days | 90 GPU-hours | $719.10 | $494.10 |
The last two rows show why billing granularity matters as much as the rate. The same B200 costs about eight times less per month when it runs only while traffic needs it.
Running a model on B200s with fal
The B200 is available on fal Serverless as the GPU-B200 machine type: 192 GB of VRAM, 210 GB of system RAM and 19 CPU cores per GPU.5 Serverless access is approved per account, and you can request it from the Serverless get-started page.
Select the B200 in your app definition:
import fal
class MyApp(fal.App):
machine_type = "GPU-B200"
num_gpus = 1
def setup(self):
... # load your model once per runner
@fal.endpoint("/")
def generate(self, input: dict) -> dict:
...Then deploy it with the CLI:
fal deploy my_app.py::MyAppFrom there:
- Start from the Quick Start to install the CLI and deploy a first app.
- Bring an existing container if you already have one. fal builds from your Dockerfile or runs an existing image. See Use a Custom Container Image.
- Tune scaling to control cost.
keep_alive,min_concurrencyandmax_concurrencydecide how many runners stay warm, which is most of what you pay for. See Scale Your Application. - Watch cost per app in App Analytics, which breaks down requests, runner utilization and spend.
FAQ
How much does an NVIDIA B200 cost?
NVIDIA does not publish a list price for the B200, and it is mostly sold inside 8-GPU servers. For most teams the useful number is the rental rate. fal publishes its B200 list rate and its lowest committed rate on fal.ai/pricing, and the price section above shows both as of the date given there.
Is B200 rental billed per hour or per second?
It depends on the platform. Hourly plans round up to the hour, and some rent only whole 8-GPU servers. fal Serverless bills per second while a runner is alive, and it does not bill for scheduling or image-pull time.
How much memory does a B200 have?
192 GB of HBM3e per GPU. In 8-GPU DGX B200 servers, NVIDIA lists 1,440 GB in total, about 180 GB usable per GPU.
Should I choose a B200 or an H200?
Choose the B200 when the model needs more than 141 GB, when the workload is limited by memory bandwidth, or when it can run in FP4. For models that fit comfortably on an H200 and run in FP8 or BF16, an H200 at a lower hourly rate can cost less per request.
Can I reserve B200 capacity on fal?
Yes. Reserved runners and usage commitments lower the effective hourly rate toward the "as low as" price on fal.ai/pricing. Contact the fal sales team for terms.
