GB200 is not a single GPU. It is a superchip: two Blackwell B200 GPUs and one Grace CPU on one board, linked by NVLink. GB200 NVL72 is the rack built from 36 of those superchips, which makes 72 Blackwell GPUs and 36 Grace CPUs in one liquid-cooled system. That is why a search for a GB200 price rarely ends in a per-hour number.
GB200 vs GB200 NVL72
- GB200 superchip: 2 Blackwell GPUs + 1 Grace CPU (72-core Arm) on one module.
- GB200 NVL72: 36 superchips in a rack, so 72 GPUs and 36 CPUs in one NVLink domain.
- Cooling: liquid. The rack draws on the order of 120 kW, which is a datacenter project, not a server you slot in.
- Memory: roughly 13 TB of HBM3e across the rack, addressable by all 72 GPUs over the NVLink fabric.
The point of NVL72 is the interconnect. All 72 GPUs behave like one very large accelerator, with NVLink bandwidth that no InfiniBand cluster of ordinary 8-GPU servers matches inside a node. That matters for trillion-parameter models and large mixture-of-experts serving. It does not matter for a LoRA fine-tune.
GB200 price: what is actually known
NVIDIA does not publish a list price for GB200 or NVL72. You buy through an OEM or a cloud quote, and press and analyst estimates put a full NVL72 rack in the millions of dollars. Treat any single number as an estimate, not a catalog price. If a page quotes an exact GB200 hourly price, check whether it is on-demand, reserved or a sales quote.
That is also why we do not rank GB200 as an hourly row. We only rank published on-demand prices, and GB200 capacity is mostly sold as reserved clusters. See how we rank an hour.
GB200 vs B200: which one do you need
- B200 is a GPU. You rent it as a single card or an 8-GPU server by the hour. See the B200 price per hour.
- GB200 is a superchip or a rack. You reserve it, usually for weeks or months, through a neocloud or a hyperscaler.
- If your model fits in 8 GPUs, B200 is almost always the right call. NVL72 pays off when one model needs more than 8 GPUs with very fast GPU-to-GPU traffic.
- If you want more memory per GPU on the same generation, look at B300 with 288 GB.
Where to rent Blackwell today
For on-demand Blackwell, start with the B200 and B300 tables on GPUBeacon. For cluster-scale reserved capacity, the usual names are neoclouds such as CoreWeave and Nebius. Ask each for the same shape in writing: GPU count, interconnect, region and commitment. Compare them in CoreWeave vs Nebius.
If you are sizing a model before you commit, run it through the VRAM fit tool. Most teams asking about GB200 would be served by 8x B200 or 8x H200 at a fraction of the commitment. See the H100 vs H200 vs B200 guide.
FAQ: GB200 and NVL72
Is GB200 a GPU?
No. GB200 is a superchip with two B200 GPUs and one Grace CPU. The NVL72 rack combines 36 of them.
How much does a GB200 NVL72 cost?
NVIDIA does not publish a price. Estimates put a rack in the millions of dollars, and cloud access is usually a reserved quote rather than an hourly rate.
Can I rent a GB200 by the hour?
Rarely. It is mostly sold as reserved capacity. For hourly Blackwell, rent B200 or B300 instead.
Is GB200 faster than B200?
Per GPU the silicon is the same family. The gain is the 72-GPU NVLink domain, which helps very large models and big serving workloads, not small jobs.
Some outbound links in these articles may be affiliate links. Rankings on GPUBeacon stay independent.