The NVIDIA B200 is a Blackwell GPU used in HGX and DGX server systems. For a buyer, the first decision is the system: GPU count, server vendor, cooling, networking, and support terms. Those choices determine the price and delivery date.
The essentials
| Item | HGX B200 |
|---|---|
| GPU memory | 180 GB HBM3e per GPU |
| Eight-GPU node memory | 1.44 TB HBM3e |
| Memory bandwidth | Up to 8 TB/s per GPU |
| GPU interconnect | Fifth-generation NVLink and NVSwitch |
| GPU power | Configurable up to 1 kW per GPU |
| Cooling | Air or liquid, depending on the OEM server |
These are NVIDIA's HGX B200 specifications. The 180 GB figure is for the B200 in an HGX system. GB200 NVL72 is a different rack-scale platform with Grace CPUs and should be quoted separately.
What changes from H100 and H200
H100 SXM has 80 GB per GPU. H200 SXM has 141 GB. HGX B200 has 180 GB. That extra memory can help with larger models, longer contexts, and batch size, but a model that fits in memory still needs a workload benchmark to establish which system will be faster or cheaper per result.
B200 also adds FP4 support and a faster GPU interconnect. The benefit depends on precision, model, software stack, and how many GPUs participate in the job. Ask for a benchmark using your workload if cost per token or training time drives the purchase.
| Question | Why it matters |
|---|---|
| Does your model require CUDA-specific libraries or custom kernels? | This may favor staying within NVIDIA's stack. |
| Will you use one GPU, one eight-GPU node, or several nodes? | The server and network design change with scale. |
| What precision will you actually run? | FP4 gains do not apply to every model or quality requirement. |
| What cooling can your facility support? | OEM B200 servers can be air- or liquid-cooled. The specific chassis decides the facility requirement. |
HGX, DGX, and GB200
HGX B200 is NVIDIA's eight-GPU baseboard integrated into servers from OEMs. CPU, memory, storage, NICs, cooling, and service coverage vary by server. NVIDIA's reference architecture also discusses four-GPU designs.
DGX B200 is NVIDIA's complete eight-GPU system. Its configuration and support package are part of the quote.
GB200 NVL72 is a rack-scale Grace Blackwell system. It has different infrastructure, delivery, and service requirements from an HGX B200 server. Do not use a GB200 quote to estimate the cost of an HGX build.
Pricing and delivery
B200 pricing and delivery depend on the OEM build, GPU allocation, quantity, destination, support, and quote date. A quoted GPU price alone is not the delivered server cost.
For a useful quote, send:
- The number of GPUs or servers and whether you need HGX, DGX, or a specific OEM chassis.
- Required CPU, RAM, storage, networking, and cooling, if already known.
- Delivery location and the date you need the equipment.
- Any requirements for OEM warranty, on-site support, or installation.
If you are still choosing the configuration, tell us the workload and facility limits. We can use those details to request the right build and return the confirmed price, lead time, and warranty terms in writing.
Request a B200 quote | Browse hardware
