NVIDIA B200
The B200 is NVIDIA's first-generation Blackwell data center GPU: two reticle-limited dies joined by a 10 TB/s NV-HBI link and packaged as one SXM6 module with 180 GB of HBM3e. It sells as a bare module only inside an 8-GPU HGX B200 baseboard, never standalone, and that baseboard is the building block for DGX B200 and third-party OEM systems (Supermicro, Dell, HPE, Lenovo). In a private AI build it is the training and large-batch inference workhorse for shops that need FP4 throughput and 180 GB of HBM per GPU but do not need the 288 GB tier or the extra power headroom of B300. As of mid-2026 it sits one rung below Blackwell Ultra (B300) in NVIDIA's current lineup and OEMs are shifting new quotes toward B300, but B200 fleets already deployed remain the majority of installed Blackwell capacity.
- The 180 GB vs 192 GB memory figure is a real discrepancy across NVIDIA's own materials, not a typo you introduced. Design and quote against 180 GB software-visible HBM3e; treat 192 GB as the physical stack size, not usable capacity.
- At 1000 W per GPU and eight GPUs per baseboard, air cooling is thermally possible but tight; most new HGX B200 deployments at rack scale go direct-liquid-cooled from day one rather than retrofitting later.
- H200 still makes sense if your workload is memory-bandwidth-bound inference on 70 to 100B-parameter dense models and you want to stay on a Hopper-compatible stack with no kernel retuning; B200 wins once you can exploit FP4 or need the larger NVLink domain for training.
- FP4 throughput numbers are peak, structured-sparsity figures. Real inference gains depend on whether your quantization pipeline actually produces NVFP4-compatible weights; dense FP8 remains the safer default for production serving until FP4 accuracy is validated on your model.
- Because B200 only exists on an 8-GPU baseboard, any capacity planning in units smaller than 8 GPUs is a fiction; plan procurement, power, and cooling in whole-baseboard increments.
How much does an NVIDIA B200 cost?
Reported pricing lands around $30,000 to $40,000 per GPU when purchased inside an 8-GPU HGX cluster, with an 8-GPU HGX B200 server running roughly $400,000 to $500,000 fully configured. Loose modules on secondary channels have traded higher, $45,000 to $55,000. These are reported market figures, not published NVIDIA list prices.
How much memory does the B200 actually have?
180 GB of HBM3e is the software-visible, spec-sheet number used by NVIDIA's own OEM documentation (Lenovo, Dell). Some earlier marketing referenced 192 GB, which reflects the physical HBM3e stack size before reserved capacity, not what a workload can allocate.
B200 vs B300: which one should I buy?
B300 (Blackwell Ultra) adds 288 GB of HBM3e versus B200's 180 GB, a higher 1,400 W power ceiling, and roughly 1.5x the dense NVFP4 throughput per GPU. Choose B200 for training and inference workloads that fit comfortably in 180 GB per GPU and where OEM allocation favors it; choose B300 for large-context or trillion-parameter-class reasoning workloads that need the extra memory headroom.
Can I buy a single B200 GPU?
No. B200 is an SXM6 module soldered to an 8-GPU HGX baseboard; there is no standalone PCIe B200 card. The practical purchase unit is the 8-GPU baseboard or a full DGX/OEM server built around it.