NVIDIA B200: the Blackwell workhorse, explained

What the B200 is

Blackwell is NVIDIA's current architecture, and the B200 is its volume part, the generation that retired Hopper as the default for new AI clusters. Two reticle-size dies sit under one lid, joined by a 10 TB/s link so software sees a single 208-billion-transistor GPU, wired to 192 GB of HBM3e at 8 TB/s. It is also NVIDIA's first chip with a hardware FP4 path: the second-generation Transformer Engine runs 4-bit floating point natively, and that rewired the economics of inference.

Within Blackwell the B200 is the one that ships in quantity. It stands a full generation above the H100 and one memory tier below the B300 (Blackwell Ultra, 288 GB per GPU). For the large majority of training runs and production inference on current-size models, the B200 is the part you actually get, in the standard 8-GPU HGX B200 node that drops into conventional AI datacenters.

Against the H100 the B200 is a genuine generational jump, not a refresh. Memory more than doubles (192 GB versus 80 GB) and bandwidth rises from 3.35 TB/s to 8 TB/s, so larger models and longer contexts stay resident on a single GPU. Fifth-generation NVLink connects the eight GPUs in a node at 1.8 TB/s each, roughly double the H100's NVLink, which is what lets a node behave as one large memory pool rather than eight separate cards. For anyone standing up new capacity, the B200 is the baseline the market moved to.

The FP4 path is the other half of the story. Running production inference at 4-bit floating point roughly doubles throughput per GPU over FP8 and cuts memory for weights in half, which shows up directly as tokens per second and cost per token. The practical rule: choose the B200 as the current-generation default for training and high-volume inference; stay on the H100 only where it is materially cheaper and the workload already fits its 80 GB; step up to the B300 when the workload is memory-bound, long contexts, very large mixture-of-experts models, or big-batch serving, where the extra 96 GB per GPU is the deciding factor.

An HGX B200 node pools 8 GPUs into roughly 1.5 TB of NVLink-connected HBM3e (8 x 192 GB). Today's frontier open models are mixture-of-experts, so every expert has to sit in HBM even though only a slice fires per token: total parameters, not the active count, set the memory floor. Because the B200 runs FP4 natively, quantizing to 4-bit roughly halves the GPU count for the same weights.

Weights-only figures at published parameter counts. Production serving adds KV cache, activations and framework overhead, so real deployments need more headroom than the raw weights suggest. The 1.6T-parameter tier tips past a single B200 node at FP8, which is precisely the case the B300's larger 288 GB per GPU (a 2.3 TB node pool) is built to hold in one node.

B200 specs

Official NVIDIA Blackwell figures; compute is dense (sparse roughly doubles it). Exact memory-per-GPU varies by form factor and OEM (early DGX B200 units shipped 180 GB per GPU, 1.44 TB per node). Real inference throughput and cost per token depend on the model and serving stack, not raw FLOPS alone.

B200 vs H100

Against the H100 the B200 is a full generation ahead on every axis that matters for large models: more than double the memory, more than double the bandwidth, and a native FP4 path the Hopper H100 does not have. For new capacity the B200 is the baseline; the H100 stays relevant only where it is materially cheaper and the workload fits its 80 GB.

B200 vs B300

Same Blackwell generation, two differences: memory and dense compute. The B300 (Blackwell Ultra) carries 288 GB versus 192 GB and delivers roughly 1.5x the dense FP4 compute. If you are compute-bound on current-size models, the B200 is the volume part and usually the better economics; if you are memory-bound or your model spills past a single B200 node, the B300 consolidates the deployment.

What a B200 costs

The B200 prices exactly where the volume Blackwell part should: a clear step above the H100 generation, a step below the B300. On-demand rides above reserved, spot appears only in windows, and reserved terms land well below on-demand, which is where serious volume transacts. Era Compute tracks all of it daily across the major clouds and European datacenters and publishes it as one live index.

Open the live B200 price index for today's composite, the chart, and on-demand, spot and reserved levels side by side with what each provider charges. The full index covers the rest of the fleet, H100 through GB300.

B200 availability in Europe

Era Compute sources B200 capacity in European datacenters built for Blackwell-class density. As the volume Blackwell part it is the most deployable option in the region: standard 8-GPU HGX B200 nodes fit conventional air- and liquid-cooled facilities, and configurations run from a single node to multi-thousand-GPU clusters, with per-GPU pricing that falls as the cluster grows.

This is reserved capacity: contracted on multi-year terms and qualified per customer, not on-demand rental. For European buyers the model pays off twice. EU placement keeps training data and inference traffic inside GDPR and data-residency boundaries, and European facilities frequently undercut equivalent US capacity.

Questions, answered