Learn / GPU hardware
NVIDIA B300: the Blackwell Ultra GPU, explained
288 GB of HBM3e, roughly 1.5x the compute of the B200, and the accelerator behind today's frontier training clusters. When a model is too big or a context too long to fit anywhere else, the B300 is the answer. Here is what it does, what it costs, and how to reserve it in Europe.
B300 package and board. Click to explore the full module in 3D.
What the B300 is
The B300 is NVIDIA's Blackwell Ultra: the B200 with the memory turned up. Same generation, same FP4 Transformer Engine, but 288 GB of HBM3e per GPU, half again more than the B200. That headroom is the whole product. It lets a single node hold larger models, longer contexts, and bigger KV caches without sharding, which is why the B300 became the default request for frontier training and high-end inference in 2026.
It ships two ways: as HGX B300, the 8-GPU server node that slots into conventional AI datacenters, and inside GB300 NVL72, the liquid-cooled rack that fuses 72 GPUs into one NVLink domain. The GPU is the same; the form factor decides your facility requirements and your procurement path.
Before the numbers, look at the thing itself. The exploded view above shows every physical layer, from the anodized top plate through the fin stack, vapor chamber, HBM3E stacks and compute dies down to the mezzanine connectors that drop the module onto the HGX baseboard. Click it to take the module apart yourself in 3D.
B300 specs
| Architecture | Blackwell Ultra, dual reticle-size dies, 10 TB/s die-to-die |
|---|---|
| GPU memory | 288 GB HBM3e per GPU (12-high stacks) |
| Memory bandwidth | 8 TB/s per GPU |
| Dense FP4 compute | 15 PFLOPS per GPU (1.5x B200) |
| Power (TDP) | 1,400 W per GPU |
| Scale-out networking | ConnectX-8, up to 1.6 Tb/s per GPU |
| Node | HGX / DGX B300, 8 GPUs, 2.1 TB GPU memory |
| Rack | GB300 NVL72, 72 GPUs, liquid-cooled, 1.1 exaFLOPS dense FP4 |
Official NVIDIA Blackwell Ultra figures; exact clocks vary by form factor and OEM. Real inference throughput and cost per token depend on the model and serving stack, not raw FLOPS alone.
B300 vs B200
Same generation, one deciding difference: memory. If your workload is memory-bound, long-context inference, large mixture-of-experts models, or big-batch serving, the B300's extra 96 GB per GPU is worth its premium. If you are compute-bound on current-size models, the B200 remains the value pick while B300 supply stays tight.
| Dimension | B200 | B300 |
|---|---|---|
| GPU memory | 192 GB HBM3e | 288 GB HBM3e |
| AI compute | 1x (baseline) | ~1.5x |
| Best fit | Current-gen training | Frontier training, long-context inference |
| Availability | Ramping, broadening | Tight, mostly reserved |
| Price | Lower | Premium; see the live indices |
B300 for inference
The B300 was built for the inference era as much as for training. Two things drive that. First, FP4 precision: Blackwell Ultra runs production inference at 4-bit floating point with roughly 15 PFLOPS of dense compute per GPU, which translates directly into tokens per second and cost per token. Second, memory: at serving time the KV cache, the per-conversation working memory of an LLM, competes with model weights for HBM. 288 GB per GPU means longer contexts and more concurrent sessions per GPU before you shard or evict.
In practice that makes the B300 the strongest single-node serving platform available: reasoning models that generate long chains of thought, agents holding large contexts, and high-concurrency APIs all benefit from memory headroom more than from raw FLOPS. If your inference bill is dominated by context length or batch size, the B300 premium over the B200 usually pays for itself in consolidation.
What runs on a single HGX B300 node
An HGX B300 node pools 8 GPUs into 2.3 TB of NVLink-connected HBM3e. 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, sets the memory floor. Weights-only arithmetic (1 byte per parameter at FP8, half that at FP4) shows what fits where:
| Model | Total params | Weights at FP8 | Fits on |
|---|---|---|---|
| Qwen 3.5 | 397B MoE | ~397 GB | 2x B300 (1x at FP4) |
| GLM-5.2 | 744B MoE | ~744 GB | 3x B300 (2x at FP4) |
| Kimi K2 | 1T MoE | ~1.0 TB | 4x B300 (2x at FP4) |
| DeepSeek V4 Pro | 1.6T MoE | ~1.6 TB | One 8-GPU node, KV headroom to spare |
Weights-only figures for today's frontier open models, at published parameter counts. Production serving adds KV cache, activations and framework overhead, which is exactly why the memory pool is the binding constraint: it is the difference between one node and a sharded deployment.
InfiniBand or Ethernet: what your B300 cluster actually needs
Inside a node, NVLink moves data between the 8 GPUs and the network does not matter. The question is what connects nodes to each other, and the answer depends on the workload.
Training across nodes needs a fabric. Every training step synchronizes gradients across all GPUs, so multi-node training pods are built on InfiniBand (NDR/XDR class) or 400 to 800G RoCE Ethernet tuned for RDMA. Skimping here idles the most expensive silicon you can rent: a slow fabric shows up as GPUs waiting, not as a line item.
Single-node inference does not. If the model fits inside one HGX B300 node, NVLink does the heavy lifting and the external network only carries requests and responses. Standard datacenter Ethernet is fine, and paying an InfiniBand premium buys nothing.
This matters commercially: fabric choice moves the price per GPU-hour and narrows which European facilities qualify. When you reserve through Era Compute, state whether the workload is multi-node training or single-node serving, and we match capacity with the right (and no more than the right) network.
What a B300 costs
The B300 prices at the top of the market, and while Blackwell Ultra supply stays tight it stays there: on-demand carries a clear premium over the H100 generation, spot appears only in windows, and reserved terms land well below on-demand, which is where serious volume transacts. Era Compute tracks every level daily across the major clouds and European datacenters and publishes it as one live index.
Open the live B300 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.
B300 availability in Europe
Era Compute sources NVIDIA B300 capacity in Central Europe, in Tier 3 datacenters with the power and cooling for Blackwell-class density. Two form factors, two timelines: air-cooled HGX B300 nodes are the near-term option, deployable at the scale of a few nodes; liquid-cooled GB300 NVL72 clusters and larger colocation come online from 2027 onward. Configurations run from a single node up to multi-thousand-GPU clusters, and per-GPU pricing comes down 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. Tell us the GPU count, cooling and timing you need and we match it.
Need B300 or GB300 capacity in Europe? Tell us the scale and timing, and we return matched offers from European datacenters.
Reserve B300 capacityB300 questions, answered
What is the NVIDIA B300?
The B300 is NVIDIA's Blackwell Ultra datacenter GPU: the memory-expanded refresh of the B200 with 288 GB of HBM3e per GPU. It targets frontier model training and long-context inference, and ships in HGX B300 8-GPU nodes and rack-scale GB300 NVL72 systems.
How is the B300 different from the B200?
Same Blackwell generation, two main differences: memory and dense compute. The B300 carries 288 GB HBM3e versus 192 GB on the B200 (50% more), and delivers roughly 1.5x the AI compute, with an improved thermal design. For memory-bound workloads such as long-context inference, the extra 96 GB per GPU is the deciding factor.
How much does a B300 cost to rent?
B300 rates move with supply. Era Compute publishes a live B300 price index with the daily composite, on-demand, spot and reserved levels, and what each provider charges. The durable rule while supply is tight: reserved capacity prices well below on-demand.
What is the difference between HGX B300 and GB300 NVL72?
HGX B300 is the 8-GPU x86 server building block, deployable in standard air- or liquid-cooled datacenters. GB300 NVL72 is a rack-scale system: 72 B300-class GPUs plus Grace CPUs in one NVLink domain, liquid-cooled, sold effectively as a full rack. Clusters are built from either, and the right choice depends on workload scale and facility capability.
Is the B300 available in Europe?
Yes. Era Compute sources B300 capacity in Central Europe in Tier 3 datacenters. Air-cooled HGX B300 nodes are available near-term at the scale of a few nodes; liquid-cooled GB300 NVL72 clusters and larger colocation come online from 2027. It is reserved capacity on multi-year terms, qualified per customer, rather than on-demand rental.
Should I rent B300 on-demand or reserve capacity?
For sustained training or production inference, reserved capacity is usually the better economics: multi-year terms price well below on-demand, and reservation is often the only way to get meaningful B300 volume at all while supply is constrained.