macrostack

NVIDIA · Blackwell Ultra · Layer 2

NVIDIA B300

Large-model inference, reasoning models above all, where NVIDIA’s own figures put it ahead of the B200: 1.5 times the dense FP4 compute, twice the attention performance and 2.1 TB of memory per 8-GPU board against 1.4 TB.

rent Rentable by the hour from multiple clouds; purchasable at scale.

Specifications

Memory
2.1 TB per 8 GPUs (no per-GPU figure)
Bandwidth
Not stated on NVIDIA’s DGX or HGX B300 pages
Compute
Per 8-GPU board: FP4 144 PFLOPS sparse, 108 dense; FP8/FP6 72 PFLOPS and FP16/BF16 36 PFLOPS, both sparse (NVIDIA’s HGX table)
Power
About 14 kW for a DGX B300 system of 8 GPUs; no per-GPU figure on NVIDIA’s pages
Interconnect
NVLink 5: 1.8 TB/s GPU to GPU, 14.4 TB/s across 8 GPUs; ConnectX-8 networking at up to 800 Gb/s
Software
CUDA
Form factor
accelerator
Workload
both

Verified 2026-09-09. Fields reading “not published” are exactly that — we do not estimate a figure a vendor withholds.

The catch

It is built for AI alone. NVIDIA’s HGX table cuts FP64 to 10 TFLOPS per board (the B200 board: 296) and INT8 to 3 POPS (the B200: 72), so scientific double-precision work and INT8 inference belong on other parts.

The economics

NVIDIA sells it per megawatt: Blackwell Ultra systems, it says, give up to 50 times the throughput per megawatt of Hopper on low-latency agentic work (SemiAnalysis InferenceX benchmarks, Q1 2026). That is a vendor claim for one workload, not a general multiple.

Cost per million tokens is the metric that decides most real purchases, and there is no neutral benchmark for it — every published figure comes from a company selling one side of the comparison. Read all of them, ours included, as claims rather than measurements.

The power question underneath this

A rack of frontier accelerators draws 120–200 kW against a 2026 average of about 27 kW, and the US grid interconnection queue exceeds 2,600 GW with waits approaching five years. Whether you can energise this part is now a harder question than whether you can buy it.

Layer 1 — Energy · The interconnection queue · Rack power density · Tokens per watt

Where you actually rent this

Choosing the part is the easy half. The same accelerator rents for wildly different money depending only on who you rent it from — the cheapest and dearest published on-demand price for an H100 sit 5.6× apart — and the provider, not the chip, is what decides whether you can leave later.

Layer 2.5 — Compute providers · The GPU cost calculator · B300 rental prices, every company

Related silicon

Sources

Want this whole layer as one table?

The accelerators on these pages as a single CSV — memory, bandwidth, compute, power, interconnect, software stack, availability, who each part is for and the catch — ready to sort, filter or load into a notebook.

Get the silicon table — free for research and evaluation

Four short lines and the download starts at once (CC-BY-NC-4.0, built from these pages the moment you ask). What you type is used only to answer the request and is never sold. Prefer Hugging Face? Request it there.

Using it inside a product, a client report or anything you sell? That needs a commercial licence.

Macrostack does not sell silicon and takes no commission at this layer — there is no affiliate programme for compute, which is exactly why nobody publishes a neutral comparison of it.

The Macrostack brief

New swaps, worth your inbox.

A short, occasional email when we add a high-intent alternative or ship a new head-to-head. No spam, no selling your address — unsubscribe in one click.