How to read this page. Each section starts Simple, then goes to a Deep dive: stop wherever you have what you need. The small numbers are sources: click one to open the original document. Where the maker has not published a figure -- a price, a die size, a factory address -- this page says so rather than estimate it.
1.At a glance
SimpleStart here
The B300 is NVIDIA’s Blackwell Ultra GPU, announced 18 March 2025 at GTC as part of a platform NVIDIA said would “pave the way for the age of AI reasoning” 1. It ships inside HGX B300 boards and DGX B300 systems.
Deep diveThe technical detail
NVIDIA’s own HGX product page confirms “HGX B300 and HGX B200 shipping now” 2; the DGX SuperPOD release said DGX B300 partner availability was expected “later this year” as of the March 2025 announcement 3.
2.Launch and history
SimpleStart here
B300 is a mid-cycle refresh of Blackwell rather than a new architecture generation — NVIDIA positioned it as the “Ultra” step between B200 and the next full generation, Rubin, aimed squarely at the more memory-hungry reasoning-style AI models that followed the first wave of chatbots.
Deep diveThe technical detail
Announced 18 March 2025 alongside GB300 at the same GTC keynote 1. NVIDIA framed the release around “AI reasoning” workloads — models that think through multi-step answers rather than respond in a single pass — which need more memory per GPU than B200 offered. DGX B300 is air-cooled, distinguishing it from the liquid-cooled GB300 NVL72 rack released the same day 3.
3.What’s inside it
SimpleStart here
B300 uses the same compute die as B200 — two reticle-limit dies joined by a 10 terabyte-per-second link — with more memory and a higher power limit rather than new logic 4.
Deep diveThe technical detail
NVIDIA’s own architecture page and developer blog both give B300 the identical headline figures published for B200 at GTC 2024: 208 billion transistors on a custom TSMC 4NP process, in a dual-die package 54. What changes: memory rises from 192GB to 288GB of HBM3e per GPU via twelve 12-Hi stacks (up from eight 8-Hi stacks on B200), bandwidth reaches 8 TB/s per GPU, NVFP4 compute rises 1.5x to 15 dense petaflops per GPU, and attention-layer throughput doubles via faster special-function units 4. NVIDIA lists 640 fifth-generation Tensor cores, 20,480 CUDA cores and 160 SMs across 8 GPCs 4.
4.Full spec table
SimpleStart here
An 8-GPU HGX B300 board delivers 2.1TB of total memory and up to 144 petaflops of sparse FP4 compute 2; a full DGX B300 system rates 144 petaflops of inference and 70 petaflops of training performance 6.
Deep diveThe technical detail
| Spec | HGX B300 (8-GPU board) |
|---|
| Process node | TSMC 4NP (custom) |
|---|
| Transistors per GPU | 208 billion |
|---|
| Memory per GPU | 288GB HBM3e, 8 TB/s |
|---|
| Total board memory | 2.1TB |
|---|
| FP4 Tensor (dense/sparse) | 108 / 144 PFLOPS |
|---|
| FP8/FP6 Tensor | 72 PFLOPS |
|---|
| FP32 | 600 TFLOPS |
|---|
| NVLink (5th gen) | 1.8 TB/s per GPU, 14.4 TB/s per board |
|---|
| Networking | Up to 1.6 TB/s |
|---|
Board-level figures per NVIDIA’s own HGX page 2. DGX B300 additionally lists 144 petaflops inference / 70 petaflops training at the system level, 8x ConnectX-8 (800 Gb/s) and 2x BlueField-3 (400 Gb/s) networking, and a 10U form factor 67.
5.Where it’s made
SimpleStart here
B300 is made by TSMC, on the same custom 4NP process NVIDIA used for B200 5. NVIDIA does not operate its own chip factories.
Deep diveThe technical detail
NVIDIA’s Blackwell architecture page names TSMC and the custom 4NP node for the whole Blackwell Ultra generation 5. Neither NVIDIA nor TSMC discloses which specific fab site produces B300 wafers.
6.Which systems use it
SimpleStart here
B300 ships in NVIDIA’s own HGX B300 (8-GPU board) and DGX B300 systems, and as the GPU inside the GB300 superchip. NVIDIA named Cisco, Dell, HPE, Lenovo and Supermicro as server partners 1.
Deep diveThe technical detail
Cloud partners named at launch include AWS, Google Cloud, Microsoft Azure and Oracle Cloud Infrastructure 1. On the pricing side, CoreWeave lists HGX B300 as an available node type 8. DGX B300’s host CPU is an Intel Xeon 6776P, not NVIDIA’s own Grace — that pairing is reserved for the GB300 superchip, covered on its own page 7.
7.Official pricing
SimpleStart here
NVIDIA does not publish a unit price for B300. CoreWeave’s own pricing page lists HGX B300 spot pricing at $35.84/hour (North America) and $36.70/hour (Europe) per 8-GPU node; on-demand pricing is “Contact sales” 8.
Deep diveThe technical detail
CoreWeave’s pricing page, checked September 2026, is the only official published pricing found for B300: HGX B300 spot at $35.84/hour in North America and $36.70/hour in Europe per 8-GPU node, with on-demand rates listed as “Contact sales” rather than a published figure 8. No B300 pricing was found published on AWS, Google Cloud, Azure or Oracle Cloud’s own pricing pages as of this check.
9.What came before, what came next
SimpleStart here
Came before: B200 (Blackwell), announced March 2024. Came after: Rubin, officially announced 5 January 2026 at CES 10.
Deep diveThe technical detail
B300 succeeds B200, reusing the same compute die with a memory and power refresh rather than a new architecture. NVIDIA’s own newsroom names Rubin as the platform that succeeds Blackwell 10, first shown at GTC 2025 the same day as B300 itself and declared in full production at CES on 5 January 2026.
10.Hidden in plain sight
SimpleStart here
“Ultra” sounds like a new chip. NVIDIA’s own numbers say it is not: B300 uses the identical 208-billion-transistor die NVIDIA published for B200 a year earlier.
Deep diveThe technical detail
NVIDIA’s Blackwell architecture page and its Blackwell Ultra developer blog both state 208 billion transistors on a TSMC 4NP dual-die package — the same figures NVIDIA gave for B200 at GTC 2024 54. The generational jump from B200 to B300 is almost entirely a memory and packaging refresh (192GB to 288GB of HBM3e, denser 12-Hi stacks, a higher power envelope) rather than new compute logic — which is also why cloud partners like CoreWeave and Microsoft Azure could stand up B300-generation deployments within months of their B200 rollouts.