Both use NVIDIA Blackwell Ultra GPUs. HGX B300 is an 8-GPU server, rented by the node from a few servers to large clusters. GB300 NVL72 is a liquid-cooled rack of 72 GPUs and 36 Grace CPUs in one NVLink domain, rented by the rack. As of 19 October 2026, B300 in our offers can start from Q4 2026. GB300 NVL72 racks start from Q1 2027. Pick GB300 NVL72 for large-model training and reasoning inference that gains from 72 GPUs sharing one NVLink domain; pick B300 for smaller blocks, earlier starts and sites without liquid cooling.
The GPU is the same generation. What changes is the unit you rent, how many GPUs talk over NVLink, and what the data center has to provide.
| HGX B300 | GB300 NVL72 | |
|---|---|---|
| Unit of rent | 8-GPU server (node) | Rack of 72 GPUs |
| NVLink domain | 8 GPUs | 72 GPUs |
| NVLink bandwidth | 14.4 TB/s per server | 130 TB/s per rack |
| GPU memory | 2.1 TB per server | 20 TB per rack; 37 TB of fast memory with the Grace CPUs |
| CPU | x86 host, per server design | 36 NVIDIA Grace CPUs |
| Cooling | Air or liquid, per server design | Liquid only |
| Power | About 14 kW per DGX B300 | About 132 kW per rack |
| Operators in our pipeline | 22 | 8 |
| Can start | From Q4 2026 | From Q1 2027 |
| Typical term | 1 to 4 years | 3 to 5 years |
Sources: NVIDIA HGX specifications, DGX B300, NVIDIA GB300 NVL72 specifications, Supermicro GB300 NVL72 rack (rack power), seen 13 October 2026. Offers as of 19 October 2026.
A GB300 NVL72 rack is built for about 132 kW (its power shelves total 132 kW) and is fully liquid-cooled, so it needs a site built for direct liquid cooling and high power per rack (Supermicro GB300 NVL72 rack). That is why its start dates follow data-center construction. B300 servers fit more existing halls.
Reserved B300 through Ancapex, our fee included, on terms of two years or longer: median about $5.40 per GPU-hour, middle half $4.90 to $6.10 (as of 19 October 2026). We have too few priced GB300 NVL72 offers to publish a rate without showing one operator's price; ask us for current quotes.
Requirements are handled confidentially and shared with operators only on a need-to-know basis, never with your name attached before you approve it. We reply within one business day from dima@ancapex.io.
Both use Blackwell Ultra GPUs. HGX B300 is an 8-GPU server with an 8-GPU NVLink domain; GB300 NVL72 is a liquid-cooled rack of 72 GPUs and 36 Grace CPUs with one 72-GPU NVLink domain at 130 TB/s.
Operators rent GB300 NVL72 by the whole rack, since the 72 GPUs share one NVLink domain. For smaller blocks, HGX B300 servers are the usual choice.
As of 19 October 2026 in our offers: B300 from Q4 2026; GB300 NVL72 from Q1 2027.
Yes. It is fully liquid-cooled and built for about 132 kW per rack (Supermicro), so the site must support direct liquid cooling.
We publish rates once we have enough priced offers that no single operator's price shows through. Send a request and we share current quotes after an NDA.
