Blackwell has been shipping in one form or another for a while now, and the naming has become a mess. There is a Blackwell in the card in your desktop, a Blackwell in a 14 kW server, and a Blackwell sold as an entire liquid-cooled rack. They share an architecture and almost nothing else. Here is how the parts differ and which differences actually matter.
Why does Blackwell have so many names?
Blackwell is the architecture. The B200 is the mainstream data centre part. The GB200 pairs Blackwell GPUs with NVIDIA’s Arm-based Grace CPU on one board. Blackwell Ultra, sold as the B300 and in rack form as the GB300 NVL72, is the mid-generation refresh.
The pattern is consistent once you see it. A leading G means a Grace CPU is attached. NVL72 means 72 GPUs wired as a single unit. Ultra means more memory, not a new architecture.
Is this the same Blackwell as in an RTX 5090?
Same architecture family, different silicon and a completely different memory system. The 5090 carries 32 GB of GDDR7 on a 512-bit bus at 575 W. The RTX PRO 6000 Blackwell workstation card pushes that to 96 GB of GDDR7 at 600 W, which is why it has become the default for local AI work.
The data centre parts break from that with HBM3e stacked on-package. A single B200 carries 180 GB of it at 8 TB/s, several times what even a 512-bit GDDR7 bus delivers. That memory system, not the shader count, is the real dividing line between the consumer and data centre sides of the same architecture family [HOST LINK]. Everything else follows from it.
What does a B200 actually give you?
The standard building block is the DGX B200: eight B200s in a 10U chassis, 1,440 GB of GPU memory, 64 TB/s of aggregate memory bandwidth, tied together by fifth-generation NVLink through two NVSwitches at 14.4 TB/s. NVIDIA rates it at 144 petaFLOPS of FP4. It draws about 14.3 kW, which is roughly ten houses.
FP4 doing the headline work is worth noting. The march toward smaller number formats is most of the recent performance story, and inference tolerates the precision loss far better than anyone predicted five years ago. Training still leans on FP8 and above.
What changes with GB300 NVL72?
The GB300 NVL72 is not a server, it is a rack sold as one computer: 72 Blackwell Ultra GPUs and 36 Grace CPUs, liquid cooled, joined by an NVLink fabric moving 130 TB/s. Each GPU carries roughly 279 GB of HBM3e, about 1.5 times a B200. The rack holds 20 TB of GPU memory and another 17 TB of LPDDR5X on the Grace side that the GPUs can address directly, for 37 TB of fast memory total.
The architecture matters more than any single number here. Software addresses one large memory pool rather than 72 cards that have to be coordinated across a slower interconnect. For a model that does not fit on one GPU, that is the difference between running and not running.
Why does memory matter more than FLOPS now?
Reasoning models changed the shape of the workload. A model that works through a problem step by step holds a growing context in memory for the whole generation, and serving many users at once multiplies it. The constraint that bites first is usually not arithmetic throughput, it is whether the weights and every in-flight conversation fit in memory at all.
That is why Blackwell Ultra’s headline improvement is capacity rather than a new architecture, and why NVIDIA now sells racks instead of cards. A decade of optimising for training throughput has given way to inference, and inference is a memory problem.
What about those performance numbers?
Here is the footnote that rarely survives into a headline. NVIDIA rates the GB300 NVL72 at 1,440 petaFLOPS of FP4, about 1.4 exaFLOPS in one rack. That figure is measured with sparsity enabled, a technique that skips zero-valued weights. Dense performance, every weight processed, is meaningfully lower, and published dense figures disagree between sources.
Sparsity is a real capability and NVIDIA footnotes it properly. The problem is downstream, where the footnote gets dropped. When you see two systems compared, check whether both numbers are sparse or both dense before concluding anything. Mixing them skews the comparison by a third or more.
What if you are not buying a rack?
Almost nobody is. A GB300 NVL72 is a seven-figure purchase needing liquid cooling and power density most buildings do not have, so access to this hardware is overwhelmingly rented. Providers buy the racks and sell time on them by the GPU-hour, or by the token for inference.
That has made availability rather than price the thing buyers compete over, since demand from the large AI labs absorbs much of each production run. Regional providers have become a real alternative, particularly where data protection rules make the physical location of the hardware a requirement rather than a preference. That includes European operators: capacity on a European GPU compute platform powered by NVIDIA B200 can be rented by the hour today, with GB300 NVL72 reservations opening for 2027.
So what should you take from this?
Three things. The B200 is the part you will actually encounter, and 180 GB of HBM3e per GPU is the number worth remembering. The GB300 NVL72 matters because it presents a rack as a single machine with one memory pool, which is what large reasoning models need and what no collection of PCIe cards can imitate. And every performance figure in this space carries a footnote, so find it before you quote it.