GB300 NVL72 / GB300 vs GB200
GB300 vs GB200 NVL72
Same rack shape, newer GPU. Here's what's actually different and when the older rack is the better deal.
Side by side
| Per GPU | GB200 NVL72 | GB300 NVL72 |
|---|---|---|
| GPU | Blackwell | Blackwell Ultra |
| HBM3e | 192 GB | 288 GB |
| FP4 compute | 1x | ~1.5x |
| Back-end network | 400 Gb/s | 800 Gb/s |
| GPUs per NVLink domain | 72 | 72 |
Figures from Microsoft's ND GB300 v6 documentation (source). AWS gives the same 1.5x memory and 1.5x FP4 comparison for P6e-GB300 against P6e-GB200 (source).
What the difference means in practice
Memory is the big one. 50% more HBM per GPU means larger models or longer contexts per GPU, and fewer GPUs for the same deployment. For inference on frontier-size models that can change how many racks you need.
Networking doubles. 800 Gb/s per GPU helps multi-rack training, where traffic between racks is the bottleneck.
Throughput. Microsoft reports about 27% more LLM inference throughput per rack on GB300 than on GB200 in its own test (source). Your gain depends on the model and how memory-bound it is.
When GB200 is the better rent
GB200 NVL72 racks have been in service longer, so more of them come free as early contracts end. If your model fits comfortably in 192 GB per GPU, or you need racks sooner than GB300 can be delivered, a GB200 quote is worth asking for alongside GB300. Our form lets you request both.
Questions
Is GB300 worth it over GB200?
Can GB300 and GB200 racks share a cluster?
RackShore runs this site and may be paid a referral fee by the provider you choose. Buyers pay us nothing. Spot a mistake? Tell us.