B300 vs B200: which node should you reserve?

The B300 has 60% more memory per GPU (288 GB vs 180 GB) and 1.7× the dense FP4 compute. On Avian's 12-month term it costs 15% more per GPU-hour ($4.95 vs $4.32), which makes it 28% cheaper per GB of GPU memory. Prices as of October 2026.

B300 pricing B200 pricing

Specs side by side

Per GPU unless noted. Both ship as complete 8-GPU HGX nodes with NVLink across all eight.

	NVIDIA B200	NVIDIA B300
Architecture	Blackwell	Blackwell Ultra
GPU memory	180 GB HBM3e	288 GB HBM3e
Memory per 8-GPU node	1,440 GB	2,304 GB
Memory bandwidth	8 TB/s	8 TB/s
Dense NVFP4	9 PFLOPS	15 PFLOPS
NVLink	1.8 TB/s	1.8 TB/s
Power	1,000 W	1,400 W
Host memory per node	2 TB+	3 TB+

Reserved price by term

Per GPU-hour, prepaid for the whole term and locked from the day you reserve. Per-node figures are for one 8-GPU node per month (730 hours).

Term	B200 / GPU-hr	B300 / GPU-hr	B300 premium	B200 node / month	B300 node / month
1 Month (sold out)	$5.40	$5.40	0%	$31,536	$31,536
3 Months (sold out)	$4.95	$5.25	6%	$28,908	$30,660
6 Months (sold out)	$4.77	$5.15	8%	$27,857	$30,076
12 Months	$4.32	$4.95	15%	$25,229	$28,908
24 Months	$3.87	$4.59	19%	$22,601	$26,806
3 Years	$3.42	$4.23	24%	$19,973	$24,703

What you get per dollar

On the 12-month term. Memory usually decides how many nodes a model needs; FP4 compute decides how fast it serves.

	B200	B300
Price per GB of GPU memory, per hour	$0.0240	$0.0172
Price per dense NVFP4 PFLOP, per hour	$0.48	$0.33
Price per GPU-hour	$4.32	$4.95

Which one to reserve

If you can't decide from the lists below, the GPU memory calculator shows whether your model fits on each node.

Reserve the B300 if

Your model doesn't fit in 1,440 GB, or only fits with little room for KV cache.
You serve long contexts or many concurrent requests, where memory goes to KV cache.
You run FP4 inference and want the most throughput per node.
One B300 node replaces two B200 nodes for your model.

Reserve the B200 if

Your model fits comfortably in 1,440 GB with room to spare.
You train or fine-tune mid-size models where memory isn't the limit.
The lowest cost per node matters more than memory per GPU.
You want the longest track record: the B200 has been in production longer.

NVIDIA B300 pricingFrom $4.23 per GPU-hour, every term. NVIDIA B200 pricingFrom $3.42 per GPU-hour, every term. Reserved vs on-demandWhen prepaying for a node beats paying by the hour.