Is H200 worth it over H100?

Yes when your workload is memory-bound. H200 has 76% more memory and 43% more bandwidth per GPU on the same Hopper architecture, so large models and long contexts run on fewer GPUs.

Can I buy B300 servers in Australia?

Yes, through OEM channels such as Supermicro, Dell and Lenovo, with lead times set by allocation. Send us your configuration and we quote a landed AUD price with the lead time.

Buyer guide

H100 vs H200 vs B300: which GPU server should an Australian team buy?

By Nicholas Dodemaide, Director · Updated 15 November 2026 · 6 min read

Buy H200 if your models are memory-bound, B300 if you need frontier scale and can wait for OEM allocation, and a used H100 node if cost per GPU matters most. In late 2026 an 8-GPU HGX H200 server lists for about US$320,000 to US$420,000, a used HGX H100 for about US$225,000 to US$275,000, and Supermicro's HGX B300 from US$536,771.

The short answer

Memory decides it. H100 has 80 GB per GPU, H200 has 141 GB and B300 has 288 GB. Pick the smallest platform that holds your model, context length and batch size in GPU memory on a single node; crossing nodes costs more than the price gap between platforms.

Specifications side by side

8-GPU platform	Memory per GPU	Memory per node	Bandwidth per GPU	NVLink per GPU
HGX H100	80 GB HBM3	640 GB	3.35 TB/s	900 GB/s
HGX H200	141 GB HBM3e	1,128 GB	4.8 TB/s	900 GB/s
HGX B300	288 GB HBM3e	about 2.3 TB	8 TB/s	1.8 TB/s

NVIDIA published specifications.

What they cost in late 2026

Server	Indicative price (USD)	As at	Source
HGX H100 8-GPU, used	US$225,000 to US$275,000 (median US$250,000)	July 2026	RackRate
HGX H200 8-GPU, new	US$320,000 to US$420,000	2026	Mercatus
HGX H200 8-GPU, used	about US$348,500	August 2026	American Compute
HGX B300 8-GPU, new	from US$536,771 to US$662,846	Sep to Oct 2026	Supermicro eStore via ancapex

List, index and asking prices, not completed sales. Australian landed cost adds freight, insurance, clearance and 10% GST.

Rental prices give a second check. H100 rental has fallen to roughly US$1.70 to US$2.63 per GPU-hour on specialist clouds, while H200 lists at about US$2.60 per GPU-hour on-demand at GMI Cloud. If your utilisation is steady, compare three years of rental against the landed price of a node before you decide.

How to choose

Model fits in 640 GB with room for context: a used HGX H100 node is the lowest cost per GPU.
Model or KV cache needs more than 640 GB, or you serve long contexts: HGX H200.
Reasoning models, very long context or frontier training: HGX B300, ordered through the OEM channel.
No HGX baseboard in your rack: H200 NVL or RTX PRO 6000 PCIe cards in a standard server.

Buying safely

Ask for serial numbers before you pay, and have the OEM or NVIDIA confirm them against an authorised sale. Data-centre GPUs held offshore may carry export-licence conditions, so the paperwork matters as much as the price.

Common questions

Is H200 worth it over H100?: Yes when your workload is memory-bound. H200 has 76% more memory and 43% more bandwidth per GPU on the same Hopper architecture, so large models and long contexts run on fewer GPUs.
Can I buy B300 servers in Australia?: Yes, through OEM channels such as Supermicro, Dell and Lenovo, with lead times set by allocation. Send us your configuration and we quote a landed AUD price with the lead time.