GPU families

The six accelerators we surface in the catalog, with VRAM, architecture, and indicative pricing.

GPU Outlet’s catalog is intentionally narrow. We surface six flagship NVIDIA families. Anything older than RTX 4090 doesn’t make the cut — the price / performance vs the 4090 is just too unfavorable.

The six families

Family	Arch	VRAM	From $/hr	Best for
RTX 4090	Ada Lovelace	24 GB	$0.65	Hobby fine-tuning, inference at 7B–13B params
RTX 5090	Blackwell	32 GB	$1.15	Better-than-4090 single-card workloads
A100 80GB	Ampere	80 GB	$1.40	Large fine-tunes, multi-GPU training
H100 80GB	Hopper	80 GB	$2.80	FP8 training, frontier-model fine-tunes
B200	Blackwell	192 GB	$6.00	Trillion-parameter training, large LLM inference
B300	Blackwell	288 GB	$7.50	The largest models the open ecosystem ships today

Prices are from — the minimum hourly rate currently offered by any upstream provider. The dashboard shows the live cheapest listing when you pick a card. See Pricing for current values.

Multi-GPU configurations

Every family supports 1×, 2×, 4×, and 8× configurations. Some providers cap available count per pod — the dashboard greys out unavailable multipliers in real time.

RTX 4090: 1× / 2× / 4× / 8×
RTX 5090: 1× / 2×              (4×, 8× rarely available)
A100:     1× / 2× / 4× / 8×
H100:     1× / 2× / 4× / 8×
B200:     1× / 2× / 4× / 8×
B300:     1× / 2× / 4×         (8× very limited)

What’s NOT in the catalog

We’re often asked about the L4, L40, L40S, V100, T4, and consumer cards below 4090. Short answer: we tried them, the price/perf isn’t there for our buyer profile. If you have a specific need, email [e-mail].

Availability tiers

The landing page shows three dots next to each GPU as an at-a-glance signal:

• 3 dots / High — at least 100 listings live across all upstream providers
• 2 dots / Medium — 20–99 listings
• 1 dot / Low — under 20 listings, may be hard to source on demand

Live availability behind the login at gpuoutlet.ai/app/create is precise. The landing dots are a marketing summary.