Real NVIDIA AI hardware — from a desktop card to a full datacenter rack. Prices below are researched estimates; verify exact figures before you buy.
Desktop card
RTX 5090
NVIDIA's fastest consumer desktop graphics card, built on the Blackwell architecture.
32 GB GPU memory
GPU memory decides how big a model you can run — bigger models need more of it.
$1,999
Price is what you pay once to own the hardware outright — there's no subscription.
575 W
Power draw is how much electricity the hardware uses at full load — it drives your electricity bill and how much cooling you need.
Workstation card
RTX PRO 6000 Blackwell
A professional workstation card with much more memory than a desktop card, aimed at engineers and small AI teams.
96 GB GPU memory
GPU memory decides how big a model you can run — bigger models need more of it.
$8,565
Price is what you pay once to own the hardware outright — there's no subscription.
600 W
Power draw is how much electricity the hardware uses at full load — it drives your electricity bill and how much cooling you need.
Datacenter card
H100
NVIDIA's previous-generation datacenter GPU, still widely used for training and running large models.
80 GB GPU memory
GPU memory decides how big a model you can run — bigger models need more of it.
$27,500
Price is what you pay once to own the hardware outright — there's no subscription.
700 W
Power draw is how much electricity the hardware uses at full load — it drives your electricity bill and how much cooling you need.
Datacenter card (Blackwell)
B200
A current-generation Blackwell datacenter GPU with roughly 2.4x the memory of an H100.
192 GB GPU memory
GPU memory decides how big a model you can run — bigger models need more of it.
$35,000
Price is what you pay once to own the hardware outright — there's no subscription.
1,000 W
Power draw is how much electricity the hardware uses at full load — it drives your electricity bill and how much cooling you need.
Datacenter card
B300 / Blackwell Ultra
NVIDIA's highest-memory single datacenter GPU, built for the largest models and reasoning workloads.
288 GB GPU memory
GPU memory decides how big a model you can run — bigger models need more of it.
$50,000
Price is what you pay once to own the hardware outright — there's no subscription.
1,400 W
Power draw is how much electricity the hardware uses at full load — it drives your electricity bill and how much cooling you need.
Full rack (72 GPUs)
GB200 NVL72
A full liquid-cooled rack of 72 Blackwell GPUs, NVLink-connected so they share memory as one system.
72 GPUs, NVLink-pooled
13,400 GB GPU memory
GPU memory decides how big a model you can run — bigger models need more of it.
$3,000,000
Price is what you pay once to own the hardware outright — there's no subscription.
120,000 W
Power draw is how much electricity the hardware uses at full load — it drives your electricity bill and how much cooling you need.
Full rack (72 GPUs, Blackwell Ultra)
GB300 NVL72
NVIDIA's newest full rack: 72 Blackwell Ultra GPUs NVLink-connected into one giant shared-memory system.
72 GPUs, NVLink-pooled
20,700 GB GPU memory
GPU memory decides how big a model you can run — bigger models need more of it.
$3,850,000
Price is what you pay once to own the hardware outright — there's no subscription.
136,000 W
Power draw is how much electricity the hardware uses at full load — it drives your electricity bill and how much cooling you need.