Five real products, small to enormous. Every number below has a plain-words explanation next to it.
Desktop
GeForce RTX 5090
The entry point -- a serious gaming card that hobbyists also use to experiment with AI.
$2,999
Launch price was $1,999, but a 2026 GDDR7 memory shortage has pushed real street prices up -- as high as $4,600+ at some retailers. $2,999 shown here is a genuine current price, not the cheapest or most expensive you'll see.
Memory32 GB
32GB holds a small AI model entirely in fast on-card memory. It's plenty for gaming or light AI experiments, but too little on its own for any model on our Model Advisor below.
Power draw575 W
About the same draw as 8 incandescent light bulbs running at once.
Workstation
RTX PRO 6000 Blackwell (Workstation Edition)
NVIDIA's professional desktop card -- built for engineers, not gamers.
$12,000
Launch MSRP was $8,565 in 2025; current 2026 retail runs $11,000-$13,300 depending on the seller, driven by the same memory shortage affecting the RTX 5090. $12,000 is a representative current price.
Memory96 GB
96GB is the largest memory of any single desktop-style card NVIDIA sells. That much room on one card means fewer cards needed to fit a model.
Power draw600 W
About the same draw as a small window air-conditioning unit.
Datacenter GPU
NVIDIA H200 SXM
A purpose-built datacenter chip -- not sold standalone, only inside servers.
$35,000
NVIDIA doesn't sell this chip by itself -- it only ships inside 8-GPU servers. $35,000 is an estimated per-GPU price backed out of real full-server quotes (around $308,000-$315,000 for an 8-GPU H200 server), so treat it as a solid estimate rather than a firm quote.
Memory141 GB
141GB of ultra-fast HBM3e memory -- built specifically to hold and run large AI models, unlike the GDDR7 memory in the desktop and workstation cards above.
Power draw700 W
About the same continuous draw as a small electric space heater.
Server
HGX B200 8-GPU Server
A complete, ready-to-run server holding 8 datacenter GPUs already wired together.
$394,406
Real starting price from a posted reseller listing (Supermicro SYS-A21GE-NBRT-G1, an NVIDIA HGX B200-class platform, the same generation as NVIDIA's own DGX B200). Configurations with more storage or networking can run well over $700,000 -- $394,406 is the realistic starting point.
Memory1440 GB
8 GPUs x 180GB each, all connected inside one box with NVIDIA's NVLink so they act like one giant pool of memory instead of 8 separate ones.
Power draw14,300 W
NVIDIA's own published figure for this server class, drawn continuously while running.
Rack-scale
GB300 NVL72 Rack
An entire rack: 72 GPUs and 36 CPUs, liquid-cooled, wired as one machine.
$3,800,000
NVIDIA does not publish a list price for a fully built rack. $3.8 million is the middle of current analyst estimates ($3.7M-$4.0M for standard configurations; inference-tuned configs have been quoted as high as $6.5M). Treat this as a serious estimate, not a firm quote.
Memory20000 GB
72 GPUs wired together with NVLink into one 20 terabyte (20,000GB) memory pool -- confirmed directly on NVIDIA's own product page. This is what lets a single rack run models far larger than any individual server above.
Power draw142,000 W
NVIDIA's own published figure -- this is why racks like this need liquid cooling.