Five real products, spanning desktop to full datacenter rack. Prices and specs are current NVIDIA/market figures as of August 2026 — datacenter gear moves in price with supply, so treat these as representative, not a live quote.
Price
What you pay once, up front, to own the hardware. Datacenter gear is usually sold to companies, not individuals — street prices move with supply and demand.
GPU memory (GB)
GPU memory (VRAM) is the model's desk space. The entire model has to fit on it at once, or the hardware simply cannot run it — no amount of raw speed fixes too little memory.
Power (watts)
Watts drawn every second the hardware is switched on and working. This is what shows up on the electric bill, and at rack scale it decides whether a building even has enough power coming in.
An average US home draws about 1200W continuously, and a typical electric-car
battery holds about 90 kWh — we use both below to make watts feel real.
Desktop GPU
NVIDIA GeForce RTX 5090
$1,999
The card a serious hobbyist or small dev team buys off the shelf. Great for trying a model out; too small in memory for the big open leaderboard models.
GPU memory32 GB
Power draw575 W
That's as much continuous power as
0.5 average homes.
Other specs
21,760 CUDA cores · 82.6 TFLOPS FP32 · GDDR7
Workstation GPU
NVIDIA RTX PRO 6000 Blackwell
$8,565
A professional card built for AI work. Several of these bolted into one PC is the cheapest way to get into triple-digit GPU memory — with real trade-offs (see Clusters below).
GPU memory96 GB
Power draw600 W
That's as much continuous power as
0.5 average homes.
Other specs
24,064 CUDA cores · ECC GDDR7 · workstation/server form factor
Datacenter GPU
NVIDIA H200 SXM
$35,000
The chip datacenters actually run big models on. Sold as bare silicon that must live inside an NVLink-capable server — you don't plug this into a desktop.
GPU memory141 GB
Power draw700 W
That's as much continuous power as
0.6 average homes.
Other specs
4.8 TB/s memory bandwidth · 1,979 TFLOPS FP8 · NVLink/NVSwitch capable
Datacenter Server
8× H200 Server (HGX H200 node)
$370,000
A single machine, one memory pool. This is real clustering: the 8 GPUs act as one giant GPU, not eight separate ones.
GPU memory1,128 GB
Power draw10,200 W
That's as much continuous power as
8 average homes.
Other specs
8 × H200 fused by NVSwitch into ONE 1,128 GB memory pool · dual Xeon Platinum host CPUs
Datacenter Rack
NVIDIA GB300 NVL72 Rack
$4,000,000
The largest single coherent machine NVIDIA sells: a whole rack wired as one computer. This is what it takes to serve the very largest open models.
GPU memory20,736 GB
Power draw135,000 W
That's as much continuous power as
112 average homes.
Other specs
72 Blackwell Ultra GPUs + 36 Grace CPUs · single NVLink domain · ~20.7 TB unified GPU memory