Hardware catalog
Real NVIDIA hardware, from a single desktop card up to a full datacenter rack. Every number has a plain-English explanation next to it.
NVIDIA GeForce RTX 4090
The card hobbyists and small teams actually own. Great for learning, prototyping, and running smaller open models entirely on one desk — but its 24 GB of memory is nowhere near enough for the largest open models on its own.
- GPU memory: 24 GB total
- Power draw: 450 W (~0.4 homes)
- Price: $1,999
NVIDIA RTX A6000
An older-generation (Ampere) professional card that still sells briskly because 48 GB on a single card is a sweet spot for mid-size open models and doesn't need datacenter power or cooling.
- GPU memory: 48 GB total
- Power draw: 300 W (~0.2 homes)
- Price: $4,500
NVIDIA RTX 6000 Ada Generation
The current-generation workstation flagship. Same 48 GB as the A6000 but newer silicon and faster memory bandwidth — the card a well-funded small team buys instead of building a server rack.
- GPU memory: 48 GB total
- Power draw: 300 W (~0.2 homes)
- Price: $6,800
NVIDIA H100 SXM Server (8-GPU)
The workhorse of the last two years of AI infrastructure: eight H100 GPUs wired together with NVLink so they act like one 640 GB memory pool. This is the entry ticket to running genuinely large open models.
- GPU memory: 640 GB total
- Power draw: 10200 W (~8.5 homes)
- Price: $300,000
NVIDIA H200 SXM Server (8-GPU)
Same 8-GPU server design as the H100 box, but each GPU carries 141 GB of newer, faster memory instead of 80 GB — almost double the pool for the same power envelope, which is why it costs more per server but less per gigabyte of memory.
- GPU memory: 1128 GB total
- Power draw: 10200 W (~8.5 homes)
- Price: $380,000
NVIDIA DGX B200 (8-GPU)
NVIDIA's newest Blackwell-generation factory-built server. 192 GB per GPU across eight GPUs gives 1.5 TB of pooled memory in one box — enough to run most open models available today on a single machine.
- GPU memory: 1536 GB total
- Power draw: 14300 W (~12 homes)
- Price: $515,000
NVIDIA DGX H100 SuperPOD (4-node cluster)
Four DGX H100 servers linked by high-speed networking and managed as one cluster. This is the shape a mid-size company buys to serve real traffic with headroom to grow, instead of one box running flat out.
- GPU memory: 2560 GB total
- Power draw: 44000 W (~37 homes)
- Price: $1,400,000
NVIDIA GB200 NVL72 Rack
A single liquid-cooled rack wiring 72 Blackwell GPUs into one machine with nearly 14 TB of pooled memory. This is what frontier AI labs buy dozens of at a time — one rack alone can serve enormous traffic for the largest open models.
- GPU memory: 13824 GB total
- Power draw: 120000 W (~100 homes)
- Price: $3,000,000