NVIDIA Hardware

10 products. Real specs. Real prices. From a desktop card to a full datacenter rack.

Entry

NVIDIA GeForce RTX 4090

Desktop Workstation
$2,399

The most powerful desktop GPU ever made. 24 GB GDDR6X. Runs GLM 5.1 on a workstation — no datacenter required.

VRAM 24 GB GDDR6X
Power (TDP) 450 W
Form Factor PCIe 4.0 x16
NVLink No
Power equivalent: 0.4 average homes · charges an EV battery every 200.0 hrs
Best for: single developer, proof-of-concept, sub-20B models
Relative Scale (logarithmic)
VRAM
24 GB GDDR6X
Power
450 W
Price
$2,399
VRAM Power Price
Pro

NVIDIA RTX 6000 Ada

Professional Workstation
$6,800

Professional workstation card with ECC memory for error-free operation. NVLink-capable for 96 GB in pairs.

VRAM 48 GB GDDR6 ECC
Power (TDP) 300 W
Form Factor PCIe 4.0 x16
NVLink Yes
Power equivalent: 0.3 average homes · charges an EV battery every 300.0 hrs
Best for: workstation AI dev, up to 40B parameters
Relative Scale (logarithmic)
VRAM
48 GB GDDR6 ECC
Power
300 W
Price
$6,800
VRAM Power Price
Inference

NVIDIA L40S

Inference Accelerator
$22,000

Purpose-built for AI inference. Ada Lovelace architecture, optimized for serving LLMs to many users simultaneously.

VRAM 48 GB GDDR6 ECC
Power (TDP) 350 W
Form Factor PCIe 5.0 x16
NVLink No
Power equivalent: 0.3 average homes · charges an EV battery every 257.1 hrs
Best for: production inference servers, up to 40B models
Relative Scale (logarithmic)
VRAM
48 GB GDDR6 ECC
Power
350 W
Price
$22,000
VRAM Power Price
Data Center

NVIDIA H100 PCIe 80 GB

Data Center GPU
$30,000

Industry-standard AI GPU. 80 GB HBM3 at 3.35 TB/s. Runs models up to 65B parameters per card. Build multi-card systems freely.

VRAM 80 GB HBM3
Power (TDP) 350 W
Form Factor PCIe 5.0 x16
NVLink No
Power equivalent: 0.3 average homes · charges an EV battery every 257.1 hrs
Best for: training, large inference, scalable clusters
Relative Scale (logarithmic)
VRAM
80 GB HBM3
Power
350 W
Price
$30,000
VRAM Power Price
SXM High-BW

NVIDIA H100 SXM5 80 GB

Data Center GPU — High Bandwidth
$40,000

SXM5 form factor with NVLink 4.0 fabric. The foundation of DGX H100 systems. 900 GB/s GPU-to-GPU bandwidth when NVLink-connected.

VRAM 80 GB HBM3e
Power (TDP) 700 W
Form Factor SXM5 (NVSwitch fabric)
NVLink Yes
Power equivalent: 0.6 average homes · charges an EV battery every 128.6 hrs
Best for: DGX systems, tight multi-GPU coupling
Relative Scale (logarithmic)
VRAM
80 GB HBM3e
Power
700 W
Price
$40,000
VRAM Power Price
Flagship

NVIDIA H200 SXM5 141 GB

Data Center GPU — Flagship
$45,000

NVIDIA's most powerful single GPU. 141 GB HBM3e at 4.8 TB/s — nearly double H100 bandwidth. Runs models up to 115B parameters per card.

VRAM 141 GB HBM3e
Power (TDP) 700 W
Form Factor SXM5 (NVSwitch fabric)
NVLink Yes
Power equivalent: 0.6 average homes · charges an EV battery every 128.6 hrs
Best for: largest single-card inference, DGX H200 systems
Relative Scale (logarithmic)
VRAM
141 GB HBM3e
Power
700 W
Price
$45,000
VRAM Power Price
NVL Dual

NVIDIA H100 NVL

Data Center GPU — Dual NVLink Module
$65,000

Two H100 GPUs in a single PCIe card, NVLink-coupled so the full 188 GB acts as one memory pool. The smallest single unit that comfortably runs 130B-parameter models at INT8 — no DGX chassis required.

VRAM 188 GB HBM3 (2 × 94 GB, NVLink)
Power (TDP) 600 W
Form Factor PCIe 5.0 x16 dual-GPU module
NVLink Yes
Power equivalent: 0.5 average homes · charges an EV battery every 150.0 hrs
Best for: models in the 100–160 GB range; single-unit alternative to multi-card PCIe setups
Relative Scale (logarithmic)
VRAM
188 GB HBM3 (2 × 94 GB, NVLink)
Power
600 W
Price
$65,000
VRAM Power Price
DGX System

NVIDIA DGX H100

Full DGX System
$300,000

8 H100 SXM5 GPUs wired together by NVSwitch. 640 GB total. 900 GB/s all-to-all GPU bandwidth. A complete AI supercomputer in a single box.

VRAM 640 GB (8 × 80 GB HBM3e)
Power (TDP) 10,200 W
Form Factor 8× SXM5, NVSwitch
NVLink Yes
GPU Count 8 GPUs
Power equivalent: 8.5 average homes · charges an EV battery every 8.8 hrs
Best for: 100B–600B models, first serious cluster step
Relative Scale (logarithmic)
VRAM
640 GB (8 × 80 GB HBM3e)
Power
10,200 W
Price
$300,000
VRAM Power Price
DGX Flagship

NVIDIA DGX H200

Full DGX System
$450,000

8 H200 SXM5 GPUs. Over 1.1 TB of combined GPU memory — enough to run a 900B-parameter model at INT8. The most memory-dense box we sell.

VRAM 1,128 GB (8 × 141 GB HBM3e)
Power (TDP) 10,200 W
Form Factor 8× SXM5, NVSwitch
NVLink Yes
GPU Count 8 GPUs
Power equivalent: 8.5 average homes · charges an EV battery every 8.8 hrs
Best for: sub-trillion models, maximum single-node memory
Relative Scale (logarithmic)
VRAM
1,128 GB (8 × 141 GB HBM3e)
Power
10,200 W
Price
$450,000
VRAM Power Price
Cluster

NVIDIA DGX SuperPOD H200

Datacenter Cluster
$3,500,000

10 DGX H200 nodes laced together by Quantum-2 InfiniBand. 80 GPUs, 11.28 TB of unified memory, real cluster failover and scheduling.

VRAM 11,280 GB (80 × H200 141 GB)
Power (TDP) 102,000 W
Form Factor 10× DGX H200, InfiniBand fabric
NVLink Yes
GPU Count 80 GPUs
Power equivalent: 85.0 average homes · charges an EV battery every 0.9 hrs
Best for: multi-trillion parameter research, fleet serving
Relative Scale (logarithmic)
VRAM
11,280 GB (80 × H200 141 GB)
Power
102,000 W
Price
$3,500,000
VRAM Power Price
Full Rack

NVIDIA MGX Full Datacenter Rack

Full Datacenter Rack
$6,500,000

20-node DGX H200 SuperPOD with full networking fabric, NVMe storage, and cooling infrastructure. 160 GPUs, 22.56 TB GPU memory.

VRAM 22,560 GB (160 × H200 141 GB)
Power (TDP) 204,000 W
Form Factor 20× DGX H200, InfiniBand, Storage
NVLink Yes
GPU Count 160 GPUs
Power equivalent: 170.0 average homes · charges an EV battery every 0.4 hrs
Best for: hyperscale inference, top-of-the-line AI datacenter
Relative Scale (logarithmic)
VRAM
22,560 GB (160 × H200 141 GB)
Power
204,000 W
Price
$6,500,000
VRAM Power Price