NVIDIA HGX H100 8-GPU
Server nodeShippingHopper2023
The system that launched the AI infrastructure boom. 8x H100 SXM GPUs connected via NVLink 4.0. Still the most widely deployed AI training system globally. Air-cooled, fits standard datacenter racks. The reference design that Dell, HPE, Supermicro, and Lenovo all build around.
8
GPUs per system
32 FP8 PFLOPS
Total HBM
0.64 TB
Host memory
2 TB
Interconnect
NVLink 4.0
7.2 TB/s
Networking
400 Gbps
Storage
30 TB NVMe SSD
Form factor
8U node
Weight
140 kg
Rack units
8U
Performance
Manufacturer datasheet values · aggregate system compute
| FP4 PFLOPS | TBD |
| FP8 PFLOPS | 32 |
| FP16 PFLOPS | 16 |
| BF16 PFLOPS | 16 |
| Training effective PFLOPS | 12 |
Power and cooling
Thermal envelope · cooling requirements · efficiency
Rack power
10.2 kW
Per GPU
700 W
Cooling
air
PUE estimate
1.3
Power draw relative to tracked systems10.2 kW / 2500 kW max
3.14 FP8 PFLOPS per kW · average across all systems is 4.81
TCO analysis
Hardware amortized over 3 years · power at $0.05/kWh
List price
$300,000
Per GPU effective
$37,500
Cost per GPU per month
$1,042
TCO per PFLOPS per year
$3,306
PFLOPS per kW
3.14
63% below the average TCO of $9,046/PFLOPS/year across all tracked systems
Available from
NVIDIAD
DellH
HPES
SupermicroL
LenovoKnown deployments
Disclosed in press releases, SEC filings, and conference talks
Quantity
600,000+ H100 GPUs
Source
Meta Q4 2025 earningsQuantity
hundreds of thousands
Source
Microsoft FY2025 10-KQuantity
A3 instances
Source
Google Cloud blogSources
Every data point on this page is reproducible
Other AI systems
Compare across the system landscape
TPU v5p Pod8960x Google TPU v5p · 8100 PFLOPS
Pod / clusterShipping
DGX GB300 NVL7272x NVIDIA B300 · 1440 PFLOPS
Full rackAnnounced
DGX GB200 NVL7272x NVIDIA B200 · 720 PFLOPS
Full rackShipping
TPU v6e Pod256x Google TPU v6e · 230 PFLOPS
Pod / clusterShipping
Maia 100 Rack32x Microsoft Maia 100 · 96 PFLOPS
Full rackRamping
DGX B2008x NVIDIA B200 · 80 PFLOPS
Server nodeShipping