AWS Trainium2 UltraServer
Server nodeShippingTrainium 22025
AWS custom silicon training server. 16x Trainium2 chips in a single UltraServer node connected via NeuronLink. Designed to compete with NVIDIA DGX for training Amazon's own foundation models and offered to AWS customers as Trn2 instances. Manufactured at TSMC on N4 with HBM3.
16
GPUs per system
48 FP8 PFLOPS
Total HBM
1.5 TB
Host memory
2 TB
Interconnect
NeuronLink
6.4 TB/s
Networking
1600 Gbps
Storage
20 TB NVMe SSD
Form factor
8U node
Weight
130 kg
Rack units
8U
Performance
Manufacturer datasheet values · aggregate system compute
| FP4 PFLOPS | TBD |
| FP8 PFLOPS | 48 |
| FP16 PFLOPS | 24 |
| BF16 PFLOPS | 24 |
| Training effective PFLOPS | 18 |
Power and cooling
Thermal envelope · cooling requirements · efficiency
Rack power
12 kW
Per GPU
500 W
Cooling
air
PUE estimate
1.2
Power draw relative to tracked systems12 kW / 2500 kW max
4 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
TBD
Lease monthly
$120,000/mo
Cost per GPU per month
$7,500
TCO per PFLOPS per year
$30,131
PFLOPS per kW
4
233% above the average TCO of $9,046/PFLOPS/year across all tracked systems
Available from
AWSKnown deployments
Disclosed in press releases, SEC filings, and conference talks
Quantity
Nova model training
Source
AWS re:Invent 2025
Quantity
Claude training (AWS)
Source
Anthropic AWS partnershipSources
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