Data as of April 14, 2026 · every change kept
TPU v5p Pod
DGX GB300 NVL72DGX GB200 NVL72
Maia 100 RackDGX B200
MI325X Platform
Google Cloud TPU v6e (Trillium) Pod
Pod / clusterShippingTrillium2025
Google's latest custom AI accelerator in pod configuration. 256 TPU v6e chips connected via custom ICI (Inter-Chip Interconnect). Optimized for both training and inference of Google's Gemini models. Available on Google Cloud as Cloud TPU v6e. Significant perf/watt improvement over v5e.
256
GPUs per system
230 FP8 PFLOPS
Total HBM
8 TB
Host memory
16 TB
Interconnect
ICI 4.0
51.2 TB/s
Networking
1600 Gbps
Storage
100 TB Persistent SSD
Form factor
Multi-rack pod
Weight
n/a
Rack units
n/a
Performance
Manufacturer datasheet values · aggregate system compute
| FP4 PFLOPS | n/a |
| FP8 PFLOPS | 230 |
| FP16 PFLOPS | 115 |
| BF16 PFLOPS | 115 |
| Training effective PFLOPS | 90 |
Power and cooling
Thermal envelope · cooling requirements · efficiency
Rack power
60 kW
Per GPU
200 W
Cooling
liquid
PUE estimate
1.1
Power draw relative to tracked systems60 kW / 2500 kW max
3.83 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
n/a
Lease monthly
$280,000/mo
Cost per GPU per month
$1,094
TCO per PFLOPS per year
$14,734
PFLOPS per kW
3.83
63% above the average TCO of $9,046/PFLOPS/year across all tracked systems
Available from
GC
Google CloudKnown deployments
Disclosed in press releases, SEC filings, and conference talks
Quantity
Gemini 2.5 training
Source
Google I/O 2025Quantity
Public cloud availability
Source
Google Cloud blogSources
Every data point on this page is reproducible
Other AI systems
Compare across the system landscape
8960x Google TPU v5p · 8100 PFLOPS
Pod / clusterShipping
72x NVIDIA B300 · 1440 PFLOPS
Full rackAnnounced
72x NVIDIA B200 · 720 PFLOPS
Full rackShipping
32x Microsoft Maia 100 · 96 PFLOPS
Full rackRamping
8x NVIDIA B200 · 80 PFLOPS
Server nodeShipping
8x AMD MI325X · 48 PFLOPS
Server nodeShipping