Compute
CPU and GPU compute pricing including NVIDIA T4, A100, and H100 instances
Hanzo Platform provides both CPU and GPU compute. CPU instances are available on-demand. GPU instances are available in select regions.
CPU Compute
Standard CPU compute is covered by the Cloud VM plans. All CPU instances run on AMD EPYC or Intel Xeon processors.
| Tier | vCPU Type | Use Case |
|---|---|---|
| Shared | Time-sliced vCPU | General workloads, web apps, APIs |
| Dedicated | Reserved physical cores | Databases, latency-sensitive services |
Shared instances are available on all plans (Nano through Power). Dedicated cores are available on the Power Dedicated plan ($160/mo).
GPU Instances
GPU instances are designed for machine learning inference, training, and GPU-accelerated workloads.
| GPU | VRAM | vCPU | RAM | Storage | Price |
|---|---|---|---|---|---|
| NVIDIA T4 | 16 GB | 4 | 16 GB | 200 GB NVMe | $0.50/hr |
| NVIDIA A100 40GB | 40 GB | 12 | 85 GB | 500 GB NVMe | $2.50/hr |
| NVIDIA A100 80GB | 80 GB | 12 | 170 GB | 1 TB NVMe | $3.80/hr |
| NVIDIA H100 | 80 GB | 16 | 200 GB | 2 TB NVMe | $5.50/hr |
NVIDIA T4
Entry-level GPU suitable for inference workloads and light fine-tuning. Good price-to-performance for serving models up to 7B parameters.
- FP16 performance: 65 TFLOPS
- INT8 performance: 130 TOPS
- Best for: Model inference, video encoding, small-scale training
NVIDIA A100
The workhorse for ML training and inference. Available in 40 GB and 80 GB VRAM configurations. Supports multi-instance GPU (MIG) for running multiple models on one card.
- FP16 performance: 312 TFLOPS
- TF32 performance: 156 TFLOPS
- Best for: Training up to 70B parameter models, high-throughput inference
NVIDIA H100
The latest generation GPU for frontier model training and large-scale inference. Transformer Engine with FP8 support.
- FP16 performance: 989 TFLOPS
- FP8 performance: 1,979 TFLOPS
- Best for: Training 70B+ models, multi-GPU distributed training, lowest latency inference
GPU instances are available in US-East, US-West, and EU-West regions. Availability varies by GPU type. Contact us for reserved capacity.
Multi-GPU
For distributed training, provision multiple GPU instances and connect them via high-bandwidth networking:
# Launch a 4x A100 training cluster
hanzo vm launch \
--plan gpu-a100-80 \
--count 4 \
--region us-east \
--name training-clusterMulti-GPU instances within the same region communicate over 100 Gbps InfiniBand (where available) or 25 Gbps Ethernet.
Reserved Instances
For sustained GPU workloads, reserved instances offer significant savings:
| Commitment | Discount |
|---|---|
| 1 month | 10% off |
| 3 months | 20% off |
| 12 months | 40% off |
Contact sales@hanzo.ai for reserved pricing.
Regions
| Region | CPU | T4 | A100 | H100 |
|---|---|---|---|---|
| US-East (NYC) | Yes | Yes | Yes | Yes |
| US-West (SFO) | Yes | Yes | Yes | Limited |
| EU-West (AMS) | Yes | Yes | Yes | -- |
| AP-South (SGP) | Yes | Yes | -- | -- |
GPU availability is subject to capacity. Use hanzo vm plans --gpu to check real-time availability.
How is this guide?
Last updated on