Hyperscaler
Google Cloud
The only place to rent TPUs, plus a full NVIDIA lineup on A3/A4 instances.
GCP's differentiator is TPU pods, v5e for cost, v6e (Trillium) for performance, with pricing per chip-hour that undercuts equivalent NVIDIA FLOPS when your stack speaks JAX/XLA. NVIDIA capacity (H100 A3, B200 A4) is competitive with other hyperscalers. DWS (Dynamic Workload Scheduler) eases short-term capacity.
GPU pricing on GCP
On-demand list per GPU-hour, last reviewed 2026-09. Methodology.
| GPU | $/GPU-hr | Notes |
|---|---|---|
| T4 | $0.35 | n1 + T4 |
| L4 | $0.71 | g2-standard-4 |
| TPU v5e | $1.20 | per chip-hour, us-west4 |
| TPU v6e | $2.70 | per chip-hour, us-east5 |
| A100 40GB | $2.93 | a2-highgpu ÷ 8 |
| A100 80GB | $3.93 | a2-ultragpu ÷ 8 |
| H100 SXM | $5.90 | A3 ÷ 8 |
| B200 | $8.20 | A4 ÷ 8 |
Strengths
- TPUs, exclusive
- Strong Kubernetes (GKE) GPU story
- Committed-use discounts apply to GPUs
Watch for
- TPU lock-in is real (XLA)
- Regional GPU availability varies widely