Source: Vertex AI — Provisioned Throughput (Overview and GSU Sizing)
Vertex AI's fixed-term reserved capacity, measured in generative AI scale units (GSUs) with per-model burndown rates.
Contents
- What: vendor documentation
- Author: Google Cloud
- Date: undated, read 2026-11-14
- Original: Provisioned Throughput overview
- Original: Calculate Provisioned Throughput requirements
Summary
The overview positions Provisioned Throughput (PT) as a fixed-cost, fixed-term subscription. The sizing page explains GSUs and burndown arithmetic.
Key claims
- PT reserves throughput for a model in chosen locations at a fixed weekly or monthly price; aimed at real-time production and predictable cost.
- A GSU is a unit of throughput. Inputs and outputs are converted to input-token-equivalents per second via burndown rates.
- Gemini 3.6 Flash example: 675 tokens/s per GSU, minimum purchase 1 GSU; input tokens 1×, cached input 0.1×, output 5×.
- Required GSUs = (input + output-equivalent tokens per query × QPS) / per-GSU throughput, rounded up (worked example: 30,000 tokens/s → 45 GSUs).
- PT requests are served ahead of on-demand; unused throughput doesn’t roll over.