Self-Host an LLM on a Rented GPU or Pay per Token? A Break-Even Calculator (2026)

说明与用途

文章说明了如何计算在租用 RTX 5090 32G GPU 上运行 Qwen3 32B AWQ 模型的盈亏平衡点，并将其成本与按令牌计费的 API 进行比较。持续流量超过约 600 tokens/s 时，自托管更划算，否则更适合使用 API。

问题

Determine when self-hosting an LLM on a rented GPU is cheaper than using a per-token API.

结果

The calculation shows that for the Qwen3 32B AWQ model on an RTX 5090 32G, self-hosting becomes cheaper than per-token APIs when sustained throughput exceeds roughly 600–700 tokens/s (depending on utilisation). Below that, the API wins.

工作流

• 1. Gather GPU hourly price and per-token API prices.
• 2. Compute blended price per token using input:output ratio.
• 3. Calculate break-even tokens per hour: hourly_gpu_price / blended_price_per_token.
• 4. Adjust for GPU utilisation: divide by utilisation factor.
• 5. Compare required throughput to actual traffic to choose strategy.
• 6. Optionally run provided Python script to automate calculations.

证据

Self-hosting on a rented GPU is cheaper than a per-token API only when you keep the GPU busy: the break-even is the hourly GPU price divided by the API’s price per token, expressed as tokens you must actually generate and process every hour.

来源

break_even_tokens_per_hour = hourly_gpu_price / blended_price_per_token

来源

If your GPU is busy only 40% of the hours you rent it, the throughput you need while busy is break-even ÷ 0.4.

来源

Qwen3 32B is an open-weight model (Apache-2.0) that is not offered on the CloudGPU API, so self-hosting is the way to run it here.

来源

At a fixed $0.59 per hour, the cost per million tokens depends only on the throughput you sustain.

来源

These are different models from Qwen3 32B, so the table compares cost only: decide on quality first, then use the arithmetic.

来源