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

Описание и применение

В статье описывается, как вычислить точку безубыточности при запуске модели Qwen3 32B AWQ на арендованном GPU RTX 5090 32G, сравнивая её стоимость с оплатой за токены в API. При устойчивом трафике свыше ~600 токенов/с самохостинг становится выгоднее, иначе лучше использовать API.

Задача

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

Результат

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.

Рабочий процесс

• 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.

Доказательства

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.

Источник

break_even_tokens_per_hour = hourly_gpu_price / blended_price_per_token

Источник

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

Источник

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.

Источник

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

Источник

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

Источник