OpenRouter finds rented GPUs can lose to AI APIs even at full load

Renting a GPU does not guarantee a lower AI bill—even if it runs nonstop. OpenRouter’s newly published cost analysis found that the tested DeepSeek V4.1 Flash configurations on H200 chips could not beat the cheapest eligible API service. A smaller Qwen model had a clearer path to savings, but with a response-speed trade-off.

The same calculation, two different answers

The analysis uses API prices checked on October 8, 2026. Its basic comparison is straightforward: divide hourly GPU rent by what an API would charge for the same hour’s token output and input. The result is the share of rented time the GPU must stay busy before self-hosting becomes cheaper.

For Qwen3.8 27B on a $0.50-an-hour RTX 3090, OpenRouter estimated break-even at 8% utilization for short prompts against the cheapest zero-data-retention endpoint. An endpoint is a provider’s access point for a model; zero data retention means a policy against storing prompts and responses. Using OpenRouter’s longer chat mix raised the estimate to 12%.

DeepSeek’s larger model produced the opposite result. Four-H200 configurations cost more than the cheapest qualifying API even at full utilization. Two B200 chips offered a possible saving, but only at 76% utilization at the lowest listed rent of $3.70 per GPU-hour.

A cheaper endpoint can erase the advantage

The study itself shows how quickly those thresholds age. On October 19, a new Qwen endpoint offered output at $0.50 per million tokens. OpenRouter calculated that this moved the short-prompt self-hosting threshold to about 29%—above the maximum 24% utilization possible for traffic confined to a 40-hour working week.

The 3090’s apparent bargain also depends on batching 64 requests together, yielding about 15 tokens per second for each user. Its benchmark used a heavily modified serving setup and short prompts, which OpenRouter calls a best case. By comparison, the median Qwen API endpoint delivered 45 tokens per second during its measurement window.

Token accounting creates another trap. OpenRouter notes that a published DeepSeek H200 estimate of $0.012 per million tokens mostly counts input tokens, many reused from a cache. Counting output instead, the same benchmark at $2.90 retail hourly rent costs $3.49 per million output tokens before input costs. The cheapest qualifying API charged $0.18.

Privacy controls are not hardware control

OpenRouter also presents its own data controls as alternatives to self-hosting, giving the analysis a commercial dimension. It describes two distinct filters:

• Retention: its zero-data-retention setting restricts requests to qualifying providers. Allowing Safety Retention endpoints can permit storage for abuse detection; implicit caching may temporarily retain derived cache data.
• Location: Business and Enterprise customers can use EU or US domains that restrict processing to endpoints in that region. Requests fail rather than leave the selected region.

Those restrictions can narrow model choice. In OpenRouter’s October 8 test, DeepSeek V4.1 Flash had no eligible EU endpoint, so the request failed. And no-retention policies still leave processing on a provider’s GPU. OpenRouter acknowledges that if data must never leave hardware a customer operates, self-hosting remains necessary regardless of the price comparison.