Atom XII® · Private AI Cost Calculator
Cloud AI API or your own GPUs: which costs less for your workload?
Enter your monthly volume and prices. The calculator compares your API bill with renting GPUs for an open model and shows the volume where they cost the same. The defaults are published prices and a published benchmark, listed below with their sources. Nothing you type leaves this page.
Result
Cloud API: $4,500 a month (blended $450 per 100M tokens).
Self-hosted: $14,279 a month for 1 deployment of 4 GPUs (each handles about 5B tokens a month at your utilization).
Break-even: about 3.2B tokens a month on one deployment.
The API costs about $9,779 less a month.
Formulas: API = tokens × (input price × input share + output price × output share). Capacity = tokens/s × 3,600 × hours × utilization. Self-hosted = deployments needed × GPUs × GPU price × hours + other costs. Break-even = one deployment's monthly cost ÷ blended API price.
What the defaults are, and where they come from
- Cloud API price: $2.00 per 1M input tokens, $12.00 per 1M output tokens (Gemini 3.1 Pro Preview, standard paid tier, prompts up to 200k tokens). Source: Google AI Studio, Gemini API pricing.
- GPU rental price: $4.89 per GPU-hour (NVIDIA H100 SXM, 1x instance, plus applicable tax; prices effective Nov 16, 2026). Source: Lambda, GPU cloud pricing.
- Self-hosted throughput: 3,795 tokens/s for Llama 3.1 70B Instruct on 4 x H100 80GB, fp8, at 100 concurrent requests of 1,000 input / 1,000 output tokens (NIM 1.3.0). Source: NVIDIA NIM for LLMs benchmarking, performance.
What this leaves out: an open model is not the same model as the API you use today, so compare models of the quality your task needs. Throughput depends on your prompts, output length and concurrency; the benchmark is one published configuration, not a measurement of your workload. Owned hardware, reserved pricing, power, staff and egress change the answer; enter them under other costs. Spiky or small workloads often stay cheaper on an API.
Want the numbers checked for your workload?
Send us what you run today and we reply with a free one-page readout: the same workload on rented or owned GPUs at published prices, what would have to be true for self-hosting to pay off, and what we would check first.
Get the free readout