AI Inference Capacity & Unit Economics

Plan a self-hosted LLM fleet across VRAM admission, concurrent requests, token capacity, GPU spend, cost per million output tokens and margin-aware minimum price.

How this calculator works

Enter gpu count, vram per gpu (gib), model weights per gpu (gib), runtime reserve per gpu (gib), kv cache per active request (gib), generated tokens/s per gpu, scheduled utilization (%), fleet availability (%), average output tokens per request, gpu cost per hour ($), target gross margin (%). Select Calculate to apply the displayed formula and review the labeled results.

Formula / method

Fleet concurrency = GPUs × floor((VRAM − model weights − runtime reserve) ÷ KV cache/request); monthly tokens = GPUs × tokens/s × utilization × availability × 730 × 3,600; cost/M = monthly GPU cost ÷ monthly tokens × 1,000,000

Worked example

Eight 80 GiB GPUs running 70 GiB model weights, 6 GiB runtime reserve and 2.5 GiB KV cache per request admit 8 concurrent requests; at 120 tokens/s per GPU, 70% utilization and 99.5% availability, they produce roughly 1.76B output tokens each month.

Assumptions and limitations

Weights are replicated on each GPU, KV cache is evenly distributed, throughput is generated-output throughput, and utilization represents productive scheduled time.

Tensor/pipeline parallelism, prefills, batching, queueing, variable sequence lengths, interconnect, CPU/network costs, orchestration, depreciation, taxes and provider discounts are not modeled. Treat capacity as a planning estimate, not a serving SLO.