Documentation

Costs

ConfigIQ compares three ways to serve the same open-weight model and workload:

  1. a hosted token API, billed per token;
  2. rented GPU infrastructure, billed per instance-hour; and
  3. purchased GPU servers, amortised over a hardware life.

The ordering is an output of the model, not an assumption. Nothing in the calculation forces hosted, rented, or purchased infrastructure to win.

What the number includes

Lens Includes Excludes
Hosted Input and output token prices for the selected model and feed Egress, storage, rate-limit overage, managed-service fees
Rented Complete instance price for the sized topology Egress, storage, support uplift, staff
Purchased Hardware amortisation, power and cooling (PUE), maintenance, and the planning overheads you enable Networking, datacentre space, staff, unless you add them

Cost is calculated only for whole replicas, instances, or servers. A partial GPU is not a purchasable unit, so the result rounds up and states the headroom that rounding created.

Planning assumptions

Assumptions are edited on the Cost sources page under cost assumptions. They are stored in this browser, not on the server, and they apply to this user's calculations only. Defaults are defined in lib/costing-assumptions.ts.

The assumptions that move the result most are hardware life, electricity price and PUE, annual maintenance, loaded staff cost, and the utilisation you assume. Changing them changes the comparison; they are not hidden constants.

Pricing data

Hosted model prices and shared GPU prices come from the aicostings service, through selectable pricing feeds. The selected feed is named on the result, and the Cost sources page records where each figure came from.

Pricing is cost data. It never changes a sized GPU count: the topology comes from AISimulators, and the price is applied to it.

What this is not

Hybrid Savings is an industry-aligned planning model, not a quote. A configuration marked Tested elsewhere in ConfigIQ has real-hardware benchmark evidence; a sizing estimate is still an estimate. Before a purchase or a production commitment, confirm current commercial rates and run a workload-matched benchmark.