Private AI cost calculator: hardware, API and operations

Estimate private AI costs from actual usage, cancelled seats, model hosting, GPU utilization, retained tools and the work needed to operate it.

Last updated

In short

  • A self-hosted chat interface can still send prompts to an external API.
  • Only subscriptions or API spend actually removed count as avoided cost.
  • Evaluate cost per accepted task, including retries and human correction.
On this page (5)

Costs and savings · Part of Private AI

Interactive worksheet · no sign-up

Estimate Private AI costs

Use the same currency and scope for both options. Inputs stay in this page; they are not included in analytics or sent with an enquiry.

1. Your current bill
2. The replacement

Business management: CHF 1,590/month for up to 3 apps; onboarding from CHF 9,900. Infrastructure and paid licences are extra.

GPU capacity reserved

Count idle reserved hours as well as useful inference. 730 hours represents one GPU reserved for an illustrative month; it is not a capacity recommendation.

Remote API input tokens

For API or hybrid inference; use your provider’s model-specific, cache-aware quote.

Remote API output tokens

Local-only inference uses zero remote tokens. Embedding and reranking costs belong in infrastructure or licence costs.

A private front end using a remote API is not local inference. Choose API usage, GPU capacity, or a declared hybrid below; check model quality, concurrency and data routing before comparing savings.

3. Transition and overlap
Optional: value internal operating time

Time released is capacity, not automatically payroll or cash saved. Keep time already billed by Pilae or another supplier out of these fields.

How the calculation works
  1. Current monthly invoices = billed units × unit price + additional current spend.
  2. Replacement monthly invoices = Pilae management + additional infrastructure + target licences + workload usage + retained services.
  3. Transition = plan onboarding + new hardware + extra migration + training + parallel months × max(0, current invoices − retained services). Parallel months are capped at the comparison horizon.
  4. Total invoices = monthly invoices × horizon, plus transition for the replacement. Payback uses the full transition commitment divided by positive monthly invoice saving, including all entered overlap months even beyond the selected horizon.
  5. Economic saving adds the change in valued internal operating time. It does not assume headcount reduction.

Business includes up to 3 apps, then a published per-app fee (CHF 390 in CHF). Starter’s published CHF price is CHF 449/month. All Pilae figures come from the same price data as the pricing page. Workload inputs are your estimates, not supplier quotes. Taxes, price growth, discount rates, opportunity costs and differences in functionality are outside this arithmetic.

Private AI can make sense because of data placement, control or a particular workload. It is not automatically cheaper than a hosted assistant. The calculator on this page compares explicit costs; it cannot establish that two models perform the same work or that a GPU can serve a particular number of employees.

Choose one of three architectures

Architecture You are paying for Boundary to confirm
Hosted assistant seats A provider-operated application, models and its included tools Which users and features can be discontinued
Operated interface plus API Interface operation, infrastructure, API requests and tools Prompts reach the selected provider
Operated interface plus local inference Application stack, model serving, GPUs, storage and operation All dependencies must follow the required placement policy

Open WebUI is an interface, not a bundled frontier model. Its licence includes branding conditions that should be checked alongside model licences and any paid components. An API-backed deployment must not be represented as local inference merely because the login page runs on your server.

Measure workload before counting seats

Collect monthly input and output tokens, peak concurrency, typical and long context sizes, response-time requirements and the proportion of tasks accepted without substantial correction. Include retries, document extraction, embeddings and tools where used. Preserve the distribution: an average of one request a second can hide a morning burst that sets GPU capacity.

The API price schedule distinguishes billable categories. A useful worksheet keeps input, output, cached input and other fees separate. With local inference, divide the complete recurring cost by accepted tasks, not by the maximum tokens a benchmark generated.

For illustration, 20,000 tasks with 1,500 input tokens and 400 output tokens each imply 30 million input and eight million output tokens before retries. Multiply each by the applicable rate. These are hypothetical workload counts, not an observed Pilae deployment.

Avoid the employee-count trap

A company with 200 employees may only discontinue 35 assistant subscriptions. The avoided seat bill is those 35 seats at their actual net price. Keep specialist subscriptions when the private system does not replace coding tools, connectors, research, audio or other required capabilities.

For a local system, a fixed GPU bill continues during quiet periods. For an API system, usage charges may fall when demand falls, while application operation still has a fixed cost. Evaluate a low-use month, normal month and peak month instead of comparing one favorable average.

Include costs outside inference

  • Application and vector storage, backups, identity and monitoring.
  • Model/runtime upgrades, evaluation and incident response.
  • Setup, document preparation, access mapping and user training.
  • Paid software or model rights where applicable.
  • Retained SaaS, API fallbacks and human review.

The cost-optimization guidance describes request reduction, smaller suitable models and asynchronous processing. Treat each as a candidate to evaluate, not a saving to assume. The vLLM benchmark interface provides measurement controls, but a benchmark should use your task mix and latency target.

A useful acceptance gate

Before buying capacity, freeze a representative task set and judge answers against an agreed rubric. Record failed tasks, latency percentiles and the rate at which a human must redo the work. Run the financial model with those results and actual hardware quotes. Do not compare a best-case local throughput figure against a full-feature SaaS seat as if they were interchangeable.

Use current Pilae pricing for the operation plan and a separate infrastructure quote. First-year cash includes onboarding and transition; a three-year amortized view should remain separate. If the outcome costs more, the deployment decision may still be justified by control or placement requirements. State that benefit directly rather than inventing a financial saving.

Sources and further reading

Use the linked documentation to confirm the version, edition and contract that apply to your project.

Scope a private AI budget

Bring your current bill, required capabilities and workload assumptions. We will separate recurring operation from migration and retained services.