An overseas brand asks a production partner near its China supply chain to support regular launches. The partner proposes dedicated AI capacity to reduce waiting and improve control. In a hypothetical month, product information stays in approval while the compute bill continues to grow.
The purchase may have reserved running capacity rather than bought individual completed assets. ONCE recommends keeping two ledgers: instance uptime and accepted production output. A busy machine and a useful content service are different results.
Read the billing scope before comparing prices
Replicate’s official Billing documentation, checked November 8, 2026, says public-model setup and idle time are free, with active request processing charged. Deployments are charged for instance uptime, including setup, idle and active time. Most private models also use dedicated hardware billing, with documented exceptions such as fast-booting fine-tunes. These statements apply to that service, not every AI platform.
The commercial implication is that no submitted work need not mean no compute expense. Dedicated deployment, private model and usage-based access are not interchangeable purchasing labels. Nor does “dedicated” establish confidentiality, ownership of a model or acceptable output quality; verify each separately.
Keep a capacity ledger and a delivery ledger
The capacity ledger records each instance’s setup, active and idle uptime against its actual billing rules. Add concurrent instances separately: eight elapsed project hours are not necessarily eight instance-hours. Record actual scaling rather than treating the maximum configured capacity as usage.
The delivery ledger follows generated results through review, correction and acceptance. Link each asset to its product-information version, language, format and shot task. A successful API response does not make the content usable. A machine can be fully occupied producing results that fail the brief.
| Pattern | First investigation | Procurement implication |
|---|---|---|
| Much idle time, little output | Approval timing, operating windows and reservation arrangements | Review the capacity schedule before expanding it |
| Much processing, few accepted assets | Input quality, task suitability and rejection reasons | Fix usability before celebrating high utilisation |
| Steady work and accepted output | Peak demand, waiting cost, staffing and complete expenses | A dedicated-versus-usage comparison has a firmer basis |
This is ONCE’s diagnostic method, not a provider reporting definition or a client test result. Separating the ledgers makes it easier to locate a machine, information or approval bottleneck.
Price the accepted output over a defined period
For comparable tasks and acceptance standards, usable-asset unit cost equals the period’s in-scope compute, storage, orchestration, operations, review and correction expenses divided by accepted usable assets. Count each expense once. If a service fee already includes review, do not add that review again. Group materially different tasks instead of averaging simple backgrounds with complex functional demonstrations.
For time-priced instances, reconcile applicable rates with actual billable durations, listing changing rates, minimum charges and extras under the real terms. For request- or output-priced access, use the actual bill. Compare complete scopes on both sides; a managed-service quote should not be ranked against GPU expense alone.
Illustrative arithmetic only: an in-scope period cost of RMB800 divided by 40 comparable accepted assets is RMB20 per asset. The same cost with ten accepted assets is RMB80. These figures are not market prices or ONCE costs, and increasing generation volume does not guarantee more accepted output.
Buy standby capacity only for a defined production need
Dedicated capacity deserves evaluation when inputs are stable, work is continuous, model suitability has been established and waiting threatens agreed delivery. Ask the supplier which controllable stages the arrangement improves and on what evidence. Keeping an instance available does not guarantee immediate completion of every model call, video or approval.
If approvals arrive in one monthly burst, an always-available arrangement may spend much of the month waiting for people. Compare agreed production windows, usage-based service or a smaller dedicated allocation first. These are options to investigate, not claims that every provider offers identical controls or pricing.
For overseas teams coordinating with China, align product sign-off with production windows. A machine available when the creative team is offline cannot itself make the director’s selection or the brand’s approval. Compute capacity is only one dependency.
Define who controls the capacity
The managed-service proposal should identify who may start, retain and stop instances; which windows are reserved; who reviews low demand; who approves expansion; how shared expenses are allocated; and how usable output is accepted. A monthly fee may include capacity or charge it separately, but a theoretical maximum asset count does not define the service.
Start with one real production cycle. Retain uptime, accepted output, complete costs and waiting records. Check information and staffing constraints before changing the compute allocation. Without a baseline, do not forecast a savings percentage; without stable demand, do not turn projected full utilisation into an annual commitment.
Our AIGC production articles provide context for related sourcing decisions. Batch discounts and dedicated capacity answer different procurement questions.
For ONCE, a reusable relationship between task type, uptime, review effort and accepted output could improve managed-service estimates. That is a development opportunity, not a claim of proven margin improvement. The buyer’s next step is to ask what “always ready for you” means in measurable capacity and actual usable delivery.
