Why convert at the billing line item instead of swapping in the marketplace

The marketplace "swap" route has a problem: images and SaaS subscriptions are usually billed as separate products, whether credits apply depends on that product's billing attribution, and what you get is often a software license rather than directly callable inference tokens.

A more reliable approach is to work at the billing line item level in the console. On most Chinese clouds, new-user trial credits are either general vouchers or product-restricted vouchers. As long as the target AI service's billing line item falls within the eligible scope, the usage charges after activation will be deducted from the credit first. So the core action is not "find a product" but confirm the credit's eligible product list, then align the AI service's billing line items to it.

Steps

Step 1: Find out the credit's eligible products and deduction order

  1. Log in to the console, go to Billing Center / Costs and view each voucher under Coupons / Vouchers: face value, expiry, eligible product scope, and whether it is restricted to a specific billing mode (subscription vs pay-as-you-go).
  2. Focus on two restrictions: whether it is pay-as-you-go only (AI inference is almost always pay-as-you-go, which usually helps), and whether it excludes certain high-value products (some vouchers explicitly exclude GPU instances and dedicated resource pools).
  3. Record each voucher's expiry date and sort by expiry — this determines the activation order below.

Step 2: Break the AI service into line items the credit can cover

AI capabilities on Chinese clouds usually split into several billing line items with different deduction difficulty:

  • Model API calls (per token or per call): easiest to absorb credit, a standard pay-as-you-go product.
  • Inference service instances (hourly): deductible, but check whether the "GPU instances excluded" clause applies.
  • Model training / Notebook environments: usually deductible, but high unit price, burns credit fast.
  • Storage and traffic: deductible, but a hidden drain that can eat your credit unnoticed.

Recommended conversion order: activate per-token model APIs first (slow, predictable consumption), then hourly inference instances; set separate budget alerts for storage and traffic so they don't crowd out inference quota.

Step 3: Huawei Cloud — ModelArts and Pangu/Ascend inference entry points

  1. In the ModelArts console, activate Online Services / inference deployment, choosing on-demand billing rather than subscription, so the line item falls under pay-as-you-go products.
  2. If using preset model services (such as the Pangu family or hosted LLM APIs), activate them via the ModelArts Studio / model-as-a-service entry; these are typically billed per token or per call and are the closest thing to "AI tokens".
  3. After activation, check the Billing Center: does the service's usage detail show "voucher deduction"? If no deduction appears, the line item is out of scope — stop the service immediately rather than waiting for the bill.
  4. Set auto-stop policies on online services other than on-demand-to-subscription conversion, so idle instances don't keep draining credit.

Step 4: Alibaba Cloud — Bailian (Model Studio) and PAI entry points

  1. Activate model services in the Bailian console and create an API key; billing is typically metered per token, a classic pay-as-you-go-deductible product.
  2. In PAI (Platform for AI), if you need EAS inference or DSW dev environments, also choose pay-as-you-go and confirm the instance spec does not trigger voucher exclusion clauses.
  3. Use the Billing Center's split billing to view deductions by product and confirm the credit is actually applied to model services.
  4. Alibaba Cloud often has separate "model service new-user free quota" independent of general trial credits; the two can stack: use the free quota first, then it automatically falls back to voucher deduction of pay-as-you-go bills.

Step 5: Make deduction status observable

  1. Set an alert in the Billing Center when voucher balance falls below a threshold.
  2. Set a daily cap on model API calls (at the gateway or application layer) so a runaway script doesn't burn the credit at once.
  3. Reconcile weekly: does the credit balance decline match expected call volume?

Caveats

  • Credits are not cash balance: most trial credits cannot be withdrawn, transferred, or refunded, expire when unused, and some cannot be invoiced — not a long-term budget.
  • Exclusion clauses are the biggest trap: GPU instances, dedicated resource pools, and some high-spec instances are often excluded; read the voucher terms, not the marketing page.
  • Region restrictions: some vouchers are limited to specific regions; services activated in another region may not qualify.
  • Identity and entity restrictions: new-user credits are usually bound to a verified entity; multiple sub-accounts under the same entity generally share or are restricted, so re-registering is pointless.
  • Order of free quota vs voucher: typically the product's own free quota is consumed first, then vouchers, then cash balance; but rules vary by platform — after activation, make a small test call to confirm.
  • Don't over-provision just to use the credit: idle instances bill hourly, and once credit runs out you get a cash bill.

When it fits

  • Individual developers or small teams wanting to validate a model API call chain without topping up.
  • Needing low-latency inference on domestic networks without overseas relays.
  • Already holding a Huawei Cloud or Alibaba Cloud account with unused new-user benefits, wanting to spend them on AI inference rather than generic cloud resources.

When it does not fit

  • Long-term stable large-scale inference: credits are a one-off, time-limited resource and cannot inform production capacity planning.
  • Needing specific overseas models: the model catalog on Chinese clouds differs from overseas platforms and cannot substitute for them.