What happens to a marketplace when the product has no shelf life, no weight and no return address? As of November 11, 2026, the answer was roughly 100,000 monthly AI memberships sold at 0.01 yuan each within 30 minutes, a package representing about two trillion tokens, according to reporting on this year's Singles' Day promotions. Compute has become a consumer SKU, and it is being priced like a coupon rather than like software. For brands and platform operators, the question is no longer whether AI can be sold in a shopping cart, but what category growth means when the cheapest item on the page is metered in tokens.
A product with an expiry date
Prepaid allowances behave differently from physical goods in three ways that matter to a marketplace. They occupy no warehouse, they cannot be inspected on arrival, and their value decays from the moment of purchase. A platform that lists them beside phone cases inherits a category whose disputes are about time rather than about condition, and most dispute systems were built for the second problem.
The pricing tells you what the seller is actually buying. A 0.01 yuan membership is not a revenue line, it is a customer acquisition cost booked as a discount. The logic that made coupon budgets work in grocery, pay to move a first order and recover on repeat, is now being applied to a product whose repeat rate depends on how much the buyer uses it rather than on how much it cost.
Why a prepaid allowance breaks the returns model
The refund question
Returns are the first casualty. A physical return is triggered by damage or mismatch, and the platform can inspect the item. A token package is consumed by use, so a refund request after partial use has no clean reference point: the seller has already paid for inference and the buyer has already received value. Marketplaces that publish a token-denominated cut-off remove most of the ambiguity before it reaches a support queue. The listing states how many tokens are refundable and until when, the checkout screen shows the same number, and the order page repeats it, which costs almost nothing and ends the argument about what the buyer should have known.
The utilisation question
The second question is usage, and it is harder to answer with a promotion. A package bought during a campaign and barely used renews at a low rate, and discounting the renewal does not fix it, because the problem is fit rather than price. The measurable variable is utilisation inside the validity window, and it is the only number that predicts whether the category survives its first year. Teams that track it usually discover the allowance is sized for an enthusiast and sold to a casual buyer.
Three ways compute reaches a shopping cart
Three forms of compute are now sold through retail channels, and they behave differently enough that pricing them with one template produces the wrong incentives. The table compares them on the dimensions that decide a repeat purchase, not on the headline price that decides the first one.
| Form | Where it sits | Expiry | Refund risk | What the buyer compares |
|---|---|---|---|---|
| Monthly membership | Subscription tab | Renews monthly | Low, cancellable | Monthly price and model access |
| Prepaid token pack | Product listing | Fixed window | High, consumed by use | Tokens per yuan |
| Bundled editing credits | Inside a seller tool | Tied to the tool | Medium, hard to separate | Cost per finished image |
The middle row carries most of the volume and the least transparent pricing. A bundle sourced from three model providers has no single reference price, so the buyer compares the headline number and ignores the composition. That works while the category is new and stops working as soon as two packages at the same price deliver visibly different quality, which is the point where a marketplace either publishes the composition or starts losing trust on the listing.
The bottom row explains why this matters beyond the AI category. Editing credits bundled into a seller tool convert compute into a unit of production, and the seller stops comparing models and starts comparing output per yuan. Reporting an image edit at roughly 0.28 yuan sets a reference point that a human retoucher cannot match on time, and once a production cost is expressed that way it tends to migrate into every listing description that touches creative work.
A weekly check on a metered listing
None of this requires a new system. It requires treating a token package as a product with an expiry date and checking it on a fixed cadence, the way a grocery team checks on-shelf availability. The sequence below is deliberately small enough to run every week without a project plan.
The record is small on purpose: five listings, one screen, ten minutes a week. What it produces is evidence of whether the category is being sold on price alone, which is the question every new SKU eventually faces. Category growth in compute retail will be settled by whether the second purchase happens without a discount, and that is observable from outside the platform long before it reaches a quarterly disclosure.
FAQ
Is a token package a subscription?
Usually not. Most packages are prepaid allowances with an expiry date, which makes them closer to a gift card than to a monthly plan. The distinction matters for refunds, for accounting and for how the buyer perceives value.
Why would anyone buy compute on a marketplace?
Because the marketplace solves discovery and payment for a buyer who is already there. The same reason people buy software licences next to phone cases: the friction of a separate account and a separate checkout is higher than the price difference.
Does this category need its own return policy?
Yes. Applying the physical-goods policy produces disputes the platform cannot adjudicate, because there is no item to inspect. A separate rule, stated on the listing, is cheaper than case-by-case resolution.
What is the first metric to watch after launch?
Utilisation within the validity window. A package that is bought and barely used renews at a low rate, and no amount of promotional pricing fixes that.
