A top Chinese livestreamer sold 2 trillion AI tokens in thirty minutes during the Singles' Day run-up, a volume that makes sense only once you stop reading it as a product review and start reading it as a promotional clip with a voucher attached. As of November 9, 2026, the packages sit on dedicated shelves at Tmall and JD.com, and the sellers are discovering that a digital credit behaves like a fast-moving good in the catalogue and like a liability in the accounts.
What the clip actually sold
Two trillion units sound enormous until the unit is priced. A token is a fraction of a cent, so the headline number describes a modest basket sold to a very large audience in a very short window. What matters for the catalogue is not the volume but the packaging: the clip sold a denomination, a validity window and a promise about which models the credit can be spent on.
The mechanism is familiar from physical goods. A host holds up a price card, a countdown runs, and the audience buys before the card changes. Applied to a digital credit, the same mechanism produces a different accounting problem, because nothing ships and nothing can be returned in the ordinary sense. The seller has booked revenue for usage that may happen next week, next quarter, or never.
Voucher mechanics on a digital good
The discount is a liability, not a markdown
Cutting the price of shampoo moves units and clears stock. Cutting the price of tokens moves units and creates an obligation, because every discounted token sold is capacity the platform must serve later at full cost. A promotion that looks like a 30 percent markdown is closer to a 30 percent prepayment of future compute, and the margin only appears if the buyer actually spends the balance.
That changes how a promotional calendar should be read. For physical stock, the question is how much inventory the discount clears. For tokens, the question is how much capacity the discount commits, and whether the committed capacity was priced for this. Sellers running a month-long promotion without that distinction are effectively selling forward at a fixed price into a variable cost.
Validity windows decide the refund line
Most packages carry an expiry. An expiry is what keeps the liability finite, and it is also the source of the complaints that land in the review section: buyers who bought during a clip discover a month later that part of the balance has lapsed. The refund line then moves not with product quality but with how clearly the window was stated on the listing.
Category growth in digital credits therefore depends on listing discipline more than on discount depth. A package that states the denomination, the models it covers, the expiry and the partial-use rule in the first screen generates fewer disputes than one that leads with a percentage off. The percentage drives the click; the terms drive the repeat purchase.
Where the quota breaks
The complaints on a token listing arrive from two populations. One bought a code and could not activate it. The other used the credit and disliked the output. Averaging the two into a star rating hides both, and assigns both to the seller, although only the first belongs there.
| Complaint | Who reports it | What it signals | Where to look |
|---|---|---|---|
| Code will not activate | First-time buyers | A fulfilment defect at the store | Order-level refund notes |
| Answer quality | Heavy users | A property of the model | Long reviews with usage detail |
| Balance disappeared | Both groups | The expiry was not read | Listing screenshots by date |
| Price fell after purchase | Returning buyers | A promotional hangover | Repeat purchase rate by cohort |
Split this way, the rating becomes actionable. Activation failures belong to the store and can be fixed with a code test before the clip goes live. Balance complaints belong to the listing and can be fixed with a clearer first screen. Output complaints belong to the provider and cannot be fixed by the seller at all, which is worth stating plainly rather than absorbing into a general apology.
What to log before the next stream
The four fields below are ordered by how cheaply they can be checked. Running them after a clip is useful; running them before the next one is what changes the outcome.
- Record the denomination and the validity window for every package on the shelf, and check that the listing states both above the fold.
- Compare the promotional price with the price the same model sells for directly, and keep the gap as a field rather than a one-off note.
- Track refunds by reason at order level, separating activation failures from lapsed balances, because the two are fixed by different teams.
- Split first purchase from repeat purchase by cohort, and report them as two numbers rather than one growth figure.
| Field | Example value | Reading |
|---|---|---|
| Denomination | 10 million tokens, 30-day validity | Shorter windows lift disputes |
| Price gap versus direct | 39 yuan promotional, 59 yuan list | Above 30 percent invites resale |
| Refund rate by reason | 2.4 percent, half from lapsed balance | Points to the listing, not the model |
| Repeat purchase share | 31 percent in month two | The cleanest demand signal available |
Run the four together and the category becomes legible in a way the festival headlines are not. A clip that moves 2 trillion units is an event; a repeat purchase share in month two is a verdict. Only the second tells a merchant whether the shelf deserves more space or the package needs to be repriced.
FAQ
Is a token package really a product listing?
On a marketplace it behaves like one. It has a category node, a comparable price, a sales count and a review stream, and each of those creates the same monitoring questions a physical item creates.
Why does a discount on tokens cost more than the margin?
Because every discounted token sold is capacity the platform must deliver later at full cost. The promotion sells forward at a fixed price into a variable cost base.
Should output complaints count against the store?
No. Activation and balance problems belong to the seller; answer quality belongs to the model provider. Mixing them sends the fix to the wrong team.
Does a 2 trillion unit clip prove the category is established?
Not on its own. It measures one window with one host. Whether the category holds depends on the second and third purchase, which the current disclosures do not cover.
