At three in the morning the shop is still talking. Nobody is watching in large numbers, nothing is trending, and the presenter repeats the same eight facts about a rice cooker to whoever drifts in. That is the shift it was built for.
JD.com's Double 11 progress report, covering the evening of 12 October to 6pm on 30 October, put the numbers for its digital human (数字人) presenters at scale. The headline figures are about volume: more than 100,000 merchants served, more than nine million hours of streaming, more than RMB 1.3 billion (≈ US$183M) in sales attributed to those streams.
Divide the last two and something quieter appears. Nine million hours and RMB 1.3 billion works out to roughly RMB 144, or about US$20, of sales for every hour a synthetic presenter was on air.
Key takeaways
- JD's JoyAvatar has served more than 100,000 merchants; in the first half of Double 11 its presenters streamed over nine million hours.
- Sales attributed to those streams topped RMB 1.3 billion, about US$20 for each streaming hour.
- That is roughly 90 hours and RMB 13,000 (≈US$1,830) of sales per merchant across the 18-day window.
- In 2024, JD's earlier digital-human line reported RMB 10 billion (≈US$1.4B) of sales over 400,000 hours, or about RMB 25,000 an hour.
- The product has moved from a premium substitute for a top host to a filler for hours no human would work.
What JD actually disclosed
The figures come from JD's own promotional reporting, relayed by trade press. The company said:
- JoyAvatar, running on its JoyAI model, had passed 100,000 merchants in total;
- those merchants streamed more than nine million hours in the window;
- the streams produced more than RMB 1.3 billion in sales;
- a separate digital worker product handled more than 684 million personalised customer communications;
- a financial AI assistant grew users tenfold year on year.
Nine million hours is 1,027 years of continuous broadcast. As a supply statistic, it is extraordinary. As a demand statistic, it says nothing.
The two sets of numbers, side by side
In 2024, JD's cloud unit reported its digital humans had served more than 5,000 brands, streamed over 400,000 hours, drawn 100 million views and driven more than RMB 10 billion (≈ US$1.4B) in sales. It also said the presenter cost about one tenth of a human one.
| 2024 disclosure (Yanxi) | 2026 disclosure (JoyAvatar) | |
|---|---|---|
| Merchants or brands | 5,000+ | 100,000+ |
| Hours | 400,000+ | 9,000,000+ |
| Sales attributed | RMB 10B+ | RMB 1.3B+ |
| Sales per hour | ≈RMB 25,000 (≈US$3,500) | ≈RMB 144 (≈US$20) |
The two figures come from the same company but not the same product label, and one is cumulative while the other covers eighteen days of a promotion. Treat the comparison as directional, not exact. Even so, the direction is not subtle: sales per streaming hour fell by two orders of magnitude.
Ninety hours per shop
The averages are more useful than the totals. Spread across 100,000 merchants, nine million hours is ninety hours each, and RMB 1.3 billion is RMB 13,000 each.
Ninety hours in eighteen days is five hours a day. That is not a store replacing its star presenter. It is a store keeping a channel warm through hours when nobody would pay a human to stand there, taking whatever orders arrive.
RMB 13,000 of sales per merchant is real money for a small shop and rounding error for a brand. It is also, crucially, sales that mostly would not have existed otherwise, because the alternative was a dark screen.
Why the price per hour had to fall
A digital presenter only makes sense economically at the bottom of the market. At the top, a human with a following converts far better, and no platform claims otherwise.
So the product's natural home is the long tail: the merchant with one warehouse, the category where the script barely changes, the time slot where the audience is a trickle. Ten thousand of those merchants at a hundred hours each is a bigger number than fifty brands at a thousand hours, and a much smaller business per account.
That is the trade JD made. It gives away reach at the margin to keep merchants inside its marketplace, and books the resulting orders as platform GMV. The digital human is not the product. Traffic retention is.
The cost side nobody publishes
JD has never disclosed what a merchant pays for JoyAvatar, so the return on the tool cannot be calculated from outside. Nor does the company break out how much of the RMB 1.3 billion would have happened anyway: a store that streams to nobody still converts the occasional visitor who searched for the product directly.
There is a compliance cost too. In September, the cyberspace regulator's campaign against misuse of model tools listed, among its case types, digital humans used to make false claims while selling to elderly viewers. A synthetic presenter reading a script is a compliance surface, not just a labour saving.
Honest limitations
Every figure here is JD's own promotional reporting for a festival window, not an audited segment disclosure; the company does not report digital-human revenue or merchant retention. The 2024 comparison uses a different product name and a cumulative base, so the per-hour collapse is indicative rather than measured like-for-like. I have no data on how many of the 100,000 merchants kept streaming after the promotion ended, the number that would tell you whether this is a business or a campaign. RMB 144 is arithmetic on two disclosed totals, not a unit price, and it mixes categories with different basket sizes.
What readers can do now
- Judge a digital presenter on incremental orders in hours that would otherwise be dark, not on average conversion against a human host.
- Ask any vendor for post-promotion retention. A merchant count that spikes for eighteen days and halves afterwards is a subsidy, not adoption.
- Audit the script, not the avatar. The regulatory risk sits in the claims being read aloud, and it does not disappear because the reader is synthetic.