Why Livestreamed AI Tokens Are Suddenly Everywhere

Why it is suddenly everywhere

1. 01

   On November 2, nearly 100,000 Qwen Office monthly memberships sold in about 30 minutes during Li Jiaqi’s Double 11 livestream. Reports converted the attached entitlements into roughly 2 trillion tokens.

2. 02

   Eastday’s report draws the crucial line: that figure is an entitlement total, not the amount of compute viewers consumed in half an hour. The membership advertised unlimited standard-mode use, so “sold tokens” is retail shorthand rather than a meter reading.

3. 03

   China’s National Data Administration said daily token calls passed 140 trillion in March 2026, up from 100 billion at the start of 2024 and 100 trillion at the end of 2025.

4. 04

   The China Academy of Information and Communications Technology tracks success rate, output tokens per second, time to first token, stability, price and context length in its 2026 token-service monitoring. A token is a service unit, not a coin or a fixed slice of GPU time.

The timeline

1. 2016-08Subwords become practical

   Rico Sennrich, Barry Haddow and Alexandra Birch publish their ACL paper on byte-pair encoding for neural machine translation. Splitting rare words into reusable subword units helps models handle vocabulary they have never seen.

2. 2018-11The lipstick comparison

   Li Jiaqi sells 15,000 lipsticks in 15 minutes during a Double 11 livestream contest. The episode later becomes the reference point for his first Token sale: the presenter is familiar, but the goods are no longer physical.

3. 2020-06Model capability becomes an API

   OpenAI launches a general-purpose text-in, text-out API powered by GPT-3-family models. Developers can embed model calls inside other products, creating a natural role for usage-based accounting.

4. 2024–2026Token calls scale up

   Official Chinese figures put daily calls at 100 billion in early 2024, 100 trillion at the end of 2025 and more than 140 trillion by March 2026.

5. 2026-09-03Tokens get a retail shelf

   Tmall opens its AI Space Station token top-up center. Subscription cards and recharge packages from several model providers appear in one storefront, before the livestream headline arrives.

6. 2026-10Token service gets its own infrastructure reports

   CAICT publishes reports on token-oriented computing infrastructure and service performance, covering throughput, latency, reliability and price rather than model quality alone.

7. 2026-11-02The livestream turns it into a mass-market story

   Qwen Office CEO Chen Yusen joins Li Jiaqi. Nearly 100,000 monthly memberships sell in 30 minutes, producing the roughly 2-trillion-token entitlement headline.

What actually changed hands

The livestream did not ship a hard drive containing two trillion units of anything. Viewers ordered Qwen Office memberships. The service demonstrated in the broadcast could help organize schoolwork, draft documents, process data and make presentations; the token number described an allowance attached to that membership.

This distinction resolves the strangest part of the headline. A token here is a fragment of information that a language model reads or generates. It cannot be transferred to a crypto wallet, and it does not give the holder a tradable asset. Nor did the audience collectively burn through two trillion tokens before the broadcast ended. Eastday explicitly described the number as an entitlement total rather than actual compute consumption.

There were also two different large numbers in the coverage. The roughly 2 trillion tokens referred to the memberships sold. A separate 20 trillion tokens were advertised as livestream benefits. Combining them would confuse sales with giveaways. Even the sales figure is best read as a measure of retail reach: nearly 100,000 orders, with one per ID according to Beijing Business Today.

How a language fragment became a billing unit

First, machines needed manageable pieces of text

Language models do not naturally see a page as a sequence of familiar words. A tokenizer breaks input into units drawn from a learned vocabulary. A common English word may fit in one token; an unusual word may become several. Chinese text, punctuation, spaces and code follow their own patterns, and different tokenizers can count the same sentence differently.

The 2016 ACL paper is an important stop in this history because it made byte-pair encoding a practical answer to rare and unseen words in neural machine translation. Instead of treating every possible word as a separate vocabulary entry, the system could assemble words from smaller pieces. That approach later became common across large language models. Tokens began as an engineering compromise, not a consumer unit.

Then the model moved behind an API

OpenAI’s June 2020 API announcement shows the next step. Its general text-in, text-out interface let developers build model calls into applications without running the underlying distributed systems themselves. Once many different products shared a remote model, providers needed a consistent way to describe activity and account for usage. Input and output tokens offered a convenient meter.

The meter is useful, but incomplete. Two requests with the same token count can differ in model, context length, latency and serving cost. CAICT’s monitoring therefore uses output tokens per second and time to first token alongside success rate and stability. Its separate infrastructure report describes token-oriented computing as a stack involving chips, software, clusters and scheduling. A raw token total says little about those conditions. That meter also appears in AI comic dramas, where discarded generated shots can turn model usage into a production cost.

Why this happened now

The livestream was the visible end of a retail build-out. Tmall’s AI Space Station had already grouped offerings from Alibaba Cloud, Zhipu, Kimi, MiniMax and others. According to platform figures reported by the New Beijing News, Double 11 opening-day transaction volume for the token category on October 20 was 450% above the previous month. In other words, the shelf existed before Li Jiaqi arrived; his broadcast gave it a mass audience.

Demand supplied the second ingredient. The jump in reported token calls reflects AI moving beyond short chats into longer documents, coding tools and agents that carry out several steps. CAICT now writes about “token factories” and inference-oriented infrastructure. The hardware side of that same expansion is visible in the AI memory supercycle, where serving models puts pressure on memory bandwidth and capacity.

The final ingredient was translation into a livestream demonstration. Few people can picture a billion tokens, but most can understand a document summary or a generated presentation. The broadcast used those tasks to make an invisible service legible. What remains unproven is whether a spectacular first order becomes regular use. That is why the event matters even if the two-trillion figure is stripped of its marketing shine: an infrastructure metric has entered consumer retail, while the language used to explain it is still catching up.

Key facts

Membership orders

Nearly 100,000 / 30 min

Headline entitlement

~2 trillion tokens

Separate benefit

20 trillion tokens

Tmall token center

Launched Sep 3, 2026

China daily token calls

>140 trillion, Mar 2026

CAICT monitoring period

Jan–Aug 2026

What people are saying

Common questions

Were these cryptocurrency tokens?

No. These are language-model tokens: pieces of input or output used to measure model activity. The livestream sold access to an AI office service, not a blockchain asset or something that can be traded on an exchange.

Did viewers really consume 2 trillion tokens in 30 minutes?

No. Contemporary reports describe roughly 2 trillion tokens of membership entitlements attached to nearly 100,000 orders. Actual calls happen later, when members use the service, and the reports do not provide that consumption figure.

Does one token always equal one word?

No. Token boundaries depend on the tokenizer and the text. A word can contain several tokens, while punctuation, spaces, Chinese characters and code can be counted in ways that do not map neatly to ordinary words.

Does a bigger token number mean better AI?

Not by itself. A token total measures volume, not whether a task was completed correctly. Service quality also depends on the model, success rate, latency, throughput, context limits and how the product turns a request into a finished result.