AI News

Tokens Hit the Double 11 Shelves, AI Inference Data Volume Surpasses Training Volume for the First Time

Phoenix TechnologyUpdated Nov 12 09:39

AI summary

According to a report by Star Market Daily, this year's Double 11 saw Tokens become a new e-commerce commodity: JD's large-model Token packages sold 20 times more during a two-hour livestream compared with the previous period, and in Li Jiaqi's livestream on Tmall, Qwen Office 0.01-yuan monthly cards sold nearly 100,000 units in half an hour. Data from the China Center for Information Industry Development shows that as of June this year, China's average daily token invocation volume exceeded 175 trillion, and AI inference data volume surpassed training data volume for the first time. In addition to official stores, some unregistered AI relay stations sell Tokens at low prices, posing risks to data privacy and regulatory oversight, while chipmakers such as Intel, Qualcomm, and AMD have shifted the focus of their presentations toward on-device and edge inference.

Why read this

AI score 57

From Double 11 Token promotions and inference data volume overtaking training to chipmakers shifting toward the device side, three signals threading together the inference industry chain and the risks of gray-market relay stations.

Topics

Tags

Body · AI translation

Tokens hit the "Double 11" shelves: AI inference takes off, and gray relay stations arrive too

Science and Technology Innovation Board Daily, November 11, 2026, 23:21:40 Updated November 12, 2026, 07:11:08

Science and Technology Innovation Board Daily, November 11 (Reporter Wang Nai) — This year's "Double 11," a new type of product appeared on e-commerce shelves: Tokens.

JD.com data shows that on November 2, during a two-hour livestream in the JD 3C Digital procurement and sales livestream room, sales of large-model Token package products grew 20-fold month-on-month; from 8 p.m. on October 30 to midnight on November 1, sales of products such as AI computers bundled with Token packages sold by JD.com grew 100% month-on-month. Tmall data shows that on November 2, in Li Jiaqi's livestream room, a Qwen Office personal membership monthly card priced at 0.01 yuan sold 50,000 units within 10 minutes of going on sale, and nearly 100,000 units cumulatively within half an hour, corresponding to about 2 trillion tokens of quota.

Beyond the consumer-side change of Tokens being put on shelves, two other new developments are worth looking at together.

One is that at the end of October, data released by the China Center for Information Industry Development showed that the volume of AI inference data in China exceeded the volume of training data for the first time. As of June this year, the average daily token call volume exceeded 175 trillion, more than 1,000 times higher than at the beginning of 2024.

The other is that at the just-concluded China International Import Expo, chipmakers such as Intel and Qualcomm made agent phones, AI PCs and on-device AI solutions the focus of their displays. Wang Zhicong, chairman of Intel China, said in an interview with Science and Technology Innovation Board Daily: "In about the next five years, a larger proportion of computing power will be deployed on the inference side and extend to devices closer to the edge."

Putting these three signals together — mass consumers buying Tokens, inference data volume exceeding training data volume, and chipmakers shifting the focus of their external promotion — also reflects that inference is becoming the "prominent discipline" of the AI industry. Around AI inference, new changes are taking place across the upstream and downstream of the industrial chain, and new points of focus and risk are emerging accordingly.

Token promotion wave: channel chaos lurks beneath the discounts

When a Science and Technology Innovation Board Daily reporter searched for "Token" on Taobao, the first thing that popped up on the homepage was a dedicated "Tmall AI Space Station," labeled below with "platform-selected, buy and use immediately." Major models such as Qwen, Zhipu, MiniMax and Qoder were arrayed within it.

By comparison, buying tokens during Double 11 is indeed considerably cheaper.

Taking Qwen AI as an example, the Lite monthly card is priced at 19.5 yuan, versus 60 yuan on the official website; the Essential price is 39.5 yuan, versus 120 yuan on the official website; the Standard price is 139 yuan, versus 180 yuan on the official website.

When the Science and Technology Innovation Board Daily reporter checked the promotion pages of other model software, they found that some products were discounted by 50%, and some by as much as 70%. For example, all three MiniMax monthly cards were 50% off, Zhipu weekly cards started at 50% off, the Qwen Office basic membership monthly card started at 30% off for the first month, and the Qoder monthly card was 33% off.

In addition to Taobao, JD.com also recently launched an AI software recharge hall, including products from vendors such as Zhipu GLM, JD Cloud, Volcano Engine and DeepSeek. Among them, the Zhipu Coding personal edition GLM-5.3 monthly card package topped the IT technology e-book bestseller list.

It is understood that during Double 11, JD Cloud's newcomer points package of 10,000 points was 0.01 yuan, Kimi K3's 10,000 tokens were 0.01 yuan, and a limited edition 7-day premium experience of Zhipu AI membership could be snapped up for 0.09 yuan.

However, in addition to these official stores, some stores with small follower counts are also selling Tokens. These "AI relay stations" act as intermediaries, many of them operating in the gray areas of the internet, with operators hiding their identities throughout, no public filing information and no clear operating entity. Most of these stores have follower counts in the double digits. In the comment sections, there are also negative reviews such as "very expensive to use," "much more costly than the official website" and "the traffic is falsely labeled."

A Science and Technology Innovation Board Daily reporter randomly bought a product priced at 9.6 yuan on Taobao; the store's selling points were "the price is 30% of the official website, and the performance is the same" and "one yuan for ten million tokens."

After completing the configuration according to the operation file provided by the store, the reporter successfully called the model on WorkBuddy. The reporter then conducted a test, asking the model to write a research report on "inference data volume exceeding training data volume." The results showed that the context usage at the non-official entry point was 199.3K, costing 0.3958 yuan. The official entry point was 210.1K, a small difference between the two.

But a single test cannot show that the channel is always reliable. In a newly joined chat group, the group owner also reminded users: "It's sold too cheaply, so sometimes when there are many people the server gets crowded and connections drop and so on, please forgive us." But disconnection is only a poor experience; the problems hidden beneath the surface of "AI relay stations" include data exposure, privacy leaks, malicious implantation, remote control, data outbound transfer, and loss of management and control.

For ordinary consumers, Tokens are also not a standard commodity that can be simply compared by price. The model capabilities of different vendors are not the same, and the same amount of tokens may be able to complete different tasks under different models. The validity period of the quota after an individual recharges, refund rules and data privacy protection also await further standardization.

Wang Jiandong, deputy director of the Big Data Development Department of the State Information Center, also suggested unifying the token measurement standard as soon as possible, establishing conversion standards, and at the same time standardizing billing disclosure and introducing third-party verification.

Upstream chip and cloud vendors reshuffle their cards around "inference"

For the first time, inference data volume has exceeded training data volume, and the focus of upstream chip vendors' displays is also changing: agents, device-cloud collaboration, edge inference, and compute scheduling have become the new round of cards to play.

Wang Zhicong, chairman of Intel China, said in an interview with a reporter from the Science and Technology Innovation Board Daily that future AI capabilities need to be deployed collaboratively between devices and the cloud. Intel is working with domestic model vendors and agent vendors to extend models and applications to laptops, traditional PCs, and smaller AI Box edge computing devices.

Regarding compute solutions, Intel also exhibited a green compute warehouse that can be deployed nearby in areas with green power resources such as photovoltaics and wind power, converting green electricity locally into inference tokens and then distributing them outward through the compute network. Wang Zhicong revealed that the relevant solutions may begin to be implemented in the first quarter or the first half of next year.

"Device-side AI is evolving from single-point applications such as imaging and speech recognition to agent experiences such as intent understanding and task orchestration; the cloud takes on tasks that require stronger intelligence and compute, while the device side leverages the advantages of real-time performance, stability, privacy, and personalization, and reduces dependence on the network," said Wan Weixing, head of AI product technology for China at Qualcomm. In addition, at the 2026 Snapdragon Summit, Qualcomm also made "agents" its core narrative, with the Snapdragon X2 series processors used in agent AI PCs.

AMD's route points more directly to "locally replacing cloud tokens." The Ryzen AI Max PRO 400 series released in September 2026 is called by AMD "the world's first x86 client processor that can run 300-billion-parameter models without cloud offloading." On shipment data, Michael Nordquist, vice president of global client product marketing at AMD, revealed that AMD AI PCs have cumulatively shipped tens of millions of units with more than 250 designs; among them, agent PCs capable of running hundred-billion-level parameter models have shipped more than 500,000 units.

MediaTek, meanwhile, has officially brought the RTX Spark superchip, designed in cooperation with NVIDIA, to Windows 11 PCs, with the first batch of models covering six brands: ASUS, Dell, HP, Lenovo, Microsoft, and MSI. According to the introduction, the biggest change in the implemented experience is that agents have moved from the cloud to local devices. On the development side, it assists in writing code and testing; on the creation side, it handles video editing and 3D rendering, and can process locally offline.

For cloud services, "in the era of large models, the delivery unit of cloud services is migrating toward tokens. What cloud vendors provide is no longer just basic resources such as computing, storage, and networking, but also services including model deployment, inference acceleration, elastic scheduling, and security compliance," said Liu Hua, deputy general manager of the emerging business division at UCloud. In his view, for customers, what they buy is not just compute resources, but model inference capabilities that can be called stably.

Image source: "China Token Factory Competitiveness Ranking"

Source

Phoenix Technology