Alibaba starts selling AI tokens on Taobao October 20 after a Z.ai storefront lifted token sales 160%
Alibaba's Taobao and Tmall open an AI-token product category on October 20, 2026, selling compute bundles built with Z.ai and MiniMax. It lands days after Meta put a day-one price on its biggest open model and Reflection shipped one pitched on cost per finished task.
Vincent Jiang · 4 min read
Taobao puts AI tokens on the shelf
Alibaba's−1.61% — Alibaba, down 1.61 percent today Taobao and Tmall marketplaces will sell AI tokens as a product category when Singles' Day, its biggest shopping festival, opens official sales on Tuesday, October 20, 2026, with a record 1 million brands taking part 1. The tokens bundle computing power from the Chinese labs Z.ai and MiniMax for students, researchers and small teams; Tmall will also sell Intel−6.29% — Intel, down 6.29 percent today and AMD−6.27% — AMD, down 6.27 percent today personal computers with Alibaba's Qwen models preinstalled 1.
AI will "comprehensively enter" merchants' internal workflows from 2027, Tmall president Liu Bo said at the Singles' Day briefing, led by Qwen Work, Alibaba's enterprise AI assistant, whose e-commerce version began testing this week 1. The push follows a year of AI rollouts on the platforms that "exceeded expectations", he said 1.
The shelf already moves tokens
Tmall opened an "AI Space Station" in early September, selling subscriptions and token top-ups from Alibaba Cloud and the Chinese labs Z.ai+0.88% — Z.ai, up 0.88 percent today, MiniMax and DeepSeek. Within 48 hours of Z.ai's storefront opening, transactions in AI tokens and subscriptions across Taobao and Tmall jumped more than 160%, a Tmall spokesperson said 2.
The suppliers grow fast and lose money at it: Z.ai's first-half 2026 revenue ran about $142 million, up about 400%, against a net loss of roughly 2 billion yuan, and MiniMax's adjusted net loss more than doubled to $293 million 2. Labs that take the shelf become "more dependent on traffic controlled by the platform", analyst Poe Zhao warned 2.
Chinese open models took the usage
DeepSeek, the Chinese lab praised for its models' efficiency, handled 17% of the tokens routed by Vercel, a service that passes developers' requests to many models, by May 2026, up from under 1%, while keeping only about 1% of what those developers spent, per Rest of World 3. On OpenRouter, a second routing service, open-weight models, whose published files anyone can download and run on their own servers, carried 60% of September's token volume from US requests, with Chinese labs supplying most of it 4.
Buyers choose them for price: Pinterest's−3.65% — Pinterest, down 3.65 percent today chief executive told investors in November 2025 that fine-tuned open models cut its AI costs by 90% versus leading proprietary models 3. Alibaba pays for Qwen out of e-commerce, and its capital spending reached $18.2 billion in the March 2026 quarter, up 58% from $11.5 billion a year earlier 5.
Alibaba's capital spending has more than quadrupled in two fiscal years
Data
| Capital spending | |
|---|---|
| FY2021 | $6.27 B |
| FY2022 | $8.41 B |
| FY2023 | $4.9 B |
| FY2024 | $4.39 B |
| FY2025 | $11.51 B |
| FY2026 | $18.21 B |
Meta priced its answer on day one
Meta's−4.28% — Meta, down 4.28 percent today answer arrived priced. On October 15, 2026, it released Muse Spark Open, an 816-billion-parameter open-weight model it calls the most capable US one to date, listed by providers on day one at $1.25 per million input tokens and $4.25 per million output 34. The next morning Reflection AI$2.1B — Reflection AI, private, latest valuation $2.1B, a two-year-old startup that has raised about $4.7 billion with Nvidia−1.59% — Nvidia, down 1.59 percent today among its backers, released Beam 36.
Beam has 501 billion parameters, 23 billion of them active, and ships under Apache 2.0, a license that lets anyone host it commercially 36. Amazon−2.33% — Amazon, down 2.33 percent today put it into SageMaker JumpStart, AWS's managed model catalog, the same day; Reflection trained it with more than 100 million reinforcement-learning rollouts on 10,500 Nvidia GB300 GPUs 37. Its launch chart runs down two versions of Alibaba's Qwen 3.8, yet on Terminal-Bench 2.1, a command-line agent test, every published Chinese result still beats both US releases 348.
Chinese open models still beat both new US releases on Terminal-Bench 2.1
Data
| Terminal-Bench 2.1 score | |
|---|---|
| DeepSeek V4.1 Flash | 90.6% |
| Moonshot Kimi K3 | 85% |
| Z.ai GLM 5.3 | 83.9% |
| Alibaba Qwen 3.8 (2.4T-A95B) | 82% |
| Reflection Beam | 80.1% |
| Meta Muse Spark Open | 76.4% |
No one has priced a finished task
Reflection's hosted service is free in beta, and it has no dollar figure for cost per finished task 3. Its compute comparison is an estimate that skips the cost of reading the prompt, and its product chief concedes DeepSeek and Qwen may have "better kernels or fused kernels" 3. Meta's license demands "Built with Muse Spark Open" attribution and a separate agreement from businesses above $20 million in revenue that resell the model as a service 48.
For Alibaba holders, the readings after October 20 are opening-day AI-token sales on Tmall 1, then the cloud and AI revenue line when December-quarter results land. Qwen Work follows into merchant workflows from 2027 1. Whether the free models' usage turns into paid token sales is the question the October 20 opening begins to answer.
Deepdive
AI-generated from this story and its cited sources. Not investment advice.
