Tokens Hit the E-Commerce Shelf: The Era of Compute Commoditization Begins

On September 2, 2026, Zhipu — the Hong Kong-listed "first stock of large models" — opened an official flagship store on Tmall. The goods on the shelf were not models or API documentation, but four clearly priced token subscription plans: Lite for individuals at 118 yuan per month, Pro at 538 yuan, Max at 1,078 yuan, and a team edition at 598 yuan per seat. Users add AI compute to their shopping cart like topping up a phone, and credits land in their account right after payment.

One day later, Tmall launched the "AI Space Station" token top-up center, where subscription plans from Alibaba Cloud, Zhipu, Kimi, MiniMax and other leading model vendors went on sale together, supporting both card-code and direct top-up delivery. For the first time, large-model capability has been packaged as a standardized retail product and placed on traditional e-commerce shelves. The era of compute commoditization has officially begun.

From Selling Projects to Selling Subscriptions

Zhipu's store opening is not a simple channel expansion — behind it is a dramatic shift in revenue structure. Its interim results show first-half 2026 revenue of 954 million yuan, up 399.7% year over year. The open platform and API business alone brought in 825 million yuan, up 2,735.7%, jumping from 15.2% of total revenue a year earlier to 86.5%. Meanwhile, traditional project-based revenue such as local deployment fell by more than half.

This is the transition from "deploying models into customer data centers and booking revenue per contract" to "putting models on the cloud and charging continuously by usage." The former is a business of delivery, security and client relationships; the latter lives or dies by model capability, developer ecosystem and inference cost — calls keep coming, revenue keeps flowing.

Zhipu even wrote a self-defined formula into its report: "AGI business value = intelligence ceiling × token consumption scale." Whether the formula holds is debatable, but the signal is clear: token consumption is becoming the new anchor for valuation narratives in the LLM industry.

Why E-Commerce Shelves: Traffic, Mindshare and Positioning

Large-model subscriptions used to be sold almost exclusively through official websites, reaching professional developers. What Zhipu lacked was a "store" with foot traffic: individual developers, small teams and SMEs with procurement needs could hardly discover or trust an API service through a website.

Tmall had exactly what was needed: search traffic, mature payment and invoicing systems, membership and customer-service mechanisms, plus exposure from promotions and enterprise procurement. On opening day, Zhipu's brand search volume surged 40x, and within 48 hours, token-related product transactions on Taobao and Tmall grew over 160% week over week. The e-commerce toolkit — promotions, price comparison, coupons, subsidies — is being replicated for AI compute as a virtual good.

For the platform, AI subscriptions are a premium virtual category: high average order value, no logistics, and built-in renewals. As early as April, Tmall issued rules for AI software listing, requiring sellers to specify token allowances, model versions and API validity periods. Whoever secures the model vendors' "first store" gets to define the category rules and show the template.

Volume Up, Price Down: The Engine of Commoditization

Tokens can be sold like phone credit only because they have become a massively consumed standardized commodity. The numbers are striking: China's daily token calls grew from roughly 100 billion at the start of 2024 to over 140 trillion by March 2026, and past 500 trillion by June — a thousand-fold jump in about two years. In March, "token" received its official Chinese name "词元," giving this measurable, priceable, tradable compute resource an identity.

Against surging consumption, prices collapsed. In early September 2026, the LLM Token Expenditure Index fell below $1 per million tokens for the first time, as domestic price wars drove costs to historic lows. Under this scissors gap of volume up and price down, subscription and top-up models have become the common choice for model vendors to lock in users and smooth cost volatility. That is also why Zhipu's average API price rose about 101% while gross margin turned positive: subscription stickiness stabilizes existing users, while new-user scale spreads fixed costs.

A Multi-Sided Positioning War

Token retail was never Zhipu's business alone. The three major telecom operators are already in: China Telecom launched commercial trial token plans in May with a personal light tier at 9.9 yuan per month, while China Mobile and China Unicom rolled out low-cost plans, some bundling tokens with cloud PCs. With massive user bases, mature payment systems and compute networks, the carriers treat tokens as the next "data package" after traditional traffic revenue peaked.

Cloud vendors are doubling down too: Alibaba set up the AlibabaTokenHub business group, and Tencent Cloud upgraded its all-domain TokenHub — "selling tokens" has become the most crowded business on the AI value chain. When compute, models and hardware meet on the same shelf, the retail race for AI services has only just begun.

Three Hard Problems Behind Commoditization

The direction is clear, but execution is not smooth. First, cross-platform fulfillment is a hurdle: after ordering on Tmall, users must still register on Zhipu's platform, obtain an API key, and configure it in their coding tools before anything works — each step filters out non-technical buyers. Second, responsibility is fuzzy: physical goods rely on logistics confirmation, while an AI subscription sells a right to use over time; how platform rules and service agreements split responsibility has no clear answer, and individual subscriptions are generally non-refundable. Third, credits are not output: a phone plan specifies data volume, but an AI plan buys credits that are burned by compute consumption without quantifying model output — the same 118-yuan tier may finish several small projects for one user and be drained by a single complex codebase for another.

These frictions show that turning compute into the "new phone credit" is about more than channels: fulfillment experience, pricing transparency and value measurement are the long-term tests.

Conclusion: Competition Shifts From Parameters to Business Models

When tokens are really sold like phone credit, the next contest in large models is no longer who has more parameters or a higher leaderboard, but who can turn "top up when you run out" into a stable, renewable business. The e-commerce shelf gives model vendors a new gateway to massive consumer and SMB demand, and for the first time presents compute's commodity nature in full.

The era of compute commoditization has arrived — it changes not just how we buy, but the industry's narrative switch from a technology race to a business race. For practitioners, understanding the token economy may matter more than chasing the next model leaderboard.