China’s two flagship AI model startups are rebuilding themselves as token merchants. A model-industry headhunter told the Silicon Lab research desk that both Zhipu and MiniMax are hiring B2B token salespeople at home and abroad, asking for three-plus years of experience at a cloud vendor or MaaS platform, ideally with coding skills of their own.
Three ways to sell tokens
Token selling is not new; Alibaba Cloud and Volcano Engine have run aggressive customer-acquisition and subsidy campaigns backed by their own compute, cloud resources and B2B experience. Zhipu and MiniMax are entering the same game. The business splits into three models. The first is metered API usage, tokens sold like utilities. Zhipu’s open platform and API now account for 86.5 per cent of revenue, replacing the local-deployment business that once delivered over 80 per cent. MiniMax’s open platform and AI enterprise services passed 60 per cent of revenue in the first half, overtaking its consumer subscription business.
The second model is subscriptions, tokens packaged like phone plans with tiers and quotas, such as Zhipu’s Coding Plan and MiniMax’s Token Plan. The third is wholesale, bundling API access for downstream AI application companies and relay stations, with model firms competing for exclusive launch rights and annual contracts.
Two strategies, one battlefield
The pair diverge in geography and emphasis. Zhipu focuses on the domestic market while MiniMax emphasises global reach, with overseas revenue at 60.8 per cent of the total in the first half. Zhipu sells tokens in laddered fashion, stressing coding and cybersecurity, and its management describes model evolution as a staircase from chat through coding and agents to autonomous AI, pricing different task difficulties differently. MiniMax sells an integrated story, pairing text models for code with multimodal models for content creation around its MiniMax-M3 push into agentic capability.
Both face the same hard reality. Cloud vendors and telecom operators are competing for the same customers. Alibaba launched a campaign this year to lift token volume, and China Mobile, China Telecom and China Unicom now sell token packages. Against them the model makers lack two things: massive compute networks and deep enterprise service channels.
Their advantages are token pricing power and control of unit intelligence cost. Zhipu says it now runs inference at hundred-thousand-card scale on domestic chips with unit token cost down 80 per cent since the start of the year. MiniMax founder Yan Junjie says the company has moved from buying compute from cloud vendors to building its own high-speed networks, servers and storage, and both are raising their use of domestic chips toward the fourth quarter.
The accounting behind the hype
The reported numbers are striking. MiniMax claims August ARR above USD 800 million with token consumption 20 times January’s level; Zhipu claims August ARR of USD 1.6 billion, up 60 per cent in a month. The caveat is that ARR reflects revenue at a point in time, not confirmed financial income or true profit structure.
Token selling is drifting toward a capital-heavy business that demands channel resources and frontline customer demand, and its gross-margin ceiling depends less on how smart a model is than on the combined cost advantage of model and compute. For Zhipu and MiniMax the fight is unavoidable, and it is getting harder. Whoever controls the ability to keep delivering cheaper intelligence inside real workflows holds long-term pricing power; everyone else is renting time.
Editor’s note: This is an adapted translation of the original Sohu Tech report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://www.sohu.com/a/1071683580_116132.