The AI Model Was the Moat. In China, It Became a Commodity.

The arithmetic cannot work. Every major cloud in China now resells the same set of models, including Zhipu's GLM, DeepSeek, Moonshot's Kimi, ByteDance's Seedance, and a thicket of near-clones. So when a sales team is told to multiply daily token revenue several times over in a single quarter, the only tools left are impossible quotas, cloud-product vouchers booked against the AI line, and a workforce that already knows the books will not close. That, per reporting in Leiphone, is roughly where China's AI cloud sales force arrived in 2026.

The human crack showed first. At one of China's first-tier cloud vendors, a regional team of more than ten AI-cloud salespeople turned in resignations within days of each other during the Q2 push. Management classified the episode as a "management incident," and the regional lead was told that one more departure would end his own job. The trigger, the salespeople said, was a quota regime that asked each of them to clear roughly 10 million yuan (about $1.4 million) in model-as-a-service (MaaS) revenue, up from the high-six-figure to low-seven-figure range that top performers cleared the year before. The math, they told Leiphone, could not be made to work without selling on terms the company would later refuse to honor.

The structural reason sits above the sales floor. Token-call switching cost is effectively zero. When a model improves, customers move on the same week, and long-term retention is thin. A salesperson quoted in Leiphone described the two phases as the same game played twice. That symmetry is the problem. The cloud giants initially told their AI sales force to push proprietary models only. When those models could not match Zhipu's GLM on coding or Seedance 2.0 on multimodal, the rule flipped. Third-party model resale revenue now counts fully toward AI sales credit. Everyone sells everyone else's models. Nobody has a moat.

That collision has produced an explicit price war. Zhipu cuts the vendor price to roughly 80% of list, and the cloud reseller hands the customer a quote at 60%. The 20-point spread is the channel's working margin, but it leaves no room for a salesperson to discount further, so vendors and resellers have moved the subsidy off the AI line and onto the rest of the cloud product. Vouchers of 20,000 to 30,000 yuan (about $2,800 to $4,200) on a deal are common. Large accounts take vouchers in the high six figures. Quarter-end, regions rotate who subsidizes whom. The reported AI-revenue number on the cloud's MaaS line is, in part, the cash value of vouchers booked against other products and rebooked as AI sales. It is real money, but it is not model revenue.

Inside that compressed margin, two model franchises are carrying most of the load. Salespeople at multiple vendors told Leiphone they believe Zhipu has captured more than half of China's domestic coding-model demand, a sales-side estimate rather than an audited share, supported by H1 reporting on Zhipu's revenue trajectory and MaaS commercialization data. On the multimodal side, Seedance 2.0, released on February 12, 2026, drew a rush from short-drama and film customers. Some were willing to prepay 3 million yuan (about $420,000) for a whitelist seat and a guaranteed call quota, and Seedance was sold at near list price with almost no discount available.

A new variable is now pulling demand out of the system entirely, per the same Leiphone reporting. OpenAI's Codex has introduced a reset mechanism that grants a fresh subscription window to users who run a referral campaign. Chinese enterprises, blocked from a direct local subscription, are reportedly issuing overseas bank cards to employees and routing coding work through those subscriptions instead of paying domestic enterprise API fees. The scale of the leak is not quantified, but it lands directly on the same quota math that already failed in Q2.

The cloud giants will keep reporting AI-cloud revenue. The line will grow. The per-salesperson target next year is unlikely to fall. The cloud's own finance organization can move vouchers around. The sales team cannot move the switching cost.