Zhipu Hits Taobao, Moonshot Courts Microsoft for Royalties — China’s AI Race Has a New Scorecard
China's large language model companies are abandoning the growth-at-all-costs playbook, with Zhipu AI opening a retail storefront on Alibaba's Tmall and Moonshot AI negotiating revenue-sharing arrangements with Microsoft, Amazon and Google — a structural shift that signals the sector's transition from capability demonstration to sustainable unit economics.
The twin moves, executed within days of each other in early September 2026, represent the most concrete evidence yet that China's AI foundational model companies are under mounting pressure to convert astronomical usage growth into durable cash flows. Zhipu's first-half results, released just ahead of its Tmall launch, showed the company's API and open-platform revenue exploding 27-fold year-on-year — yet its adjusted net loss still widened 12.1% to RMB 1.964 billion (US$272.8 million), and its shares fell more than 5% after earnings missed analyst consensus by roughly 30%.
The market's verdict was unambiguous: volume alone is no longer sufficient.
Zhipu Reshapes Its Revenue Stack Around Subscription Tokens
Zhipu AI's decision to list its GLM Coding Plan on Tmall — making it one of the first domestic LLM vendors to treat token bundles as a consumer retail product — is less a marketing stunt than a deliberate restructuring of its revenue architecture.
The financials justify the urgency. In the first half of 2026, Zhipu generated total revenue of RMB 954 million (US$132.5 million), up 399.7% year-on-year and already exceeding the company's full-year 2025 revenue of RMB 724 million (US$100.6 million). The engine behind that growth was almost entirely its open-platform and API business, which contributed RMB 825 million (US$114.6 million) — a 27-fold increase — lifting its share of total revenue from 15.2% a year ago to 86.5%.
The operational metrics are equally striking. By end-August 2026, Zhipu's MaaS platform had accumulated more than 7.4 million registered users, up 144% from the start of the year, while paying daily active users surged 603%. Token call volume expanded more than 40-fold, and the top-10 revenue-generating clients saw average daily call volumes rise 98-fold. On an annualized basis, the platform's monthly ARR reached US$1.6 billion, with weekly ARR touching US$2 billion.
Yet the structural tension is hard to ignore. As high-margin localized deployment contracts — the legacy government and enterprise project business — shrank as a share of revenue, the fast-growing but lower-margin API segment dragged overall gross margin from 50% in H1 2025 down to 26.4%. Sales costs surged 635.4% year-on-year to RMB 702 million (US$97.5 million). The API business did cross into positive gross margin territory — reaching 24.6% versus -0.4% a year earlier — but the trajectory underscores how much cost leverage remains to be unlocked.
The Tmall storefront addresses this directly. Subscription tiers for the GLM Coding Plan — built on the latest GLM-5.3 model and compatible with more than 20 agentic coding tools including ZCode, ClaudeCode and Codex — are priced from RMB 118 per month for the Lite tier to over RMB 1,000 per month for the Max tier. Notably, pricing has been revised sharply upward from earlier in the year: the Pro tier jumped from RMB 149 to RMB 538, a 3.6-fold increase, justified by a shift to a token-credit system with off-peak discounts and upgraded model capabilities.
The commercial logic mirrors a mobile data plan: fixed monthly fees convert volatile per-call revenue into predictable subscription cash flows, while the Tmall channel provides a lower-friction acquisition funnel for individual developers and independent software vendors — a segment that is difficult to reach through enterprise sales motions. For Zhipu, the path from government AI contractor to token-economy platform is now a matter of margin management, not business model validation.
Moonshot Designs a Royalty Layer Inside Global Cloud Infrastructure
Where Zhipu is pushing tokens toward end consumers, Moonshot AI is engineering a fundamentally different revenue topology — one that does not require the company to own the customer relationship at all.
Reuters reported on August 26, 2026 that Moonshot is in early-stage negotiations with Microsoft (for Azure), Amazon (for AWS) and Alphabet's Google Cloud to embed its Kimi K3 model into their respective inference infrastructure, with Moonshot seeking a revenue share of up to 30% on K3-related services generated through those platforms. All three hyperscalers and Moonshot declined to comment.
The mechanism is not improvised. When Moonshot released Kimi K3 with open weights, it embedded a commercial gate in the license: any entity whose affiliated revenue from K3-related MaaS services exceeds US$20 million over a rolling 12-month period must enter into a separate commercial agreement with Moonshot before deploying K3 for commercial purposes. This clause effectively converts open-source distribution into a deferred monetization funnel — the model proliferates freely until usage reaches commercial scale, at which point Moonshot re-enters the transaction.
Reuters had flagged a parallel track as early as August 7, 2026, reporting that Moonshot was designing revenue-sharing structures for K3's large commercial users at the same 30% ceiling. On July 20, 2026, Chinasoft International disclosed via Hong Kong Stock Exchange filing that it had signed a token revenue-sharing and joint innovation agreement with Moonshot, under which the two parties will co-develop enterprise-grade agents for the energy, power and financial sectors — Chinasoft handling industry delivery, Moonshot providing the Kimi model layer, with token consumption revenue split at an agreed ratio.
The global inference ecosystem is already taking shape. Moonshot's developer portal lists dedicated entry points for third-party inference providers; Together AI, Fireworks, DigitalOcean, Modal, Baseten and DeepInfra are among more than 10 publicly identified providers already serving or supporting K3. Together AI announced a strategic partnership with Moonshot in late July 2026 to natively serve Kimi models on U.S. infrastructure. Modal partnered with Moonshot and vLLM at K3's launch to offer both shared API and dedicated deployment options.
This architecture inverts the traditional software export model. Rather than selling a product to overseas users, Moonshot is embedding its model as a revenue-generating asset inside third-party commercial infrastructure — collecting a royalty each time the model is consumed, regardless of which platform intermediates the transaction.
The strategic risk is equally clear: Moonshot's negotiating leverage is a direct function of K3's technical differentiation. Should a competing model surpass K3's capabilities, the 30% revenue-share ask becomes untenable. The company's reported confidential A1 filing with the Hong Kong Stock Exchange — disclosed by media on September 2, 2026, the same day as the Tmall launch — alongside a concurrent fundraising round at a pre-money valuation of US$50 billion, suggests Moonshot is racing to institutionalize these commercial structures before the competitive window narrows. Moonshot declined to confirm or deny the IPO reports.
Sector-Wide Profitability Remains Elusive Despite Hypergrowth
The urgency driving both strategies is visible across the broader Chinese LLM landscape, where revenue growth is dramatic but profit conversion remains the defining unsolved problem.
DeepSeek generated approximately RMB 475 million (US$66 million) in revenue in the first seven months of 2026 — roughly 10 times its full-year 2025 figure — with an annualized run rate of US$400 million to US$500 million, according to The Information. Its API business carried a gross margin of 82.9%, with overall gross margin at 44.6%, making it the sector's clearest example of unit economics working in favor of the model provider.
MiniMax posted H1 2026 revenue of approximately US$120 million, up 283.1% year-on-year and already exceeding its full-year 2025 revenue of US$79 million. By August 2026, its annualized revenue run rate had crossed US$800 million. However, H1 cost of sales reached US$95.76 million — up 62.6% against full-year 2025 — and net losses remained at US$358 million for the period, illustrating that scale alone does not resolve the cost structure.
The pattern is consistent: token volume is compounding, but compute costs, customer acquisition costs and revenue quality are compounding alongside it. Zhipu's post-earnings stock decline — despite a near-400% revenue increase — reflects market skepticism that the current growth trajectory translates into the kind of unit economics that justify current valuations.
The next competitive dimension, then, is not which model scores highest on a benchmark, but which company can simultaneously compress inference costs, lock in recurring revenue relationships and convert ARR into free cash flow. The companies that establish durable distribution — whether through Tmall's 900-million-user retail funnel or through royalty embeds in hyperscaler infrastructure — will carry a structural advantage that pure technical capability cannot easily replicate.
Model capability sets the ceiling. Unit economics determine the floor.
Related Coverage:
Zhipu’s H1 Revenue Surges 400% as API Pivot Cuts Gross Margin in Half
Moonshot AI Files Confidentially for Hong Kong IPO at $50 Billion Valuation