1. TL;DR — Capital Markets and Industry Converge on Three Tracks

The third day of China’s 2026 National Day holiday closes the curtain on the Q3 earnings season. Zhipu, DeepSeek, Moonshot AI, StepFun, and MiniMax have all completed funding rounds, listings, or confidential filings within weeks. U.S. and Chinese foundation-model labs are simultaneously sprinting to public markets. At the same time, 2026 has been widely recognized as the scale-out year for on-device AI: agent-native smartphones, intelligent cockpits, and AI PCs are ramping on three parallel tracks, with 47% of enterprises already running end-cloud collaborative deployments. This article traces two intertwined storylines: the Q3 valuation reset of China’s “Six Tigers” and the on-device LLM penetration inflection.

2. Q3 Capital Moves Across the Six Tigers

2.1 Zhipu: A US$5B Round Resets the Bar

On September 13, Zhipu announced a roughly US$5 billion financing package — about US$2 billion in share placement plus approximately US$3 billion in zero-coupon convertible bonds (conversion price HK$892.50, a 12.55% premium to the prior close). Just two months earlier, in July, Zhipu had raised around US$4 billion via placement. Cumulative 2026 proceeds now approach US$9 billion. Zhipu’s 2026 H1 interim report shows revenue of ¥954 million (+399.7% YoY), with open-platform and API revenue reaching ¥825 million (+2,735.7%) and lifting from 15.2% to 86.5% of total revenue. Platform token consumption grew more than 40× since the start of the year, paying DAU rose 603%, enterprise and developer users exceeded 7.4 million, and the average API selling price climbed 101%. On September 11, Zhipu’s HKEX-listed shares closed at HK$831.5, lifting total market cap to HK$387.166 billion — a cumulative 6.2× gain over its IPO price.

2.2 DeepSeek: From Research Lab to STAR Market

DeepSeek executed a dramatic corporate reversal between April and September 2026. In June it closed a roughly ¥50 billion first round at a post-money valuation of about ¥350 billion, with Tencent, the CATL ecosystem, JD.com, NetEase, IDG Capital, and the National AI Industry Fund on the cap table. A second round opened in July at a pre-money valuation near ¥500 billion, with SMIC Capital, Boyu Capital, CPE Summit, and Hefei state-backed funds joining. On September 9, DeepSeek engaged CITIC Securities to prepare a STAR Market IPO filing. Revenue for the first seven months reached ¥475 million — roughly 10× the full-year 2025 figure — alongside a net loss of ¥715 million and infrastructure spend up nearly 10× year on year. Founder Liang Wenfeng personally committed about ¥20 billion to the first round.

2.3 Moonshot, StepFun, and the Remaining Two

Moonshot AI submitted a confidential A1 filing to HKEX on September 2, targeting roughly US$3 billion in proceeds and an earliest listing window of Q1 2027. A subsequent STAR Market dual listing has also been floated. StepFun completed share-structure reform and offshore-architecture unwinding in April and closed nearly US$2.5 billion in May, with an HKEX target before end-2026. Its B2B2C model is embedded in over 60% of China’s top-tier smartphone brands, with cumulative installed base surpassing 42 million units. Baichuan has pivoted to serious medical AI; 01.AI has wound down base-model pretraining and shifted to enterprise solutions. Both have placed IPOs in 2027.

2.4 Already Listed: MiniMax Draws Down, Zhipu Hits New Highs

MiniMax priced its January IPO at the top of the range at HK$165, closed day one at HK$345 (+109%), and pushed market cap above HK$100 billion. By September 11, total market cap stood at HK$102.535 billion — a roughly 78% drawdown from the March 18 peak of HK$464 billion. The two newly listed names, both from January, have seen first-day gains diverge by 8×, and secondary markets are repricing each business model very differently.

3. Across the Pacific: Anthropic Targets a US$2 Trillion Valuation

SemiAnalysis projects Anthropic Q3 2026 GAAP EBIT above US$1 billion, implying roughly 6% margin. ARR has surged from US$9 billion at end-2025 to over US$60 billion today. NNARR is running around US$15 billion per month; if sustained, ARR could break US$300 billion by end-2027, implying a US$6 trillion valuation. Anthropic confidentially filed for IPO on June 1. Claude Code now accounts for over 7% of all commits on GitHub, directly driving monthly ARR from US$3 billion in January to US$11 billion in March. Structurally, 75-85% of Anthropic ARR comes from usage-based APIs, with consumer subscriptions at only 5% — a sharp contrast with OpenAI’s revenue mix.

4. On-Device LLMs: 2026 as the Scale-Out Year

4.1 Agent-Native Smartphones Go Mass-Market

Three agent phones shared the WAIC 2026 spotlight: StepFun’s STEPX Neo (running the agent-native “Step AOS,” first to clear L3 certification), Honor’s RobotPhone (embodied Agentic OS plus a hidden motorized gimbal), and Nubia’s NaviX Ultra (system-level GUI Agent powered by Doubao, Snapdragon 8 Elite, 16GB + 512GB). All three place agents as the first-class OS entry point, with cross-app multi-step tasks running automatically.

4.2 Intelligent Cockpits and AI PCs in Parallel

The Eva super-agent jointly built by Qianli Technology, Geely, and StepFun is in mass production on the Zeekr 8X, with 34B-grade models deployed locally on cockpit hardware — a new benchmark. On the AI PC side, 7B models now run stably on thin-and-light laptops without discrete GPUs, leveraging integrated CPU NPUs to deliver a fully offline office AI loop.

4.3 Headline Numbers

  • IDC forecasts 2026 China AI smartphone shipments at 147 million units (+31.6%), with penetration above 53%.
  • Counterpoint projects 432 million Gen-AI smartphones shipped globally; on-device AI market size surpasses ¥800 billion.
  • 47% of enterprises have deployed end-cloud collaborative architectures.
  • The September flagship refresh wave — Xiaomi 18 Fold, iPhone 18 Pro, OPPO Find X9 Ultra — all bundle on-device LLMs with cross-app agents.

5. Three Convergence Signals

  • Capital track: U.S. and Chinese players race to public markets simultaneously. China’s 2026 cumulative LLM sector funding is on track to exceed ¥100 billion. Valuation multiples are now set by four metrics: token consumption, ARR, installed base, and API selling price.
  • Industry track: On-device LLMs have spread from flagships to thousand-yuan phones. In Q2 2026, over 60% of newly shipped smartphones globally shipped with pre-installed open-source on-device inference engines.
  • Regulatory track: The National AI Industry Fund, as the only direct state investor in DeepSeek’s first round, signals “patient capital” as the Q4 main theme.
  • Application track: Agent-native operating systems (STEPX Neo / Eva / Wukong) are replacing the app-entry paradigm; cross-app tasks are moving from “demo” to “mass production.”

6. FAQ

Q1: Why are U.S. and Chinese LLM companies concentrating their public-market push in Q3-Q4 2026? A: Compute, talent, and data thresholds have risen to the US$1 billion-per-round level. Private markets alone cannot finance the next generation of frontier training. At the same time, narrowing expectations on NVIDIA H-series access for China are pushing Chinese players to lock in overseas cloud-compute contracts and listing valuations before the window potentially closes.

Q2: How did Zhipu grow API revenue 2,735.7% YoY? A: Three reinforcing factors: ① Token volume +40×, driven by the GLM open-source ecosystem; ② Paying DAU +603%, with enterprise and developer users past 7.4 million; ③ Average API selling price +101% — MaaS platform pricing power scales in lockstep with usage.

Q3: How is the 2026 “scale-out year” different from earlier on-device pilots? A: The watershed combines three forces: NPU compute democratization, mature open-source lightweight models, and the emergence of agent-native OS. In Q2 2026, over 60% of new smartphone shipments globally pre-installed on-device inference engines. Thousand-yuan phones now run full LLM capability smoothly, with the legacy trio of pain points — disconnection, privacy, latency — largely resolved.

Q4: Why has MiniMax drawn down 78% while Zhipu hit new highs? A: Business model divergence. Zhipu runs a MaaS platform with 80%+ revenue from enterprise APIs directly tied to token consumption. MiniMax rides an emotional-companion + overseas consumer track, where ARR swings with consumer subscription volatility and the broader “Six Tigers” valuation digestion cycle. Public markets are pricing growth certainty differently.

Q5: What are the core variables of the LLM “endurance race”? A: Three industry voices converged on: ① cash-generation capability (API + ARR price simultaneously rising); ② compute moat (self-built data centers + domestic chip share); ③ ecosystem penetration (B2B2C installed base). The era of pure parameter/compute competition is over; the second half will be won by whoever first converts model capability into sustainable cash flow.

7. Key Terminology

  • Six Tigers: China’s six leading foundation-model startups — Zhipu, MiniMax, Moonshot AI, StepFun, Baichuan, 01.AI.
  • MaaS: Model as a Service — cloud-side business model charging by token consumption.
  • ARR: Annual Recurring Revenue — the core growth metric for SaaS / MaaS companies.
  • GAAP EBIT: Earnings Before Interest and Taxes under generally accepted accounting principles; a standard measure of core-business cash generation.
  • NNARR: Net New ARR — the absolute incremental figure; Anthropic currently runs about +US$15 billion per month.
  • End-Cloud Collaboration: Cloud handles training and complex generation; devices handle high-frequency, real-time, privacy-sensitive local inference.
  • Agent-Native OS: An OS that treats the agent as the first-class entry point (STEPX Neo / Wukong), enabling cross-app tasks without launching individual apps.
  • NNARR vs. NRR: NNARR is the absolute increment; NRR (Net Revenue Retention) is the retention multiplier on the same customer cohort — Anthropic’s NRR is reported at 500%.

8. References

Industry Reports

Official Sources

Authoritative Media