On-premise revenue weighed on first-half results, but accelerating API and ARR prompted a price-target increase to HK$2,000

Overview

The report reviews Zhipu AI's 1H26 results and shifts the analytical focus from the on-premise deployment revenue miss to the rapid expansion of its cloud/API business. JPMorgan believes model iterations are simultaneously driving increases in usage, pricing, and ARR, while inference efficiency and commercial gross margin are also improving. It therefore raises its revenue forecasts and price target, but emphasizes that profit improvement will continue to lag revenue growth.

Core views

The apparent 1H26 revenue miss came almost entirely from the on-premise deployment business. Zhipu AI generated revenue of Rmb954mn, up 400% year over year, but 22% below JPMorgan's forecast and 30% below consensus expectations. Within this, on-premise deployment revenue declined 20% year over year to Rmb129mn, 64% below JPMorgan's forecast; the report views this as a business the company is proactively scaling back. By contrast, cloud/API revenue reached Rmb825mn, up 2,736% year over year and 412% sequentially, broadly in line with JPMorgan's forecast, while its share of total revenue rose from 15.2% a year earlier to 86.5%. The report therefore believes the business focus has clearly shifted toward the more scalable API model and that growth trends should not be judged solely by the on-premise revenue shortfall. The more important change occurred after June. Zhipu AI's ARR has historically increased in step changes following model and product upgrades, and the current increase is the largest to date: ARR reached US$1.6bn in August 2026, 46 times the December 2025 level, implying revenue of approximately US$133mn in August alone, already exceeding total 1H26 cloud/API revenue of US$123mn. Management expects ARR to rise by another 50% to US$2.4bn by year-end, significantly above its previous guidance of US$1bn. The acceleration is being driven by both usage and pricing: MaaS Token consumption increased by more than 40 times from the beginning of the year, Coding Plan usage rose by more than 23 times, average API pricing increased by approximately 101%, paid daily active users grew by 603%, and average daily usage among the top ten customers by revenue increased by 98 times. Based on this, JPMorgan concludes that GLM-5.2/5.3 has not only generated more usage but has also improved realized pricing and paid conversion. Revenue expansion is beginning to improve unit economics, but profit realization will still take time. Cloud/API gross margin rose from -0.4% in 1H25 and 18.9% in FY25 to 24.6% in 1H26, while group gross margin increased by 16 percentage points year over year to 26.4%. Adjusted net loss was Rmb2.0bn, below R&D expenditure of Rmb2.1bn, indicating that commercial gross profit, after sales and administrative expenses, can already fund part of the R&D investment. However, R&D expenditure still exceeds revenue by more than two times and absolute investment requirements remain high, so the report expects profit improvement to lag the recent revenue acceleration. The financial model shows adjusted net losses of Rmb3,547mn and Rmb1,968mn in 2026E and 2027E, respectively, before turning to adjusted net profit of Rmb6,651mn in 2028E; operating cash flow is expected to shift from -Rmb1,278mn in 2026E to Rmb4,721mn in 2027E and Rmb20,572mn in 2028E. Inference efficiency and computing-power supply are another key determinant of whether API growth can continue. Management stated that inference cost per Token has declined by approximately 80% year to date, while end-to-end serving performance on the same domestic hardware has improved by approximately 3 times. Zhipu AI has achieved production inference at a scale of approximately 100,000 domestic accelerator cards, and GLM-5.3 Flash processed more than 60 trillion Tokens on domestic chips in six days; management expects to obtain more advanced domestic GPUs over the next 3—6 months. JPMorgan believes the increase in physical computing power and higher throughput per card will jointly expand the supply of billable Tokens and support rapidly growing API demand. At the application level, management summarizes the capability evolution path as Chat→Coding→Agent→Co-work→Autonomous AI. Coding has become the largest revenue source while also providing a training environment for long-horizon planning, tool use, verification, and error recovery. Cybersecurity is currently the clearest adjacent use case: since GLM-5.2, Zhipu AI has identified 2,400 expert-screened vulnerabilities across 269 projects, of which more than 1,000 were high-risk vulnerabilities. Legal, financial, and data-analysis applications remain at an earlier stage and have not yet generated meaningful revenue. The report is therefore optimistic about Coding's migration toward higher-value professional workflows, but commercialization in non-Coding use cases still needs to be validated. Products and channels continue to expand. While retaining the approximately 745bn-parameter base model of GLM-5.2, GLM-5.3 improved end-to-end task completion by more than 50% through deeper post-training. The next-generation GLM foundation model has entered training, with objectives including greater effective scale, a longer native context, and native multimodal capabilities; the longer-term direction is to enable models to generate data, build training environments and verifiers, and optimize their own infrastructure. Overseas, the company is discussing open-weight models, local hosting, and revenue-sharing arrangements with overseas cloud service providers. Management expects further progress over the next 1—2 months, potentially expanding distribution without requiring Zhipu AI to bear all incremental computing-power needs by itself. Based on the assumption of US$2.4bn in year-end ARR, JPMorgan raises its 2026—2030E revenue forecasts by 14%, 18%, 11%, 14%, and 14%, respectively, to Rmb6,102mn, Rmb17,759mn, Rmb45,589mn, Rmb127,003mn, and Rmb220,397mn. 2026E adjusted EPS was lowered from -Rmb7.68 to -Rmb8.02, a change of -4.5%; 2027E adjusted EPS improved from -Rmb4.39 to -Rmb4.27, a change of +2.8%, again demonstrating that revenue improvement precedes profit improvement. The report maintains Overweight and raises its December 2026 price target from HK$1,800 to HK$2,000. The price target is based on 2030E normalized adjusted net profit of Rmb61,138mn, adjusted EPS of Rmb133, and a 20 times P/E multiple, discounted back at a 15% WACC; the 20 times valuation exceeds that of China's leading internet companies, with the report explaining the premium through its forecast of more than 100% revenue CAGR over 2026—2030E.

Analysis framework

The report first compares actual 1H26 results with JPMorgan's forecasts and consensus expectations, then breaks down the revenue shortfall between on-premise deployment and cloud/API. It subsequently assesses growth momentum using ARR, Token usage, customer usage, and API pricing before and after model releases, while evaluating scalability economics through gross margin, inference costs, computing throughput, and R&D expenditure. Finally, the report incorporates US$2.4bn in year-end ARR into its 2026—2030 financial forecasts and calculates the price target using a forward P/E and WACC discounting framework.

Methodology notes

• Industry/Sector Analysis FrameworkVolume-price decomposition

  API usage and pricing decomposition

  The report separately examines MaaS Token consumption, Coding Plan usage, paying users, usage by major customers, and average API pricing to show that revenue acceleration is being driven simultaneously by greater usage scale and higher realized pricing.

• Event-driven Strategy and Behavioral FinanceEvent-driven analysis

  Relationship between model releases and step changes in ARR

  The report treats successive model and product upgrades as triggers for step changes in ARR and uses changes in usage, pricing, and ARR following the releases of GLM-5.2/5.3 to assess the new models' commercial conversion.

• Competitive and Strategic FrameworkEconomies of scale / learning curve

  Inference unit economics and scalability

  An approximately 80% decline in cost per Token and an approximately 3 times improvement in serving performance on the same hardware indicate that algorithmic and engineering efficiency gains can increase billable supply per card and improve API gross margin.

• Valuation MethodPE/PEG valuation

  Discounted forward normalized P/E valuation

  The report applies a 20 times P/E multiple to 2030E normalized adjusted EPS and discounts it back to December 2026 using a 15% WACC; the valuation premium primarily reflects its forecast of more than 100% revenue CAGR over 2026—2030E.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• Zhipu AI (Z AI Co Ltd - H, 02513.HK)

  The report's sole primary covered company; cloud/API, Coding, and model iterations are the core sources of growth, while declining on-premise deployment revenue accounted for the 1H26 results shortfall.

  Strengths

  The GLM series has repeatedly achieved a leading domestic position; model capabilities can translate into paid usage and pricing improvements; cloud/API gross margin and inference efficiency are improving; Coding has become the largest revenue source.

  Weaknesses

  R&D expenditure still exceeds revenue by more than two times, and profit improvement lags revenue growth; non-Coding applications such as legal, financial, and data analysis have not yet generated meaningful revenue.

  Comparison

  JPMorgan prefers Zhipu AI among Chinese large-model companies and assigns it a higher valuation multiple than China's leading internet companies to reflect its forecast of more than 100% revenue CAGR over 2026—2030E.

  Risks

  The company faces export-control, geopolitical, and entity-list risks, as well as risks related to competition, sustained high R&D investment, commercialization, customer adoption, and external computing-power supply.

Key data

• 1H26 total revenueRmb954mnUp 400% year over year, 22% below JPMorgan's forecast and 30% below consensus expectations.
• 1H26 on-premise deployment revenueRmb129mnDown 20% year over year and 64% below JPMorgan's forecast.
• 1H26 cloud/API revenueRmb825mnUp 2,736% year over year and 412% sequentially, accounting for 86.5% of total revenue versus 15.2% a year earlier.
• August 2026 ARRUS$1.6bn46 times the December 2025 level, implying August monthly revenue of approximately US$133mn.
• Year-end 2026 ARR guidanceUS$2.4bnA further 50% increase from August, versus previous management guidance of US$1bn.
• API growth metricsMaaS Token usage >40 times; Coding Plan usage >23 times; average API price approximately +101%Paid DAU grew by 603%, while average daily usage among the top ten customers by revenue increased by 98 times.
• Cloud/API gross margin24.6%-0.4% in 1H25 and 18.9% in FY25; 1H26 group gross margin increased by 16 percentage points year over year to 26.4%.
• R&D and lossesAdjusted net loss of Rmb2.0bn; R&D expenditure of Rmb2.1bnCommercial gross profit has funded part of R&D expenses, but R&D expenditure still exceeds revenue by more than two times.
• Inference efficiencyCost per Token approximately -80%; serving performance approximately 3 timesBoth represent improvements from the beginning of 2026 through the reporting period; the company has achieved production inference at a scale of approximately 100,000 domestic accelerator cards.
• Cybersecurity applications2,400 vulnerabilities, 269 projectsMore than 1,000 of these were high-risk vulnerabilities.
• 2026—2030E revenue forecastsRmb6,102mn / Rmb17,759mn / Rmb45,589mn / Rmb127,003mn / Rmb220,397mnRaised by 14%, 18%, 11%, 14%, and 14%, respectively, versus the previous forecasts.
• Price-target valuation parameters20 times 2030E P/E; 15% WACCBased on 2030E adjusted net profit of Rmb61,138mn and adjusted EPS of Rmb133, resulting in a December 2026 price target of HK$2,000.

Impact & implications

The report believes Zhipu AI's revenue mix is shifting from one-off, less scalable on-premise deployments toward cloud/API, with improvements in model capabilities translating simultaneously into higher usage and higher pricing. Improvements in gross margin, inference costs, and computing efficiency are allowing commercial revenue to begin funding part of R&D investment, but high R&D intensity means profit realization will continue to lag significantly. If API run rates, domestic computing-power supply, non-Coding applications, and overseas channels progress as management expects, the report's revenue upgrades and valuation premium will receive further support.

Risks

• Export controls, geopolitical risks, or inclusion on an entity list could restrict the company's operations and access to resources.
• Intensifying competition could weaken Zhipu AI's leadership in model capabilities, pricing power, or customer growth.
• Sustained and substantial R&D investment creates execution risk and continues to pressure profitability.
• Commercialization progress and customer adoption remain uncertain, particularly in applications beyond Coding.
• The company depends on computing infrastructure and external suppliers and therefore faces risks related to computing-power costs and availability.

What to watch

• Monitor the API revenue run rate in September 2026 and the fourth quarter to verify whether the August ARR acceleration can continue.
• Monitor adoption and revenue contributions from use cases beyond Coding.
• Monitor whether local hosting, open-weight models, and revenue-sharing partnerships with overseas CSPs can achieve commercialization.
• Monitor the next-generation GLM's capabilities, inference costs, and relative competitive positioning.