Z.AI's ARR and API commercialization are accelerating rapidly, but the cloud transition and high R&D investment continue to weigh on earnings visibility

Overview

This report reviews Z.AI's 1H26 results and analyzes the gap between ARR and revenue, the commercialization and margin trajectory, next-generation models, compute expansion, and valuation. Goldman Sachs acknowledges the company's acceleration in APIs, pricing, and enterprise adoption, but believes that weakening on-premises business, persistently high R&D investment, and cash burn leave the current risk-reward relatively balanced.

Core views

Goldman Sachs characterized the 1H26 results as mixed. The company generated revenue of Rmb954mn, up 400% year over year and 79% half over half, but 47% below Goldman Sachs' forecast of Rmb1,808mn. Open Platform/API revenue was Rmb825mn, up 2,736% year over year and 412% half over half, 4% above market consensus but 38% below Goldman Sachs' forecast due to differences between gross and net accounting and ARR growth being more weighted toward the second half. On-premises deployment revenue fell to Rmb129mn, down 20% year over year and 65% half over half, 73% below Goldman Sachs' forecast, reflecting transitional pressure as the business shifts from on-premises deployment to cloud-based MaaS. Profitability was similarly mixed. The API business gross margin reached 24.6%, above Goldman Sachs' 24.0% forecast and 2H25's 22.4%, improving despite higher compute costs. However, weaker-than-expected profitability in on-premises deployment left the blended gross margin at only 26%, below Goldman Sachs' 32% forecast, although still above MiniMax's 18%. Gross profit in 1H26 was Rmb252mn, up 164% year over year but 56% below Goldman Sachs' forecast. R&D expenses were Rmb2.131bn, up 34% year over year and broadly in line with expectations, while adjusted net loss was Rmb1.964bn. Goldman Sachs believes that gross profit from the inference business can already partially fund model R&D investment, but is not yet sufficient to change the company's overall loss-making position. ARR was the most prominent growth metric in the report. The company's ARR exceeded US$1.6bn in August; based on the latest weekly run rate, the annualized level was approximately US$2bn, both on a gross basis. ARR rose from US$250mn in March to US$1bn in July and then to US$1.6bn in August, demonstrating a rapidly accelerating growth trajectory. The drivers included both usage and pricing: year-to-date MaaS Token consumption increased 40-fold, Coding Plan usage increased 23-fold, and the average API selling price rose 101%. Enterprise penetration also deepened significantly, with average daily usage by the top ten customers increasing 98-fold from the beginning of the year, and GLM becoming a major model provider globally for four of China's leading internet platforms. Management explained that the gap between ARR and 1H recognized revenue stemmed partly from the gross-versus-net accounting difference created by including channel revenue sharing in ARR, and partly from ARR reflecting the current run rate while accounting revenue has not yet fully captured the recent acceleration. Based on continued improvement in Token usage and pricing power, Goldman Sachs raised its 2H26 revenue forecast to approximately Rmb7.0bn from Rmb6.8bn. However, because 1H revenue was below expectations, it lowered its FY26 revenue forecast to approximately Rmb7.9bn from Rmb8.6bn. The end-2026 ARR forecast was raised to US$2.7bn from US$2.5bn, above Goldman Sachs' US$1.2bn forecast for MiniMax. Management does not prioritize maximizing total Token volume, instead focusing on higher-value Token demand, using coding as the primary entry point and expanding into Agent, Co-work, and autonomous AI workflows. As model capabilities improve, the company expects its commercialization model to gradually shift from API calls and subscriptions toward delivery based on task outcomes, while releasing a model approximately every two to three months to maintain the competitiveness of its open-source models. Accordingly, Goldman Sachs raised its Open Platform gross margin forecasts for 2H26 and FY26 to 25.3% and 25.2%, respectively, from 23.4% and 23.5%, primarily based on improved pricing power, enhanced inference efficiency following model releases, and operating leverage from greater Token scale. Regarding the model roadmap, management believes that improving intelligence still depends on scaling and requires joint optimization of pre-training scale, effective depth, training depth, and task environments. The company views GLM-5.2 and GLM-5.3 as evidence that post-training continues to offer high returns on investment, while GLM-5.3 Flash uses a new architecture to improve efficiency and expand the intelligence-cost frontier. The next stage will focus on larger foundation models, next-generation pre-training that has already commenced, longer native context windows, and native multimodal capabilities, ultimately progressing toward Fully Self Training, in which AI recursively participates in generating and improving the data, environments, and infrastructure required for future training. Compute supply and efficiency represent another core pillar supporting this roadmap. The company is expanding compute capacity through self-operated clusters, leased capacity, and cloud-service procurement. A cluster containing more than 100,000 domestically produced chips already supports large-scale inference, while inference cost per Token has declined 80% year to date. Despite continued constraints in the compute environment, the company continues to optimize its software, networking, and inference technology stacks, with the API revenue monetization multiple relative to each Rmb1 of compute spending improving 14-fold year over year. Goldman Sachs raised its 2H26 and FY26 R&D expense forecasts to Rmb4.3bn and Rmb6.4bn, respectively, from Rmb4.1bn and Rmb6.2bn, with approximately two-thirds related to compute investment. The previously completed HK$31bn equity placement will fund larger-scale model training and continuously growing compute requirements. Following the 1H results, Goldman Sachs lowered its 2026-28 revenue forecasts by 8%, 4%, and 2%, respectively, due to reduced on-premises deployment revenue estimates and broadly unchanged API revenue estimates. The new forecasts are Rmb7.921bn, Rmb23.713bn, and Rmb56.521bn, corresponding to year-over-year growth of 994%, 199%, and 138%. Because the 1H loss exceeded expectations, earnings forecasts for 2026-28 were reduced by 2%, 4%, and 2%, respectively, with adjusted net losses expected to be Rmb4.618bn, Rmb5.262bn, and Rmb1.381bn. The 12-month target price remains HK$1,610, based on a DCF valuation assuming a 12% WACC, 2% terminal growth rate, a 20% market share in 2030, and a 26% long-term adjusted EBIT margin in 2035. Bull- and bear-case valuations were slightly revised to HK$2,508 and HK$839, respectively. Goldman Sachs ultimately maintained its sector-relative Neutral rating, viewing rapid commercialization and long-term model potential as broadly balanced by near-term losses, competition, and investment risks.

Analysis framework

Goldman Sachs first compared the actual 1H26 results item by item with its own forecasts and market consensus, then explained the revenue variance through gross-versus-net accounting differences, ARR's forward-looking run-rate characteristics, and the cloud transition. It subsequently broke commercialization growth into Token usage, API pricing, and enterprise adoption, and assessed the gross margin and R&D investment trajectory in conjunction with the model release cadence, inference efficiency, and compute supply. Finally, it updated its 2026-28 revenue and earnings forecasts and tested the target price using DCF and bull-, base-, and bear-case scenarios.

Methodology notes

• Valuation methodologyDCF discounted cash flow

  DCF valuation

  The report discounts the company's future cash flows to present value, using a 12% WACC and 2% terminal growth rate, while incorporating long-term assumptions including a 20% market share in 2030 and a 26% adjusted EBIT margin in 2035, to derive a 12-month target price of HK$1,610.

• Industry/Sector analysis frameworkVolume-price decomposition

  Decomposition of Token usage and API pricing

  The report decomposes MaaS growth into Token consumption, Coding Plan usage, and the average API selling price, thereby determining that the ARR increase was driven not only by traffic expansion but also by stronger pricing power.

• Company fundamentals and financial framework

  Bridge between ARR run rate and accounting revenue

  The report uses monthly ARR and the latest weekly annualized run rate to observe real-time commercialization momentum, then explains the gap between ARR and 1H reported revenue through gross-versus-net accounting differences for channel revenue sharing and the timing lag in revenue recognition.

Asset mapping & comparison

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

• Z.AI Co. (02513.HK)

  The primary company covered in the report, whose growth depends on foundation-model capabilities, MaaS/API commercialization, enterprise adoption, and compute efficiency.

  Strengths

  ARR is rising rapidly, with simultaneous growth in Token usage and API pricing; API gross margin is improving; adoption by leading enterprise customers is deepening; domestically produced chip clusters and inference optimization have significantly reduced unit costs.

  Weaknesses

  On-premises deployment revenue declined significantly, while blended revenue and gross profit fell below Goldman Sachs' forecasts; R&D spending and adjusted net losses remain high, and near-term earnings visibility is limited.

  Comparison

  The 1H26 blended gross margin was 26%, above MiniMax's 18%; R&D expenses increased 34% year over year, below MiniMax's 139%; Goldman Sachs forecasts Z.AI's end-2026 ARR at US$2.7bn, above MiniMax's US$1.2bn.

  Risks

  Global foundation-model competition, high R&D spending, cash burn and self-funding capacity, as well as geopolitical risks arising from intensifying China-US technology competition.

• MiniMax Group

  A peer company used in the report to compare gross margins, R&D investment growth, and ARR prospects.

  Comparison

  The report lists its blended gross margin at 18% and R&D expense growth at 139% year over year, while Goldman Sachs forecasts its end-2026 ARR at US$1.2bn, using these figures as benchmarks for Z.AI's corresponding metrics.

Key data

• 1H26 revenueRmb954mnUp 400% year over year and 79% half over half, 47% below Goldman Sachs' forecast
• 1H26 Open Platform/API revenueRmb825mnUp 2,736% year over year and 412% half over half, 4% above consensus but 38% below Goldman Sachs' forecast
• 1H26 on-premises deployment revenueRmb129mnDown 20% year over year and 65% half over half, 73% below Goldman Sachs' forecast
• August ARROver US$1.6bnGross basis; approximately US$2bn annualized based on the latest weekly run rate
• ARR growth trajectoryUS$250mn in March, US$1bn in July, and US$1.6bn in AugustReflects the rapid increase in the commercialization run rate during the year
• MaaS usage and pricingToken consumption +40-fold, Coding Plan usage +23-fold, average API selling price +101%All changes are year to date
• Adoption by leading customersAverage daily usage by the top ten customers +98-foldCompared with the beginning of the year; GLM has become a major model provider for four of China's leading internet platforms
• API business gross margin24.6%Above Goldman Sachs' 24.0% forecast and 2H25's 22.4%
• Blended gross margin26%Below Goldman Sachs' 32% forecast, but above MiniMax's 18%
• 1H26 R&D expensesRmb2.131bnUp 34% year over year and broadly in line with Goldman Sachs' forecast
• 1H26 adjusted net lossRmb1.964bnGross profit from the inference business can partially support model R&D, but the company remains significantly loss-making
• End-2026 ARR forecastUS$2.7bnRaised from US$2.5bn; Goldman Sachs' forecast for MiniMax is US$1.2bn
• Inference efficiencyInference cost per Token declined 80%Year to date; the compute monetization multiple improved 14-fold year over year
• 2H26/FY26 R&D expense forecastsRmb4.3bn/Rmb6.4bnPreviously Rmb4.1bn/Rmb6.2bn, with approximately two-thirds allocated to compute investment
• 12-month target priceHK$1,610Unchanged; DCF uses a 12% WACC and 2% terminal growth rate
• Valuation scenariosBull case HK$2,508; bear case HK$839The base-case target price is HK$1,610

Impact & implications

The report believes that Z.AI's cloud transition is creating a disconnect between short-term accounting revenue and the real-time business run rate: contraction in on-premises deployment is weighing on current revenue, but ARR, Token usage, pricing, and enterprise adoption indicate a clear acceleration in MaaS commercialization. Pricing power, inference efficiency, and scale benefits are expected to support API margins, although investment in larger models and compute means that near-term losses and cash burn remain significant. By expanding into Agent, Co-work, and autonomous workflows, the company aims to gradually shift from usage-based fees toward a higher-value, task-outcome-oriented business model, but the target price still depends on sustained market-share gains and the realization of long-term margins.

Risks

• Upside risk: Model intelligence may be stronger than expected.
• Upside risk: Earnings visibility may improve faster than expected.
• Upside risk: The company may establish additional non-inference revenue-sharing streams.
• Upside risk: Overall commercialization capabilities may be stronger than expected.
• Downside risk: Competition in the global foundation-model industry may intensify.
• Downside risk: High R&D expenses may keep near-term earnings visibility limited.
• Downside risk: Cash burn and self-funding capacity may be weaker than expected.
• Downside risk: Intensifying China-US technology competition may have geopolitical implications.

What to watch

• Monitor whether ARR can continue advancing from over US$1.6bn in August toward the end-2026 forecast of US$2.7bn.
• Monitor whether ARR can convert into reported revenue as differences in gross-versus-net accounting and growth timing narrow.
• Track Token consumption, Coding Plan usage, API pricing, and average daily usage by leading customers.
• Track whether API gross margin can continue improving due to pricing power, inference efficiency, and scale benefits.
• Monitor rising 2H26 R&D expenses, earnings visibility, cash burn, and self-funding capacity.
• Monitor the model release cadence of approximately every two to three months, as well as progress in larger foundation models, long context windows, native multimodality, and Fully Self Training.
• Track compute expansion, cluster utilization, inference cost per Token, and the compute monetization multiple.
• Monitor the company's expansion from Coding into Agent, Co-work, and autonomous AI workflows, as well as its progress toward an outcome-based pricing model.