China AI model ARR expectations raised; performance-to-price ratio, multimodality, and agent entry points become key competitive themes

Overview

The report tracks the latest developments in Chinese AI foundation models, multimodal generation, agent applications, cloud infrastructure, and consumer internet applications. Goldman Sachs believes Chinese models are continuing to expand usage share on the back of cost efficiency, coding capabilities, and open-weight strategies, with commercialization progressing faster than previously expected; therefore, it raises revenue forecasts for the industry and key companies. At the same time, competition among ultra-large-parameter models, pricing pressure on low-end APIs, sustained R&D investment, and potential cross-border model regulation will intensify industry divergence.

Core views

First, combined ARR of Chinese AI models is expected to reach US$13bn at end-2026, above the previous US$10bn forecast. Second, the competitive focus has expanded from simply reducing costs to high-end coding capability, multimodal capability, agent entry points, and pricing power. Third, MiniMax H3's full-modal, open-weight, and low-price strategy is expected to drive ARR growth and valuation recovery. Fourth, workspace products such as Tencent WorkBuddy and Alibaba QwenWork are forming a closed loop of model distribution, data feedback, and monetization. Fifth, improved domestic chip and computing power supply is expected to support higher capital expenditure by cloud vendors, with cloud and data centers remaining the preferred subsector.

Analysis framework

The report combines model arena rankings, OpenRouter token usage and estimated revenue, API prices, multimodal product capabilities, application visits and time spent, company ARR disclosures, earnings forecasts, and cloud vendor capital expenditure data to conduct a horizontal comparison of model vendors' performance-to-price ratio, commercialization capability, financial strength, and ecosystem entry points, and accordingly adjusts forecasts and valuations for Z.AI and MiniMax for 2026 to 2028.

Methodology notes

• Competitive analysisAI model competitive positioning framework

  Comprehensively evaluate performance, price, cost efficiency, financial strength, ecosystem distribution, and commercialization capability

  The report believes pricing power, cost efficiency, and financial strength are the three key indicators for evaluating Chinese AI model companies, and further examines performance in coding, multimodal, and agent products.

• Revenue forecastARR projection

  Estimate annual recurring revenue based on API demand, token growth, product releases, and company disclosures

  Industry and company ARR forecasts reflect the current revenue run rate and are not realized full-year revenue; private deployment of open-weight models may also cause public data to underestimate actual usage scale.

• Market trackingOpenRouter token and price tracking

  Observe model adoption through token consumption share, task mix, and price changes

  This framework helps identify share changes of high cost-performance models, but platform data cannot fully represent all direct-sales APIs, private deployments, and usage within China.

• Application trackingLeading application engagement tracking

  Assess application trends through time spent, monthly active users, daily active users, visits, and download share

  The report tracks leading Chinese mobile applications, consumer-grade AI applications, and desktop agent products to measure user adoption, competitive landscape, and potential commercialization opportunities.

• Capital expenditure analysisCapital expenditure to cloud revenue conversion framework

  Compare rolling capital expenditure lagged by one quarter with rolling incremental cloud revenue

  The report uses Alibaba, Tencent, and overseas cloud vendors as references to assess AI capital expenditure intensity, cash flow capacity, and future cloud revenue conversion efficiency.

Asset mapping & comparison

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

• MiniMax Group (0100.HK)

  Key recommended AI model company, Buy rating maintained

  Strengths

  H3 offers full-modal generation, open weights, significant cost efficiency, and stronger commercial production capabilities; API business is growing rapidly.

  Weaknesses

  Still loss-making, with high R&D and other operating expenses, and model revenue scale still below Zhipu.

  Comparison

  H3 pricing is approximately 30% to 50% of peers, and its video model ranking has improved significantly, but some image-to-video capabilities still lag Seedance.

  Risks

  Model iteration below expectations, price competition, computing power constraints, R&D spending exceeding expectations, and slower commercialization.

• Z.AI Co. (2513.HK)

  Chinese foundation model company, Neutral rating maintained

  Strengths

  ARR growth has been strong so far in 2026, the GLM series and API business have scale advantages, and the company continues to invest in domestic computing power infrastructure.

  Weaknesses

  Training costs and R&D investment are high, short-term losses are significant, and long-term revenue forecasts are affected by intensifying competition.

  Comparison

  Mid-2026 ARR run rate leads MiniMax, but the company will need to address a wave of releases of larger-parameter and high-end coding models in the future.

  Risks

  Target price cut, high-end model competition, sustained losses, uncertainty over R&D returns, and regulatory changes.

• Alibaba

  Beneficiary of AI cloud, foundation models, and workspace applications

  Strengths

  Has ample funding, cloud infrastructure, the Qwen model ecosystem, and QwenWork integration capability; domestic supernodes have been adapted to Qwen3.8 Max.

  Weaknesses

  Integration of models and product lines is still underway, and capital expenditure conversion efficiency requires continued verification.

  Comparison

  Superior to independent model labs in resources, cloud distribution, and rapid iteration, but faces competition from ByteDance, Tencent, and open-source models.

  Risks

  Capital expenditure returns, price competition, product integration execution, and model share volatility.

• Tencent

  Beneficiary of cloud infrastructure, models, and agent workspace

  Strengths

  WorkBuddy has advantages in ecosystem distribution, security governance, and multi-model support, and can monetize both proprietary tokens and third-party APIs.

  Weaknesses

  Product credit quota metrics are not directly comparable with peers, and commercialization effectiveness still needs to be validated.

  Comparison

  WorkBuddy leads Chinese desktop agent products with a 34% share and has stronger ecosystem and cross-selling capabilities than independent labs.

  Risks

  Product competition, API pricing pressure, capital expenditure returns, and insufficient user paid conversion.

• GDS, VNET, Kingsoft Cloud

  Infrastructure beneficiaries of AI token demand and rising cloud capital expenditure

  Strengths

  Agentic AI drives demand for computing power, cloud services, and data centers, with room for improvement in cloud prices and margins.

  Weaknesses

  Business is sensitive to capital expenditure cycles, customer utilization, and financing conditions.

  Comparison

  Compared with model vendors, their benefit path is more oriented toward infrastructure demand growth, but they lack direct pricing power at the model layer.

  Risks

  Capital expenditure slowdown, oversupply, utilization below expectations, financing costs, and cloud price competition.

Key data

• Forecast for combined ARR of Chinese AI models at end-2026US$13bnPrevious forecast was US$10bn.
• Forecast for Zhipu API and MiniMax Group ARR at end-2026US$2.5bn; US$1bnReflects stronger demand and faster ARR ramp-up.
• Z.AI 2026 to 2028 revenue forecast revisions+35%; +4%; -2%Earnings forecasts for the same period are revised by -3%, -3%, and -1%, respectively.
• MiniMax 2026 to 2028 revenue forecast revisionsRaised by 22% to 64%Mainly driven by the rising share of the rapidly growing API business.
• Target price adjustmentsZ.AI lowered by 14%; MiniMax lowered by 7%Ratings maintained at Neutral and Buy, respectively.
• Z.AI first-half 2026 forecastRevenue Rmb1.8bn; adjusted net loss Rmb1.8bnTraining costs are expected to be approximately Rmb2.2bn.
• MiniMax first-half 2026 forecastRevenue US$120mn; adjusted net loss US$168mnR&D expenses are expected to be approximately US$0.2bn.
• MiniMax H3 pricingApproximately 30% to 50% of peers2K per-second pricing is approximately one-third of peers, and 768P is approximately one-half of peers.
• MiniMax H3 generation specificationsUp to 15 seconds, 24 FPS, and up to 12 reference filesSupports native stereo audio, up to 9 images, 3 videos, 3 audio clips, and prompts of up to 7000 characters.
• Total time spent on China's top 400 mobile applicationsUp 9% YoY in June 2026YoY growth was 5% in June 2025.
• Total engagement of China AIGC applicationsUp 2% MoM in June 2026During the same period, time spent on Douyin's main app and Lite version increased 27% and 30% YoY, respectively.
• ByteDance top application time-spent share33%Up approximately 6 percentage points from two years ago.
• Tencent WorkBuddy desktop agent share34%Ranked first among Chinese desktop AI agent products by visits in June 2026.
• Low-price agent model API price rangeUS$0.1 to US$0.2/million tokensThe report expects prices and gross margins in this range to remain under pressure in the second half of 2026.

Impact & implications

At the industry level, improvements in model performance, API price reductions, and agent demand will expand token consumption, but will also compress gross margins for low-end models and accelerate the elimination of vendors lacking funding and ecosystem entry points. At the company level, MiniMax is expected to improve ARR and valuation with H3 and subsequent M3 series, while Z.AI, despite benefiting from short-term ARR growth, still faces competition from high-end coding models and pressure from R&D investment. At the industry-chain level, improved domestic computing power supply and ultra-large model scaling will support demand for cloud services, data centers, and domestic chips, potentially benefiting Alibaba, Tencent, and related infrastructure operators.

Risks

• Intensive releases of high-end coding and agent models in the second half of 2026 may lead to rapid user switching and market share volatility.
• Low-end API price competition may continue to depress model vendors' gross margins.
• Training and inference demand for ultra-large-parameter models may exacerbate computing power shortages and push up R&D costs.
• AI product adoption, subscription payments, and enterprise commercialization may be slower than expected.
• Overseas downloads of Chinese model weights and cross-border transfers of training data may face stricter regulation.
• Increased capital expenditure may not translate into cloud revenue and profit as expected.
• Changes in open-weight license terms may affect developer adoption and revenue recognition.
• Reduced application subsidies may lead to normalization of traffic and active user growth.

What to watch

• Performance, pricing, and release timing of Zhipu GLM-5.5, subsequent MiniMax M3 versions, and M3 Pro.
• Competitive landscape for high-end coding models with 1 trillion to 5 trillion parameters in the second half of 2026.
• MiniMax H3 open-weight implementation, API usage, usable output rate, and ARR conversion.
• Visits, subscriptions, and enterprise customer growth for agent workspaces such as WorkBuddy, QwenWork, and ZCode.
• Changes in OpenRouter Chinese model token share, coding task share, and estimated revenue.
• Capital expenditure by Chinese hyperscale cloud providers and progress in domestic chip deployment in the second half of 2026.
• Z.AI and MiniMax first-half 2026 revenue, losses, R&D expenses, and cash burn.
• Regulatory policies related to model weight downloads and cross-border transfer of training data.
• Daily active users, monthly active users, download share, and time spent for China's consumer-grade AIGC applications.