UBS is positive on China AI models around the H2 2026E themes of model capabilities, monetization, and Token ROI

Overview

This report focuses on three main themes for the China AI model industry in H2 2026E: improving model capabilities, expansion of model-layer monetization, and rising attention from enterprises and users to Token return on investment. UBS believes that after the listings of Zhipu and MiniMax, market discussion has expanded from pure model iteration to commercialization, competitive landscape, and valuation differences. Overall, the report is positive on Chinese models gaining global share through cost efficiency and application expansion, while also emphasizing that compute supply, geopolitical regulation, and competition remain key constraints.

Core views

The core views include: first, AI coding remains the clearest capability path for improving model intelligence, and SOTA models may consolidate their lead through a data flywheel, productization, and AI participation in model R&D. Second, the TAM for AI coding continues to expand from developer tools into knowledge-worker workflows and, combined with multimodal progress, drives model-layer monetization. Third, as enterprise adoption shifts from tokenmaxxing to ROI discipline, the pricing power of different models will further diverge, with SOTA models serving high-value scenarios while non-SOTA models are more likely to face commoditization pressure in mass-market Token usage.

Analysis framework

The report adopts a thematic research approach combined with company valuation updates, first analyzing industry drivers such as AI coding, productization, Token ROI, and multimodal commercialization, and then separately evaluating Zhipu and MiniMax in terms of model capabilities, ARR ramp-up, revenue forecasts, profitability pressure, and valuation multiples. Zhipu is primarily valued using P/S with SOTP as a cross-check, while MiniMax is valued using P/ARR.

Methodology notes

• Model capability analysisRL-based post-training

  Reinforcement-learning post-training

  The report believes AI coding outcomes can be validated through code execution, unit tests, and automated graders, and when combined with synthetic data and real-task feedback, this makes it a highly feasible area for continued improvement in model intelligence.

• Commercialization analysisData flywheel

  Data flywheel

  User interactions, real workflows, project context, and preferences in coding agent products can feed back into model improvement and enhance user stickiness.

• Demand analysisToken ROI discipline

  Token return-on-investment discipline

  As enterprises shift from maximizing Token usage to optimizing return per Token, they will place greater emphasis on cost efficiency and task output, benefiting Chinese models with structural cost advantages.

• Valuation analysisP/S, P/ARR and SOTP

  Price-to-sales, P/ARR, and sum-of-the-parts valuation

  Zhipu uses 160x 2026E P/S with SOTP as a cross-check; MiniMax uses 20x 2026E P/ARR, reflecting the valuation gap between SOTA and non-SOTA models.

Asset mapping & comparison

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

• Knowledge Atlas Technology (Zhipu; Z.ai; 2513.HK)

  Core beneficiary target

  Strengths

  It has domestic leadership in SOTA coding models, GLM-5.2 has gained global recognition, the ramp-up in open platform and API revenue is clear, and infrastructure optimization such as LayerSplit and ZCube improves throughput and cost.

  Weaknesses

  It still faces pressure from compute supply, commercialization execution, replacement by self-developed models from key-account customers, and long-term losses.

  Comparison

  The report compares Zhipu's model development and monetization strategy to Anthropic and assigns it a higher P/ARR or P/S valuation premium relative to MiniMax.

  Risks

  Geopolitical and regulatory uncertainty, constraints on compute infrastructure, intensifying competition, data security, and AI governance risks.

• MiniMax (0100.HK)

  Long-term beneficiary but near-term valuation discount target

  Strengths

  It has advantages in compute access and R&D efficiency, products cover both 2B and 2C, and it has differentiated potential in multimodal areas such as Hailuo.

  Weaknesses

  Its model capabilities still need time to catch up with frontier SOTA models, and heavy R&D investment will weigh on near-term profitability and cash flow.

  Comparison

  MiniMax's valuation multiple is significantly lower than Zhipu's, reflecting the market's recognition of the difference in pricing power between SOTA and non-SOTA coding models.

  Risks

  Commercialization delays, margin pressure, model training data risks, user-generated content governance risks, and disruption from key service providers.

• China AI model providers

  Potential beneficiaries at the industry level

  Strengths

  They have better structural cost efficiency and possess Token ROI advantages in high-frequency, repetitive coding and agentic reasoning tasks.

  Weaknesses

  Competition from global frontier models is intense, and compute supply may still limit the pace of ARR ramp-up.

  Comparison

  Relative to U.S. frontier models, Chinese models may gain share through cost efficiency and overseas application integration, but high-end scenarios still depend on the model intelligence gap.

  Risks

  Export restrictions, regulatory environment, rapid technological change, and price competition.

Key data

• Zhipu target priceHK$2,200Raised from the previous HK$1,160, based on 2026E revenue of Rmb5.5bn and 160x P/S.
• MiniMax target priceHK$500Lowered from the previous HK$1,000, based on Dec 2026E ARR of US$1.0bn and 20x P/ARR.
• Zhipu Dec 2026E ARRUS$1.5bnAbove the previous internal management target of US$1bn, reflecting AI coding leadership and compute optimization.
• MiniMax Dec 2026E ARRapproximately US$1.0bnThe report expects it to be broadly in line with management's internal target.
• Zhipu 2025-27E revenue CAGR332%The open platform revenue share is expected to rise from 26% in 2025 to 87% in 2027.
• MiniMax 2025-27E revenue CAGR258%The revenue mix is expected to shift more toward 2B services.
• OpenAI Codex non-developer individual user growth137xAs of early June 2026, compared with August 2025.
• OpenAI Codex non-developer organization user growth189xShows AI coding agents expanding from developers to a broader set of knowledge workers.
• Kuaishou Kling ARRUS$500mMay 2026 data, used in the report as an example of monetization progress in video generation by Chinese models.
• Zhipu GLM-5.2 capability1mn-token context windowThe report says it has further narrowed the gap with the world's top coding models while maintaining competitive pricing.

Impact & implications

The investment implication is that the market may continue to assign a higher valuation premium to SOTA coding capabilities, with Zhipu gaining stronger pricing power due to GLM-5.2, AI coding leadership, and open-platform ARR ramp-up; although MiniMax faces a valuation discount in the near term, if future model iteration, multimodal products, and commercialization are delivered, it could also catch up and see stock-price catalysts. For the industry, Token ROI discipline will drive demand segmentation, with high-value scenarios concentrated in leading models, while mass-market Token usage is more likely to see price competition.

Risks

• Evolution of the competitive landscape and intensifying industry competition.
• Rapid changes in technology trends, user demand, and preferences.
• Uncertainty in model commercialization, with ARR ramp-up potentially below expectations.
• Constraints on compute supply and critical computing infrastructure.
• Geopolitical, export-control, and regulatory uncertainty.
• Data security, model training data, and AI governance risks.
• Heavy R&D investment, hardware costs, and customer acquisition costs may pressure margins.
• Execution risk in international expansion and policy risk in overseas markets.

What to watch

• Subsequent model upgrades at Zhipu and changes in the global ranking of the GLM series in coding and agentic capabilities.
• Progress of the MiniMax M series, Hailuo video generation model, and multimodal commercialization.
• 2026 interim results, ARR disclosures, and revenue mix changes for Zhipu and MiniMax.
• The pace at which enterprises shift from tokenmaxxing to Token ROI optimization, as well as trends in API pricing and usage volume.
• The number of integrations of Chinese models into overseas applications and platforms, and their actual share of Token usage.
• Progress in compute supply, adaptation of domestic AI chips, and inference cost optimization.
• Capital-market catalysts such as Southbound Stock Connect inclusion and A-share listing.