UBS sees three key H226E themes for China AI models: coding capability, commercialization, and token ROI

Overview

This UBS report focuses on the China AI model investment theme in the second half of 2026, centered on model capability uplift, commercialization expansion, and token ROI discipline. UBS believes that after Zhipu and MiniMax listed, market discussion has moved beyond pure model iteration to monetization, competitive dynamics, and valuation divergence.

Core views

UBS believes AI coding remains the clearest path for model capability uplift because code execution, unit testing, and automated grading make outcomes more verifiable and suitable for post-RL training. China AI labs are launching coding-agent products and harnesses that can accumulate user project context, skills, and preferences, forming a data flywheel and stronger user stickiness. On commercialization, AI coding is expanding from developer tools to broader knowledge-worker scenarios, alongside progress in monetizing multimodal models. As enterprises shift from tokenmaxxing to token optimization, Chinese models, with structural cost efficiency and ROI advantages in high-frequency repetitive reasoning scenarios, have global share capture potential.

Analysis framework

The report compares Zhipu and MiniMax across model capability trajectory, productization progress, TAM expansion, enterprise adoption patterns, ARR forecasts, and P/S versus P/ARR valuation frameworks. UBS positions Zhipu as the SOTA-adjacent coding model leader with higher valuation premium, while MiniMax has computing access and R&D efficiency advantages, but its model performance still needs to prove whether it can catch up with frontier SOTA.

Methodology notes

• 模型能力评估AI coding and post-RL training

  Treating AI coding as a verifiable training setting for model capability improvement

  Coding tasks can be verified through execution, unit tests, and automated scoring; together with synthetic data and real task feedback from coding agents, this supports continuous expansion of model capability boundaries.

• 商业化评估Token ROI discipline

  Shifting from tokenmaxxing to token optimization

  Enterprises are no longer only trying to maximize token usage, but are focusing more on business output per token and ROI, which may lead to further separation in pricing power and demand tiering between SOTA models and other models.

• Valuation methodsCross-checking P/S, P/ARR, and SOTP

  Zhipu uses P/S with SOTP cross-checks, while MiniMax uses P/ARR

  UBS applies 160x 2026E P/S to Zhipu and cross-checks with open-platform P/ARR and on-premise deployment P/S using SOTP; for MiniMax, UBS applies 20x 2026E P/ARR.

Asset mapping & comparison

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

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

  Core beneficiary

  Strengths

  SOTA coding model leadership is strengthening, GLM-5.2 has gained global recognition, and the potential for open platform and API revenue growth is strong, with inference infrastructure optimization improving cost efficiency.

  Weaknesses

  Still constrained by compute availability, competitive pressure, and geopolitical regulatory uncertainty; on-premise deployment business may be a lower priority.

  Comparison

  Compared with MiniMax, UBS believes Zhipu is stronger in SOTA coding capability and commercialization clarity, and therefore assigns higher valuation multiples and premium.

  Risks

  Compute supply constraints, commercialization delays, in-house model substitution by key accounts, data security, and AI governance regulatory risk.

• MiniMax (0100.HK)

  Long-term beneficiary but valuation-discounted name

  Strengths

  Advantages in computing access, R&D efficiency, multimodal positioning, and global commercialization strategy, with faster growth in API and Token Plan revenue.

  Weaknesses

  Model capability still needs to catch up with frontier SOTA; sustained high R&D spending suppresses near-term profitability and cash flow.

  Comparison

  Compared with Zhipu, MiniMax’s valuation multiple is materially lower, reflecting market discount for weaker commercialization and pricing power of a non-SOTA model.

  Risks

  Intensifying competition, commercialization uncertainty, margin pressure, user-generated content governance risk, and disruptions from critical service providers.

Key data

• Zhipu target priceHK$2,200Raised from HK$1,160, based on 2026E revenue of Rmb5.5bn and 160x P/S.
• MiniMax target priceHK$500Lowered from HK$1,000, based on 2026E ARR of US$1.0bn and 20x P/ARR.
• Zhipu 2026E ARRUS$1.5bnExpected to be reached by December 2026, above the management's prior internal target of US$1bn.
• MiniMax 2026E ARRabout US$1.0bnExpected to be reached by December 2026, broadly in line with management's internal target.
• Zhipu 2025-27E revenue CAGR332%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 further toward B2B services.
• Zhipu 2026E revenue forecast revision+71%Mainly reflects a more constructive view on ARR ramp speed.
• MiniMax 2026E revenue forecast revision+108%Reflects faster ARR acceleration, with 2026E revenue forecast at US$461mn.
• Kuaishou Kling ARRUS$500mThe report cites its May 2026 video-generation commercialization progress as a multimodal monetization example.
• OpenAI Codex non-developer consumer growth137xFrom August 2025 to early June 2026, suggesting AI coding adoption is expanding beyond developers.

Impact & implications

The investment implication is that China AI model companies may gain global share in high-frequency, repetitive, agentic, and coding workflows through cost efficiency, while capital markets will more strictly distinguish long-term pricing power between SOTA and non-SOTA coding models. Zhipu is expected to benefit more in the near term from model leadership, GLM-5.2 credibility, and a clearer commercialization path; MiniMax's key variables are future model iteration, multimodal monetization, and ARR realization.

Risks

• The AI model competitive landscape is evolving rapidly and becoming more intense.
• Technology trends, user needs, and preferences can change quickly, potentially eroding existing product advantages.
• Commercialization cadence and ARR ramp carry uncertainty.
• Compute supply and availability of advanced hardware may constrain model upgrades and revenue growth.
• Geopolitics, export restrictions, and regulatory environments may affect global expansion.
• Data security, AI governance, and model-training-data-related risks.
• Rising costs of traffic acquisition, content, and brand promotion may pressure margins.

What to watch

• Subsequent upgrades to the Zhipu GLM series and changes in its ranking among global coding models.
• Release cadence of MiniMax M-series, Hailuo, and the potential 2.7tn-parameter open-weight model.
• Whether enterprise AI adoption continues to shift from tokenmaxxing to ROI-driven usage.
• Interim 2026 performance, ARR disclosures, and margin changes for Zhipu and MiniMax.
• Capital-market catalysts such as southbound capital inclusion and A-share listing.
• The number of overseas applications and platform integrations of China models, and penetration of high-frequency agentic workflows.
• Progress in compute supply, inference-cost optimization, and adaptation with domestic AI chips.