Capabilities and costs reshape China's large-model competitive frontier; Zhipu is preferred, while MiniMax awaits M3.1 validation

Overview

The report focuses on the shift in China's large-model competition from pure price competition toward a capability-cost Pareto frontier. JPMorgan believes that models face greater commercialization challenges if they cannot offer stronger capabilities at the same price, or comparable capabilities at a lower price. As the industry's capability frontier continues to evolve rapidly, the research team prefers Zhipu, which has sustained capability iteration and coverage of high-value workflows; it maintains a Neutral rating on MiniMax, believing it still needs M3.1 and multimodal products to demonstrate stronger independent competitiveness.

Core views

Both the capability frontier and cost efficiency can create commercial value, but amid rapid model capability upgrades, sustained capability leadership is assigned greater value.With the base model unchanged, GLM-5.3 improves coding and agent capabilities through heavier post-training; if independently validated, it could support Zhipu's adoption and retention.DeepSeek's price increase narrows MiniMax's cost disadvantage in substitutable workloads, but this improvement stems from a competitor's pricing decision and is less durable.Hailuo H3 strengthens MiniMax's multimodal positioning, but whether independent vendors can capture sufficient revenue and profit against integrated platforms such as ByteDance and Kuaishou remains the central question.

Analysis framework

The report uses a capability-cost Pareto framework for relative competitive positioning: cost is primarily measured by blended token pricing, while capability references benchmark tests and real-world product performance. The report explicitly notes that the framework is suitable for comparing relative positions, but not for precisely calculating the return on investment for an individual customer, because token mix, caching, workloads, and production-environment performance all affect actual economics.

Methodology notes

• Competitive analysisCapability-cost Pareto frontier

  A model sits on the competitive frontier when no competitor can provide stronger capabilities at the same or lower price, or comparable capabilities at a lower price.

  This framework assesses a model's relative commercial position across capability and price; models inside the frontier are more likely to be squeezed by substitutes that are stronger or cheaper.

• Valuation methodsDiscounted P/E valuation

  Based on expected 2030 P/E multiples and discounted using a 15% weighted average cost of capital.

  Target prices for both Zhipu and MiniMax are based on 20x expected 2030 P/E and discounted; the valuation premium reflects expected revenue CAGR exceeding 100% from 2026 to 2030.

Asset mapping & comparison

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

• Z AI Co Ltd - H（2513.HK）

  Core beneficiary

  Strengths

  GLM-5.3 improves coding and agent capabilities; a sustained record of launching domestic frontier models; closer to the capability frontier, with strategic flexibility to extend into cost-efficiency optimization.

  Weaknesses

  Commercialization and paid conversion require further validation; future competitive advantage depends on sustained model iteration and cannot rely solely on GLM-5.3.

  Comparison

  Compared with MiniMax, Zhipu's competitive improvement is driven more by internal capability gains, leading the report to assign it a higher rating.

  Risks

  Export controls, geopolitical and entity-list risks; intensified competition; sustained high R&D investment; customer adoption uncertainty; dependence on computing infrastructure and external suppliers.

• MiniMax Group Inc - H（0100.HK）

  Neutral watchlist name

  Strengths

  Offers a combination of large models and multimodal capabilities, 2C and 2B product coverage, international expansion, and capabilities across text, audio, images, video, and productivity use cases; Hailuo H3 provides a second growth-curve option.

  Weaknesses

  The current M3 has not established a clear advantage in capability or value for money; improvements partly depend on DeepSeek's price increases; independent vendors have weaker value-capture capabilities than integrated platforms.

  Comparison

  Compared with Zhipu, MiniMax remains in a model-capability catch-up phase; compared with platforms such as ByteDance and Kuaishou, it lacks synergistic monetization across content, distribution, advertising, and user-engagement channels.

  Risks

  Litigation proceedings with U.S. film studios; intensified competition; high R&D investment and profitability pressure; commercialization and customer-adoption uncertainty; dependence on computing power and external supply.

• DeepSeek

  Industry competitive benchmark

  Strengths

  Model architecture and system-level efficiency support lower pricing, including MoE, attention design, KV cache, quantization, and inference-serving optimization.

  Weaknesses

  API price adjustments may alter short-term relative pricing positions, but do not directly prove that other vendors have established structural cost advantages.

  Comparison

  The report views it as China's most representative credible cost leader and as continuing to set the cost benchmark for peers such as MiniMax.

  Risks

  Changes in its pricing and model iterations could rapidly reshape the competitive landscape for peers.

Key data

• Zhipu rating and target priceOverweight; HK$1,800Target price raised from HK$1,600; the report-listed share price on 2026-08-14 was HK$1,270.
• MiniMax rating and target priceNeutral; HK$260Target price raised from HK$160; the report-listed share price on 2026-08-14 was HK$329.
• Zhipu revenue forecast revisionRaised 6% to 10% for 2026 through 2030Reflects GLM-5.3 capability improvements and improved commercialization expectations.
• MiniMax revenue forecast revisionRaised 11% to 21% for 2027 through 2030Reflects more moderate large-model price competition and the multimodal option value of Hailuo H3.
• Implied ARR valuationZhipu approximately 20x 2027 US$5bn ARR; MiniMax approximately 12x US$1bn ARR at end-2026This reflects the report's presentation of ARR multiples implied by the target prices.
• DeepSeek API price increase effective date2026-08-17The report states that the announcement date was 2026-08-13 and that certain V4 API prices increased significantly.

Impact & implications

For investors, the report shifts the valuation focus from short-term usage volumes or one-off model releases to whether vendors can repeatedly remain at the capability or cost frontier. The positive case for Zhipu is that its capability improvement is driven by its own model iteration and can further improve value for money through subsequent inference optimization; MiniMax's target-price increase primarily reflects easing external price competition and multimodal option value, which is not yet sufficient to support a more positive rating.

Risks

• Large-model capability iteration and pricing are changing rapidly, and current competitive positions may quickly become obsolete.
• GLM-5.3's capability improvements still require broader independent validation and real-user feedback.
• If MiniMax's M3.1 improves only marginally and remains inside the Pareto frontier, the long-term investment thesis may lack material improvement.
• Growth in multimodal demand does not necessarily translate into revenue, profit, or pricing power for independent model vendors.
• Regulation, geopolitics, computing-power supply, R&D investment, and customer commercialization could all affect the path to profitability.

What to watch

• Independent evaluations, user adoption, and paid conversion for GLM-5.3 in real-world use cases such as coding and agents.
• Whether Zhipu can continue launching frontier models and improve inference costs while maintaining capabilities.
• Whether MiniMax M3.1's capabilities, API pricing, and value-for-money positioning improve materially.
• Hailuo H3's paid usage, pricing durability, unit economics, and multimodal frontier performance.
• Whether DeepSeek's subsequent API pricing and model releases again compress peers' profit pools.
• Competitive and monetization progress by integrated platforms such as ByteDance and Kuaishou in the multimodal content ecosystem.