HSBC initiates coverage of MiniMax: a global foundation-model pure play, but valuation is already fair

Overview

This report is HSBC's initial coverage of MiniMax (0100.HK). HSBC positions MiniMax as a pure-play foundation-model company for global markets, with multimodal capabilities across text, image, audio, and video, monetized through a diversified mix of Open Platform, generative media, Agent, and companionship products. The report acknowledges MiniMax's strengths in model capability, low pricing, inference and training efficiency, global user reach, and organizational efficiency, but argues that after the sharp share price rise since IPO, valuation is now broadly reasonable, so HSBC assigns a Hold rating.

Core views

The core views are: first, 2026 is a key year for faster penetration of foundation models and AI agents, with global token usage expected to grow exponentially and MiniMax set to benefit. Second, MiniMax's revenue is highly global, with 73% of revenue coming from overseas in 2025, and it serves individual users in more than 200 countries and regions as well as enterprise customers and developers in more than 100 countries and regions. Third, M2.7 offers performance close to leading models at a meaningfully lower price, and together with MoE, Linear Attention, CISPO, and self-developed AI infrastructure, supports pricing competitiveness and margin improvement. Fourth, despite strong fundamentals, the current valuation of around 186x 2026e PS already largely reflects scarcity and growth, and HSBC prefers Zhipu, which trades at a lower valuation.

Analysis framework

The report combines fundamental forecasts, assessment of technical competitiveness, business-segment growth decomposition, P/ARR comparable valuation, and a 10-year DCF valuation. Revenue forecasts are built around token usage growth, API demand, overseas user adoption, cloud-service partnerships, and monetization of multimodal applications; on the cost side, it focuses on training costs, inference costs, compute utilization, and the impact of organizational automation on the profitability path. Valuation is anchored by DCF, with OpenAI and Anthropic's historical P/ARR relationship used as a cross-check, alongside bull and bear scenario stress tests.

Methodology notes

• Valuation methodsDCF

  10-year discounted cash flow model

  The HKD1,000 target price comes from a 10-year DCF model, with key assumptions including 11.4% WACC and 3% perpetual growth, to reflect the company's long-term growth potential.

• Valuation methodsP/ARR comparable regression

  Comparable valuation based on OpenAI and Anthropic's historical ARR and valuation relationship

  The report uses OpenAI and Anthropic's ARR trajectories and implied P/ARR to estimate MiniMax's value, arriving at a cross-check result of about USD36bn or HKD887 per share, but notes the sample is limited and does not account for private-company liquidity discounts.

• TechnologyMixture-of-Experts

  MoE architecture

  MoE improves inference efficiency and scalability by activating only the expert networks most relevant to the input, helping lower compute cost per token.

• TechnologyLinear Attention

  Linear attention mechanism

  Linear Attention reduces the compute and memory burden of long-context tasks, allowing the model to process long text, document analysis, and software development scenarios more efficiently.

• TechnologyCISPO

  Clipped IS-weight Policy Optimization

  CISPO improves training efficiency by limiting the impact of a single sample's weight rather than discarding samples, preserving learning signals while maintaining training stability.

Asset mapping & comparison

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

• MiniMax (0100.HK)

  covered company

  Strengths

  High share of global revenue, strong multimodal model capability, standout M2.7 value for money, flat organizational structure, and clear improvements in inference and training efficiency.

  Weaknesses

  Still loss-making, high training and inference costs, 2026e-27e EPS below consensus, and may need additional funding in the future.

  Comparison

  Versus leading models such as OpenAI, Anthropic, Gemini, and Claude, MiniMax is cheaper but nearly as capable; versus Zhipu, HSBC believes MiniMax has a higher valuation and faces more IP and content risk in video generation and companionship businesses.

  Risks

  New model launches by competitors, IPOs by rivals reducing scarcity premium, compute bottlenecks, lock-up pressure, cash burn, geopolitical risk, intellectual property risk, and the risk of inappropriate content generation.

• Zhipu / Knowledge Atlas (2513 HK)

  peer comparison

  Strengths

  Its valuation is lower than MiniMax's, API pricing power is relatively strong, Tsinghua-linked talent support helps improve model capability, and coding ability is leading.

  Weaknesses

  It is also in a highly competitive, high-investment foundation-model arena.

  Comparison

  HSBC prefers Zhipu because it requires a lower valuation, API price increases have not clearly suppressed paid token growth, and ARR growth remains strong.

  Risks

  Model competition, commercialization pace, and cost investment pressure.

• OpenAI / Anthropic

  overseas peers and valuation reference

  Strengths

  Leading global foundation-model companies, serving as valuation and technology benchmarks for P/ARR.

  Weaknesses

  If their IPOs or new fundraisings change market scarcity, they could compress valuation premiums for listed AI foundation-model companies such as MiniMax.

  Comparison

  MiniMax's valuation is cross-checked using the historical P/ARR relationship of OpenAI and Anthropic; the report believes MiniMax's current valuation is already fair.

  Risks

  Competitor listings, model iterations, and pricing changes could affect MiniMax's market share and valuation multiple.

Key data

• Rating and target priceHold; HKD1,000Implies about 6% downside versus the current price of HKD1,060.
• Current valuationabout 186x 2026e PSHSBC considers it fair versus leading foundation-model companies.
• DCF assumptionsWACC 11.4%; perpetual growth rate 3%The target price is based on a 10-year DCF model.
• Revenue forecastUSD228m / USD682m / USD1.7bnThese correspond to 2026e / 2027e / 2028e revenue assumptions.
• Revenue CAGR179% in 2025-28eDriven by rapid token usage growth and gradual price increases.
• Overseas revenue mix73%Share of revenue from markets outside mainland China in 2025.
• User and customer reach236m individual users; 214k enterprise customers and developersAs of December 31, 2025, individual users covered more than 200 countries and regions, while enterprise customers and developers covered more than 100 countries and regions.
• M2 model usage growthAverage daily tokens in February 2026 more than 6x December 2025Token consumption under the Token Plan or Coding Plan more than 10x over the same period.
• ARRUSD150m in February 2026; estimated USD310m in December 2026ARR in February 2026 was 50% higher than in December 2025.
• Model value for moneyAbout 8% of the price of the three leading models; about 90% of their performanceRelative to Gemini 3.1 Pro Preview, GPT-5.4, and Claude Opus 4.6.
• Cost efficiencyM2 inference cost per 1m tokens down more than 50%Comparison between February 2026 and December 2025.
• Profitability timingExpected to turn profitable in 2029eTraining costs as a percentage of revenue are expected to fall from 241% in 2025 to 41% in 2028e, while inference costs as a percentage of revenue decline from 69% to 55%.
• Bull caseHKD1,837; about 73% upsideAssumes 2026e global LLM revenue market share is 0.1 percentage point above base case, and 2026e PS is 225x.
• Bear caseHKD67; about 94% downsideAssumes 2026e global LLM revenue market share is 0.1 percentage point below base case, and 2026e PS is 21x.

Impact & implications

The report's investment implication is that MiniMax has rare global, multimodal, and high-growth attributes among AI foundation-model companies, and will benefit over the long term from token demand expansion driven by AI agents, coding, office applications, and video generation; however, much of the growth expectation is already reflected in the share price in the near term, and competition, compute constraints, lock-up expiries, and financing pressure may bring significant volatility. For investors, it is more suitable at present to track it as a high-growth AI infrastructure name rather than add aggressively when valuation is already full.

Risks

• New model launches by competitors could dilute MiniMax's token usage growth.
• IPOs by Anthropic, OpenAI, Stepfun, Moonshot, and other peers could reduce MiniMax's scarcity premium and trigger a reset in valuation multiples.
• More severe compute bottlenecks could limit revenue upside.
• Lock-up expiries on July 8, 2026 and January 8, 2027 may create supply pressure.
• Model iteration performance may be weaker than expected.
• The pace of reaching break-even may be slower than expected.
• Heavy cash burn and refinancing needs could weigh on valuation.
• Geopolitical risk, IP infringement, and inappropriate content generation could affect the business and regulatory environment.
• Intense price competition could weaken the margin-improvement path.
• Talent attrition or weaker-than-expected hiring could affect model R&D and iteration.
• Changes in government attitudes toward AI could affect industry regulation and commercialization.

What to watch

• The release schedule and performance of MiniMax's M3 model in 1H26.
• Whether potential inclusion in Stock Connect is realized.
• Trends in token usage, ARR, and paid API growth for the M2/M3 series.
• Whether inference costs, training costs, and Hailuo video-generation latency continue to decline.
• Changes in revenue contribution from Open Platform, generative media, Agent, and companionship businesses.
• Model updates, pricing strategies, and capital-markets actions by OpenAI, Anthropic, Zhipu, and other competitors.
• Share-price and liquidity impact after the July 2026 and January 2027 lock-up expiries.
• Financing arrangements, cash burn rate, and visibility on the path to 2029e profitability.