Goldman Sachs upgrades MiniMax to Buy, deconstructs China’s internet through five AI questions

Overview

This is Goldman Sachs’ mid-term strategy and company research report on China’s internet/AI model industry. The report focuses on five key AI debates—US-China model performance gap, competitive moats for Chinese models, sustainability of token consumption, domestic chip substitution, and consumer-grade AI Agents (OS-level vs in-app). At the same time, it upgrades MiniMax’s (0100.HK) rating from Neutral to Buy, setting a target price of HK$1,000, while maintaining Alibaba’s (09988.HK/BABA) Buy rating (Conviction List) and continuing to favor China’s AI cloud and data center sectors.

Core views

The report argues that the performance gap between US and Chinese foundational models is narrowing, with Chinese models gaining competitiveness thanks to pricing, speed, and task-specific performance. However, the U.S. still leads in complex coding, top-tier training budgets, and cutting-edge models. Specifically, new models like DeepSeek V4, Tencent Hy3, Xiaomi MiMo V2.5, and Alibaba Qwen3.6 have been released one after another, intensifying competition within the 200–3 trillion parameter range and fragmenting the model landscape. Future moats will hinge on coding capabilities, multimodal abilities, and task completion rates, with billing models potentially shifting from per-token charges to success-based fees. On the demand side, token consumption continues to grow rapidly. According to the National Data Administration, China’s daily token consumption exceeded 140 trillion in March, more than a thousandfold increase since early 2024. The report anticipates that enterprise and agent applications will drive sustained token growth over the years. Chinese cloud providers’ capital expenditures are projected to account for roughly 60% of operating cash flow by 2026E, significantly lower than the U.S. peers’ ~90%, suggesting further room for increased spending in the second half of 2026 through 2028. Cloud service providers’ pricing power strengthens as demand tightens and model performance improves. Regarding MiniMax, its ARR has grown robustly year-to-date, with an annualized recurring revenue averaging approximately $150 million per day in February, and expected to reach around $250 million by late March—comparable to Zhipu’s disclosed figure. Based on this, Goldman Sachs has raised MiniMax’s 2026E/2027E revenue forecasts by 25% and 11%, respectively. With full-modal capabilities covering text, images, video, audio, and music, along with an industry-leading low-cost, efficient inference architecture (text API gross margin ~40%, multimodal 60–70%), MiniMax stands out as a rare independent AI vendor. Upcoming M3 models (including trillion-parameter versions and smaller variants) and Hailuo 3 are seen as key catalysts. Current stock prices have retreated about 40% from their March highs, improving the risk-reward ratio. As for consumer-grade AI Agents, the report suggests that OS-level agents (such as DouBao Assistant) may become major long-term traffic entry points, posing a threat to traditional apps. Giants like Tencent and Alibaba will leverage their payment, logistics, and social graph advantages to strengthen in-app agents and defend their moats. The report also tracks usage durations across various verticals of China’s internet in March: e-commerce up 11% yoy, gaming up 10% yoy, local services up 4% yoy, AIGC-to-C app usage up 36% mom, with DouBao capturing 63% of total usage time.

Analysis framework

Goldman Sachs employs a ‘five-debate’ framework to dissect the main investment themes in China’s AI model sector: first, it reviews market controversies and evidence across five dimensions—performance gaps, competitive landscapes, token sustainability, chip supply chains, and consumer-grade agents—before offering corresponding judgments. For MiniMax’s valuation, it uses a multi-scenario DCF approach: under baseline assumptions of 12% WACC and 2% perpetual growth, it projects global subscription and API revenue market share rising from 0.5% to 2.5% between 2026 and 2030, with a long-term adjusted EBIT margin of around 21%. In pessimistic scenarios, it applies a 19x EV/Sales multiple based on 2027E revenues; in optimistic scenarios, it uses a 30x P/ARR multiple for mid-2028 ARR. The report cross-validates these projections using third-party data from OpenRouter, Artificial Analysis, QuestMobile, SensorTower, and others to verify model rankings, token usage, app duration metrics, and more.

Methodology notes

• valuation methodDCF cash flow discounting

  Discounted Cash Flow (DCF) valuation

  For MiniMax’s baseline valuation, Goldman Sachs adopts DCF, setting a 12% WACC and 2% perpetual growth rate, projecting future free cash flows discounted back to 2026 to derive equity value. DCF is suitable for high-growth tech companies that have not yet stabilized profitability, reflecting long-term market expansion and potential inflection points.

• valuation methodEV/EBITDA valuation

  Relative EV/Sales valuation

  In MiniMax’s pessimistic scenario, the report benchmarks against traditional Chinese AI companies using a 19x EV/Sales multiple based on 2027E revenues, capturing short-term pricing dynamics when profits remain elusive but growth is rapid.

• industry analysis frameworksupply-and-demand framework

  Tight supply of computing power and pricing power

  The report repeatedly attributes rising AI model pricing and expanding cloud provider capital expenditures to strained computing capacity (domestic high-end chips restricted, overseas model access limited) coupled with surging demand (explosive token volume, widespread enterprise agent adoption).

• competition and strategy frameworkmoat/competitive advantage framework

  Differentiated advantages between independent AI vendors and internet giants

  The report contends that independent AI firms like MiniMax excel in organizational efficiency and swift decision-making, enabling rapid model iteration, whereas internet giants rely on robust operating cash flow from core businesses and cloud infrastructure to capture AI cloud/data center dividends—but require separate incentive mechanisms to retain AI talent.

• industry analysis frameworkpenetration S-curve framework

  TAM and penetration rate in the AI video generation market

  When analyzing the video-generation model space, the report breaks down global TAM into driving factors such as paying creator numbers, monthly ARPU, AI penetration rates, and ad/film post-production penetration rates, illustrating an S-shaped penetration trajectory.

Asset mapping & comparison

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

• MiniMax (0100.HK)

  Core asset whose rating was upgraded from Neutral to Buy; directly benefits from ARR growth, M3 and Hailuo 3 releases, and full-modal AI capabilities

  Strengths

  Full-modal AI portfolio (text, image, video, audio, music); overseas revenue share around 70%; text API gross margin ~40%, multimodal 60–70%; high cost efficiency and organizational agility

  Weaknesses

  Not yet profitable, with persistent net losses and operating cash outflows through 2023–9M25; faces short-term pressures such as lock-up expiration and intensified technical competition

  Comparison

  Among 200–3 trillion parameter models, ranks highly with stronger multimodal capabilities than most peers; however, lags behind U.S. SOTA and certain Chinese giant models in complex coding and other advanced scenarios

  Risks

  Model performance falling short of expectations; intensifying competition; unclear commercialization paths and profit visibility; IP/content generation risks; cash burn and financing challenges; geopolitical risks stemming from U.S.-China tech rivalry

• Alibaba Group (09988.HK / BABA)

  Maintained at Buy (Conviction List); Alibaba Cloud and AI model/application ecosystems are direct beneficiaries

  Strengths

  Alibaba Cloud’s growth accelerating, expected to expand by about 40% in the March quarter; leading HappyHorse video-generation model; rapid Qwen model iterations; increasing APAC IaaS market share

  Comparison

  Ranked alongside GDS, VNET, Kingsoft Cloud as preferred choices for cloud and data center sectors; fourth place in overseas IaaS markets

• Cloud & Data Centers (GDS, VNET, Kingsoft Cloud, etc.)

  Goldman Sachs’ #1 preferred sub-sector; benefiting from surging AI token demand and improved cloud pricing

  Strengths

  AI training and inference needs drive IDC and cloud service demand; domestic computing power shortages exacerbate supply-demand mismatches

  Risks

  Fluctuating capital expenditure cycles; progress in domestic chip substitution; constraints on electricity and land supplies

• Tencent (0700.HK)

  Defensive advantage in AI Agents/WeChat ecosystem; participating in Hy3 model competition

  Strengths

  Super App ecosystem of WeChat, integrated social graphs, payments, and logistics; in-app agents can protect traffic entry points

  Comparison

  Compared to independent AI vendors, organizational efficiency and incentive structures may be slower; but cash flow and ecological advantages are strong

• PDD Holdings (PDD)

  Listed as one of China’s mega-cap key ideas for 2026; Temu’s globalization and revenue re-acceleration

  Strengths

  Top-line growth expected to re-accelerate; room for valuation reassessment; Temu’s overseas expansion

Key data

• MiniMax 2026E/2027E revenue forecast increase+25% / +11%2026E revenue revised upward from $240 million to $300 million
• MiniMax 12-month target priceHK$1,000Implies approximately $4 billion valuation, representing a ~40% upside compared to the last closing price
• MiniMax Bear/Base/Bull valuationsHK$420 / HK$1,000 / HK$1,600Correspond to -41% / +40% / +124% downside risks, respectively
• MiniMax average ARR in Februaryapproximately $150 millionExpected to reach around $250 million by late March, comparable to Zhipu’s reported figure
• Alibaba Cloud’s projected March quarter growth rateabout +40% yoyHigher than the December quarter’s +36%
• 2026 Chinese cloud providers’ capital expendituresover $70 billion+Account for roughly 60% of operating cash flow; U.S. peers spend over $700 billion+, accounting for about 90% of operating cash flow
• China’s daily token consumption (March 2026)exceeds 140 trillionMore than a thousandfold increase since early 2024, up about 40% from the end of 2025’s 100 trillion
• DouBao’s March DAU/duration sharearound 150 million / 63%A super entry point for consumer AI, with steadily growing market share
• MiniMax DCF parametersWACC 12%, perpetual growth rate 2%Projects global subscription and API market share reaching 2.5% by 2030
• MiniMax API business gross margins (GSe)text around 40%, multimodal 60–70%Reflects cost efficiency and flexible computing architecture

Impact & implications

The report concludes that the gap between China’s foundational models and those of the U.S. is narrowing, with improved model pricing and ARR trends, making cloud/data centers the highest-priority sub-sectors in the AI wave. MiniMax, with its full-modal layout, high overseas revenue share (~70%), and cost efficiency, enjoys differentiated advantages; following a 40% pullback from its March peak, its risk-reward profile has markedly improved. In the medium to long term, AI Agents—especially OS-level ones—may reshape traffic entry points, compelling traditional apps to transition toward in-app agents to maintain their ecosystem footholds.

Risks

• Intensified global competition in foundational models, with model performance falling short of expectations
• Slow profit visibility, with MiniMax persistently posting net losses and operating cash outflows
• Commercialization capabilities lagging behind expectations
• IP/content generation-related risks
• Cash burn and self-financing capability risks
• Escalating geopolitical tensions and supply chain restrictions resulting from U.S.-China tech rivalry
• Supply pressures before lock-up expiration and potential inclusion in HSTECH/Hong Kong Stock Connect

What to watch

• Release and market reception of MiniMax’s M3 model (including trillion-parameter version)
• Launch of MiniMax’s Hailuo 3 video-generation model
• Monthly ARR trends and comparisons with peers like Zhipu
• Progress in Chinese cloud providers’ capital expenditures and domestic chip supply (Huawei Ascend 910C/950, Alibaba Pingtouge T-Head, etc.)
• Competitive dynamics among DeepSeek V4, Tencent Hy3, Xiaomi MiMo V2.5, and other models
• Evolution of consumer-grade AI Agents: OS-level agents (DouBao Assistant) versus in-app agents (WeChat AI Agent)
• Token growth and cloud pricing changes expected in the second half of 2026 through 2027