China Internet AI enters a phase of reassessing model competition, token volume growth, and computing capex
AI summary card
China Internet AI enters a phase of reassessing model competition, token volume growth, and computing capex
Goldman Sachs reassesses China's internet sector around five major AI debates, forecasting China's daily token usage to reach 350 trillion by end-2026, while continuing to rank cloud and data centers as the preferred direction.
- The report focuses the China internet AI discussion on five questions: the China-US model gap and pricing, China's model competitive landscape, token growth and ROI, hyperscale cloud providers' margins/capex, and the form and application impact of consumer AI agents.
- Goldman Sachs expects China's daily token usage to rise from about 140 trillion in March 2026 to 350 trillion by end-2026, equivalent to 10.5 quadrillion tokens in December alone.
- Cloud and data centers remain the preferred sub-sector, supported by expanding AI token demand, rising cloud pricing, improving margins, and higher cloud-provider capex in 2H26.
- On the defensive side, gaming and entertainment are upgraded to second place, while e-commerce and mobility fall to third, though JD.com is listed as a key idea for 2H26.
- MiniMax is a key favorite, mainly due to its full-modality capabilities, launches such as M3/Hailuo 3, ARR upside potential, and a more attractive P/ARR valuation versus peers.
Report interpretation
Overview
This Goldman Sachs industry research report reviews the key investment themes in China internet and AI models after earnings season and the Asia Communacopia + Technology conference. The report argues that China's AI model ecosystem is undergoing multiple shifts: performance catch-up, price competition, surging token demand, rising computing capex, and consumer AI agents reshaping traffic entry points. From an investment perspective, the report continues to favor cloud and data centers, raises the relative ranking of gaming and entertainment amid weak macro consumption, and offers differentiated views on companies such as Alibaba, Tencent, JD.com, and MiniMax.
Core views
The core views are: first, Chinese models are gradually becoming close to 'good enough' on general tasks, while pricing is significantly below that of leading US frontier models, reinforcing their value-for-money positioning; second, competition in China's model layer is more fragmented, and may eventually form a two-tier pricing structure differentiated by performance/multimodal capabilities; third, agentic AI will drive exponential token usage growth, but ROI depends on actual task completion rates, engineering overhead, and token unit costs; fourth, as domestic chip supply improves, Chinese hyperscale cloud providers may raise capex from 2H26 into 2027; fifth, consumer AI agents may change application traffic entry points, but will also bring inference cost and margin pressure.
Analysis framework
The report combines an industry-debate framework, sub-sector ranking, review of key company catalysts, and valuation/ARR comparisons. It evaluates model performance, token pricing, ARR targets, cloud capex, cloud margins, consumer AI entry points, macro consumption trends, and single-stock valuations within one framework, thereby forming sub-sector preferences and key stock ideas.
Methodology notes
Use five core questions to break down the main investment themes of China internet AI
The report analyzes the gap and pricing between Chinese and US AI models, China's model competition, token growth and ROI, cloud-provider margins/capex, and the form and impact of consumer AI agents.
Compare talent stability, intelligence level, pricing power, cost efficiency, and financial strength
The report believes coding, multimodal video generation, training efficiency, and balance sheet strength will determine the long-term competitive advantages of Chinese AI model companies.
Derive valuation differences among AI model companies using year-end ARR targets and market capitalization
The report compares the P/ARR of Knowledge Atlas and MiniMax, arguing that MiniMax, at around 20x year-end P/ARR, offers more upside-tilted risk-reward relative to global peers at a similar stage and to Knowledge Atlas.
Rank directions such as cloud and data centers, gaming and entertainment, e-commerce and mobility, and AI models
The report ranks cloud and data centers first, gaming and entertainment second, e-commerce and mobility third, and AI models fourth.
Asset mapping & comparison
Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).
- AlibabaChina internet giant and a key participant in cloud computing and AI models
- Strengths
- Accelerating cloud revenue, potential margin improvement from rising pricing, and AI upside from MaaS ARR trends and the Qwen model ecosystem.
- Weaknesses
- Short-term pressure on CMR, while consumer AI investments booked under 'All others' create EPS downgrade pressure.
- Comparison
- Compared with Tencent, Alibaba's share-price decline is driven more by EPS downgrades than by valuation multiple compression; compared with pure-play model companies, Alibaba has a full-stack advantage across cloud, models, and capex capability.
- Risks
- Slower retail sales, continued profit drag from AI investment, and weaker-than-expected cloud growth or pricing uplift.
- TencentChina internet giant and a participant in gaming and entertainment, advertising, cloud, and consumer AI agents
- Strengths
- Gaming and advertising are defensive; ad growth is supported by AI targeting and higher ad load, while the Hunyuan model and Weixin AI agent may create new entry points.
- Weaknesses
- The market is concerned about AI progress and visibility into consumer AI ROI, while fintech revenue is affected by weak consumption.
- Comparison
- Tencent's year-to-date decline is driven mainly by valuation compression rather than clear EPS downgrades; its valuation multiple is at cyclical lows versus domestic and global internet peers.
- Risks
- AI investment suppresses profit growth, while early-stage Weixin agent inference costs may be high and monetization uncertain.
- GDSA key idea within the preferred cloud and data center segment
- Strengths
- Benefiting from expanding AI token demand, higher cloud-provider capex, and improving sequential data-center demand in 2H26.
- Weaknesses
- 1Q26 online capacity expansion and customer move-ins were weaker than expected.
- Comparison
- Compared with internet platforms, GDS is more directly sensitive to computing capex and data-center demand.
- Risks
- Delayed cloud-provider capex, slow customer move-ins, and domestic chip supply bottlenecks.
- VNETA key idea within the preferred cloud and data center segment
- Strengths
- Could benefit from accelerating China cloud and AI infrastructure demand in 2H26.
- Weaknesses
- In the near term, the industry is still affected by capacity expansion, move-in pace, and fluctuations in customer demand.
- Comparison
- Similar to GDS, VNET is an important data-center exposure in the report.
- Risks
- Capex cycles falling short of expectations, price competition, and financing/utilization pressure.
- Kingsoft CloudA key idea within the preferred cloud and data center segment
- Strengths
- Higher cloud pricing, rising AI token demand, and improving industry margins may support growth.
- Weaknesses
- Competition in China's IaaS market remains intense, and the margin base is weaker than that of US cloud peers.
- Comparison
- Compared with large platform clouds, Kingsoft Cloud offers more pure-play cloud business elasticity.
- Risks
- Cloud pricing improvement proves unsustainable, customer demand fluctuates, and gross margins on AI-related revenue miss expectations.
- MiniMaxA key idea in the AI model segment
- Strengths
- Full-modality capabilities, launches such as M3 and Hailuo 3, and upside in API revenue gross margin and ARR are the main highlights.
- Weaknesses
- Faces price competition and rapid iteration pressure from models such as DeepSeek, Tencent, MiMo, Qwen, and GLM.
- Comparison
- The report believes MiniMax's roughly 20x year-end P/ARR is below Knowledge Atlas at about 75x and below some global peers at a similar stage.
- Risks
- Price wars erode premium, Hailuo 3 underperforms Seedance/Kling, and expiration of the six-month lock-up increases free float and share-price volatility.
- JD.comA key idea for 2H26 in e-commerce and mobility
- Strengths
- The report believes 2Q26 may be the final quarter with the toughest YoY comparison pressure, with potential recovery in YoY revenue and profit in 2H26.
- Weaknesses
- 2Q retail sales and the e-commerce advertising environment are weak, while a mid-to-high single-digit decline in JD Retail revenue is already expected by the market.
- Comparison
- Although the e-commerce and mobility sub-sector ranking was lowered, JD.com is singled out as a key idea.
- Risks
- Continued weak consumption, intensified competition, and slower-than-expected profit recovery.
- NetEaseA newly added key idea in gaming and entertainment
- Strengths
- Gaming has counter-cyclical characteristics, valuation is at cyclical lows, and new game launches may provide catalysts.
- Weaknesses
- Revenue still depends on game product cycles and launch cadence.
- Comparison
- The report upgrades gaming and entertainment from third to second place and newly highlights NetEase.
- Risks
- New games underperform expectations, or regulation/user time shifts.
- ByteDanceAn important reference entity for China's AI models, consumer AI, and token demand
- Strengths
- Doubao leads usage in China's consumer AI, while Seedance 2.0 shows strong traction in video generation and enterprise contracts.
- Weaknesses
- After Doubao introduced subscription packages, there were signs of its first month of sequential MAU decline.
- Comparison
- ByteDance creates significant competitive pressure on independent model companies such as MiniMax in consumer AI token consumption and multimodal video models.
- Risks
- Uncertainty around subscription conversion, user retention, video-model competition, and AI cost structure.
Key data
- China token usage forecast350 trillion tokens/day by end-2026, equivalent to 10.5 quadrillion tokens/month in December 2026The report expects China's token demand to continue growing from about 140 trillion/day in March 2026.
- China token scale in March 2026About 140 trillion tokens/day, or roughly 4 quadrillion tokens/monthData cited from China's National Data Administration.
- Global agentic AI token growthBy 2030, monthly token consumption will be 24x higher than in 2026, reaching 120 quadrillion/monthThe report cites estimates from Goldman Sachs' US research team, with enterprise-agent growth exceeding consumer-agent growth.
- BBAT capex forecastAbout US$100bn in 2026, skewed toward 2H26The report expects rising cloud-provider capex to support sequential acceleration in data centers in 2H26.
- China model price rangeAbout US$0.2-1 per million tokens, below the blended price of about US$4 per million tokens for US SOTA modelsLow pricing and a narrowing performance gap are the core of Chinese AI models' value-for-money positioning.
- MiniMax valuation scenarioBear/base/bull valuation of HK$420/HK$1,000/HK$1,600, corresponding to about -17%/+97%/+215%The report believes its 12-month risk-reward is skewed to the upside.
- Alibaba stock driver breakdownDown about 22% year to date, mainly driven by EPS downgradesThe report believes its EPS downgrade cycle may bottom in 2H26.
- Tencent stock driver breakdownDown about 27% year to date, mainly driven by valuation multiple compressionThe report notes Tencent trades at about 11-13x 2026E P/E, near cyclical lows versus domestic and global peers.
Impact & implications
The investment implication is that China's internet AI chain is no longer just a pure model-performance race, but has shifted toward comprehensive competition across performance, price, token demand, cloud margins, capex returns, and control of consumer entry points. The report prefers positioning in cloud and data centers that benefit from token volume growth and computing demand, while increasing the defensive weight of gaming and entertainment in a weak consumption environment; for big tech, it focuses on Alibaba's EPS bottoming and Tencent's valuation bottoming, while for model companies it places greater emphasis on ARR visibility, gross margin, and multimodal differentiation.
Risks
- China AI model price wars intensify, putting pressure on model-layer revenue and gross margins.
- Recovery in domestic high-end chip and ASIC supply is slower than expected, constraining cloud-provider capex and inference-capacity expansion.
- Token consumption driven by agentic AI fails to convert into sufficient ROI, leading enterprise customers to cut or optimize token spending.
- Consumer AI agents have limited early monetization, but significantly increase inference costs, weighing on platform margins.
- Macro consumption remains weak, dragging on e-commerce, advertising, fintech, and retail-related businesses.
- Data-center customer move-in pace, utilization, or cloud pricing improvement falls short of expectations.
- MiniMax faces risks from lock-up expiration, increased free float, pricing-strategy pressure, and validation of multimodal model performance.
- AI model safety, regulation, data permissions, and cross-border usage restrictions may affect the global adoption of Chinese models.
What to watch
- Whether China cloud providers' capex is meaningfully revised up in 2H26.
- Whether Alibaba Cloud's revenue growth, pricing, and margins improve simultaneously.
- Progress in Tencent advertising, cloud business, the Hunyuan model, and Weixin AI agent pilots.
- Whether China's daily token usage moves toward the 350 trillion target.
- Performance, pricing, and ARR delivery of MiniMax M3, Hailuo 3, and subsequent large-parameter foundation models.
- Changes in pricing and token share for models such as DeepSeek, Qwen, GLM, MiMo, Seedance, and Kling.
- Whether JD.com's YoY revenue and profit recovery appears as expected after 2Q.
- Whether consumer AI agents are dominated by OS-level entry points or by in-app agents within super apps.