China Internet and AI models: Goldman Sachs favors defensive games, AI cloud and selected China Internet ideas into year-end

Overview

This sector report reassesses China Internet and AI-model opportunities after 2Q results and investor conferences. Goldman Sachs expects soft macro conditions to create a mixed near-term earnings setup, but favors games, AI cloud, data centers and selected internet verticals for their relative resilience, AI-driven growth potential or valuations.

Core views

Goldman Sachs frames the near-term China Internet outlook around a weak consumption backdrop. Softer travel demand, cautious consumer spending and later Apple product shipments are expected to weigh on 3Q eCommerce, advertising and local-services revenue. The report nevertheless expects Games, Cloud and Internet Verticals to be relatively defensive into 3Q and 4Q. It also expects Alibaba and JD quick-commerce losses to halve year-on-year in 2H26, supporting a faster profit-growth recovery for transaction platforms. AI cloud is the principal growth counterweight. Strong token demand, high AI-server rental prices and cloud price increases since April 2026 are expected to support cloud growth and margins. Goldman Sachs estimates GPU-as-a-Service payback at two to three years, while noting that recent GPU rental pricing has shortened owner payback to roughly one to three years. The report sees scope for further cloud earnings beats as compute demand remains tight and US hyperscaler cloud growth provides a favorable reference point. However, negative recent free cash flow at Alibaba and Tencent, recent equity financings by Alibaba and Z.AI, the timing of capex peaks and the durability of elevated ROIC remain key investor concerns. If supply and demand normalize, shorter current payback periods may not be sustained. The report argues that the competitive basis for Chinese AI models is broadening beyond model quality. Narrower performance gaps and restricted access to Chain-of-Thought data increase the importance of post-training, cost efficiency, pricing power, financial strength and the harness or orchestration layer. Controlling workflow orchestration can create a feedback loop of user data, model improvement and eventual recursive improvement. Goldman Sachs believes Chinese models have reached sufficient intelligence and cost efficiency for virtual-worker tasks, expanding the focus from coding toward finance, legal and accounting workflows. It notes DeepSeek V4.1 Flash reduced KV Cache per token by 437x versus V1, while Z.AI/Zhipu raised its 2026 year-end ARR target to US$3 billion from US$2.4 billion, driven by compute investment, co-work demand and hyperscaler revenue sharing. AI agents are presented as the next commercialization battleground. Enterprise products such as Tencent WorkBuddy, Alibaba QwenWork and ByteDance Doubao Work are competing for workflow data and distribution. Goldman Sachs compares the model to eCommerce platforms: a first-party model can reinforce stickiness and post-training, while third-party models, skills, connectors and devices broaden task coverage. WorkBuddy had more than 70,000 developer-expanded skills and more than 100 ecosystem partners; QwenWork reported more than 30 million registered users, with over half enterprise accounts. Goldman Sachs estimates a current subscription-led annual revenue pool of about Rmb45 billion under illustrative assumptions, but argues the larger long-run opportunity could come from labor-budget substitution, where capturing 1% of China's white-collar wage pool would imply nearly Rmb200 billion of revenue potential. The report also stresses uncertainty around user-acquisition ROI and long-term retention. MiniMax's open-weight H3 illustrates both the benefits and limits of open model strategies. fal's post-trained H3 Max delivered a reported 35x faster generation rate, producing a five-second 768p clip in three seconds at comparable quality, while H3 Max Turbo roughly halved generation time again and listed at US$0.04 per second versus US$0.08 per second for H3 and H3 Max. Goldman Sachs sees MiniMax benefiting from broader adoption, external R&D leverage and user-feedback data, but also notes that third-party efficiency gains suggest video-generation moats remain uncertain. It expects different technical paths for premium high-fidelity content and faster-than-playback interactive uses such as livestreaming, gaming and companionship. Policy and tax developments are another material uncertainty. The report flags AI-model distillation rules, transactions with US suppliers, access to overseas leased compute and preferential high-tech tax treatment as factors that could affect compute access, market access and long-term after-tax profits. It expects manageable effects from announced eCommerce merchant taxes for shelf-based platforms and notes no year-to-date VAT rate change for games. Most China Internet mega-caps must seek renewal of High and New Tech Enterprise status by 4Q26; that status currently carries a 15% income-tax rate rather than the standard 25%. Goldman Sachs does not assume a tax-rate change, but shows that each 1 percentage-point tax-rate increase would reduce 2026E adjusted net profit by 1.1% for Tencent and Alibaba, and by 1.2% for NetEase and PDD. On valuation and stock selection, Goldman Sachs sees favorable risk-reward in China Internet mega-caps after year-to-date and post-results weakness. China Internet traded at a 13x median forward 12-month non-GAAP P/E versus 21x for US Internet on a GAAP median basis. The firm lifts Games and Entertainment to its number-one preference, favoring Tencent and NetEase for defensive growth; keeps Cloud and Data Centers second, favoring Alibaba and VNET; raises Internet Verticals/Education to third, adding New Oriental and KE Holdings; ranks eCommerce fourth with JD as the key idea; and puts AI models fifth with MiniMax as its preferred name. For Tencent, Goldman Sachs retains Buy and trims 2026E-28E net profit by 3%-8% because of investments in Hunyuan, WorkBuddy and Xiaowei. It forecasts 0% and -5% EPS growth in 3Q26E and 4Q26E, respectively, and lowers the 12-month SOTP target to HK$650 from HK$670. The report believes Tencent's core profits and US$130 billion offshore-dominated investment portfolio can fund near-term capex and prepayments without equity issuance. It identifies upcoming Hunyuan launches, WorkBuddy retention and paying-ratio data, Xiaowei adoption and evidence of Tencent Cloud's top-three hyperscaler positioning as key 6-12 month drivers. For Alibaba, Goldman Sachs retains Buy and sees cloud growth, improved capex-return visibility, narrowing quick-commerce losses and an EPS-growth inflection from the September quarter. It lifts September/December/March cloud-growth forecasts to 53%/55%/55%, from 50%/52%/52%, reflecting GPU-as-a-Service and MaaS demand. The report expects FY27E/FY28E EPS growth of 58%/27%, and sees FY27E/FY28E quick-commerce losses halving with a profit turnaround by FY29E on track. Its 12-month SOTP targets are cut to US$177/HK$172 from US$186/HK$180 because of dilution from the completed HK$80 billion equity raising. Other selected views include Buy-rated NetEase as a high-margin, cash-generative non-AI compounder, though game grossing growth may be lighter ahead of the January 2027 global launch of Ananta; JD as a multiple-repair story supported by retail recovery, cash flow and disciplined expansion; VNET as an AI-driven wholesale data-center growth story, with 862MW of year-to-date wins and 355MW of reservations; MiniMax as a potential valuation-repair candidate with multimodal commercialization and cost advantages; New Oriental for valuation, recovery and enhanced shareholder returns; and KE Holdings as a beneficiary of the "828 property policy" and a shift toward existing-home transactions and associated housing services.

Analysis framework

Goldman Sachs combines post-2Q company results, sector operating trends, app-engagement trackers, consensus revisions, valuation comparisons, cloud-capex and payback analysis, and company-level earnings and sum-of-the-parts forecasts. It then ranks sub-sectors and refreshes selected stock views based on growth resilience, AI exposure, valuation, funding capacity and execution risks.

Methodology notes

• Industry AnalysisSupply-demand framework

  AI-compute supply-demand and GPU rental-payback analysis

  The report links tight compute supply and strong token demand to GPU rental prices, cloud growth, margins, capex payback and the risk of future return normalization.

• Valuation methodsSOTP (Sum-of-the-Parts) Valuation

  Sum-of-the-parts valuation

  Goldman Sachs values Tencent and Alibaba by assigning separate values to operating segments, investments and net cash, then applying holding-company discounts to derive target prices.

• Valuation methodsDCF (Discounted Cash Flow)

  Discounted cash flow valuation

  The report uses DCF for MiniMax and other covered names where stated, discounting projected cash flows using explicit WACC and terminal-growth assumptions.

• Competition & strategyValue chain analysis

  AI model, harness and distribution-layer positioning

  The report assesses how models, post-training, workflow data, third-party ecosystems and user-distribution channels reinforce competitive positions and monetization.

Asset mapping & comparison

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

• Tencent Holdings (0700.HK)

  Preferred Games and Entertainment idea and Buy-rated AI-platform exposure

  Strengths

  Strong core profits, US$130 billion investment portfolio, AI-stack visibility and ecosystem advantages for WorkBuddy and Xiaowei.

  Weaknesses

  AI investment is pressuring near-term profits and estimates.

  Comparison

  At 13x 2026E P/E, the report cites an 11% downside to trough valuations.

  Risks

  Below-expected AI progress, higher AI investment, advertising competition, game-launch or approval delays, slower FinTech/Cloud growth and reinvestment risk.

• Alibaba Group (BABA/9988.HK)

  Preferred Cloud and Data Centers idea and Buy-rated eCommerce and AI-cloud exposure

  Strengths

  Accelerating cloud growth, improving quick-commerce economics and greater visibility of Cloud, T-Head and AI-applications value.

  Weaknesses

  Capex exceeds operating cash flow and the completed equity raise dilutes shareholders.

  Comparison

  At 14x 2026E P/E, the report cites a 16% downside to trough valuations.

  Risks

  Lower GMV growth, slower China-retail monetization, weaker strategic-investment execution and cloud-growth deceleration.

• NetEase (NTES/9999.HK)

  Preferred Games and Entertainment idea

  Strengths

  Record-high margins, free-cash-flow generation, channel-shift benefits and evergreen game franchises.

  Weaknesses

  Potentially light game pipeline and slower grossing growth in 2H26E.

  Comparison

  Positioned as a non-AI compounder relative to peers with elevated AI spending.

  Risks

  Weaker legacy-game performance, slower new-game monetization, competition and rising content or advertising costs.

• VNET Group (VNET)

  Preferred Cloud and Data Centers idea

  Strengths

  AI-investment-driven wholesale IDC demand, 862MW of year-to-date wins, reservations and potential CATL energy synergies.

  Weaknesses

  Project funding and execution remain important.

  Comparison

  Goldman Sachs forecasts 37% revenue CAGR and 35% EBITDA CAGR for wholesale IDC in 2026E-28E.

  Risks

  Financing constraints, weaker order execution, geopolitical AI risks, traditional-business weakness and changes in AI-model training demand.

• JD.com (JD/9618.HK)

  Preferred eCommerce idea

  Strengths

  Attractive valuation, cash flow, shareholder-return policies, recovering JD Retail growth and supply-chain-driven delivery opportunities.

  Weaknesses

  Retail margins face investment pressure from price competitiveness, customer experience and ecosystem investment.

  Comparison

  The report maintains Buy and unchanged 12-month targets of US$43/HK$169.

  Risks

  More intense eCommerce and food-delivery competition, weaker GMV or retail execution and margin volatility.

• MiniMax Group (0100.HK)

  Preferred AI-model idea

  Strengths

  Comprehensive multimodal offering, commercialization capability, token-cost advantage and open-weight strategy.

  Weaknesses

  Profit visibility remains limited amid high R&D and competitive intensity.

  Comparison

  The report sees a valuation discount versus Chinese video-generation peer Kling and Z.AI.

  Risks

  Weaker model performance, slower commercialization or profit path, IP/content risks, cash burn and geopolitical technology-race risk.

• New Oriental Education & Technology (EDU/9901.HK)

  Preferred Internet Verticals/Education idea

  Strengths

  Compelling valuation, prospective revenue recovery, cost optimization and enhanced shareholder returns.

  Comparison

  The report expects 1QFY27 revenue growth to exceed annual revenue growth.

• KE Holdings (BEKE/2423.HK)

  Preferred Internet Verticals idea

  Strengths

  Potential beneficiary of the 828 property policy, strong existing-home franchise and synergies in renovation, rental and living services.

  Weaknesses

  App engagement remained weak in July.

  Comparison

  The stock trades below 15x 2026E P/E and below 14x 12-month forward P/E, about one standard deviation below its historical average.

Key data

• China Internet versus US Internet valuation13x versus 21xChina Internet median forward 12-month non-GAAP P/E versus US Internet GAAP median P/E.
• Alibaba Cloud growth forecast53%/55%/55%September/December/March quarter growth forecasts, raised from 50%/52%/52%.
• Tencent target priceHK$65012-month SOTP target, lowered from HK$670.
• Alibaba target pricesUS$177/HK$17212-month SOTP targets, lowered from US$186/HK$180 primarily for equity-raising dilution.
• Z.AI 2026 year-end ARR targetUS$3 billionRaised from US$2.4 billion.
• Workplace-agent illustrative subscription TAMc.Rmb45 billion annual revenue poolIllustrative estimate; the report sees labor-budget substitution as the larger long-term opportunity.
• VNET order wins862MW year-to-date plus 355MW reservationsFull-year new wholesale order guidance is about 1.0-1.1GW, or about 1.4GW including reservations.

Impact & implications

The report's preferred positioning emphasizes relatively resilient games, AI-driven cloud and data-center growth, and selected vertical or value opportunities rather than broad exposure to consumption-sensitive eCommerce and advertising. It argues that sustained AI returns depend on compute economics, funding capacity, commercialization, workflow adoption and policy access, not simply model performance.

Risks

• Soft consumption, weaker travel demand and delayed product shipments could further pressure eCommerce, advertising and local-services growth.
• AI-cloud investment may require sustained capex or additional funding before free cash flow improves, and current high ROICs could fade if compute supply normalizes.
• AI-model competition, uncertain monetization, high R&D costs, user-acquisition ROI and long-term retention can limit returns from AI-agent investment.
• Policies on AI-model distillation, US-supplier transactions, overseas compute access, taxes and High and New Tech Enterprise renewals could affect profits and market access.
• Company-specific risks include weaker cloud, game, retail or order-growth execution and intensified competition across covered sectors.

What to watch

• 3Q and 4Q revenue trends in eCommerce, advertising, local services, games and cloud.
• Alibaba Cloud growth, GPU-as-a-Service and MaaS demand, and the relationship between cloud capex, returns and free cash flow.
• Tencent Hunyuan releases, WorkBuddy retention and paying metrics, Xiaowei adoption and Tencent Cloud positioning.
• AI-agent user growth, workflow integration, enterprise adoption and monetization progress.
• HNTE renewals by 4Q26, AI-policy developments, tax changes and access to overseas AI compute.
• VNET order wins, financing progress and CATL partnership completion.