China AI value chain Report Interpretation

Overview

This thematic report examines whether generative AI can move from heavy infrastructure spending to sustainable monetization. Nomura’s central case is that expanding token consumption, enterprise adoption and China’s AI infrastructure buildout can support a longer investment cycle, with differentiated implications for cloud platforms, IDCs and software providers.

Core views

Nomura argues that generative AI is at a turning point: adoption is broadening beyond chatbots and copilots toward agentic AI, edge AI and industrial models. It expects leading LLM platforms to refresh models every three to six months and views AI agents, edge devices and industrial use cases as key adoption drivers over the next one to three years. The report contrasts global closed-source models, which seek technological leadership and premium monetization, with China’s open-source approach, which it believes improves cost efficiency, enlarges developer ecosystems and accelerates application adoption. The report says the performance gap between leading open- and closed-source models is narrowing, although closed-source systems still lead in difficult reasoning, professional coding and knowledge accuracy. It cites open-source-model scores of 34%-36% on Humanity’s Last Exam versus 44% for GPT-5.5, 43%-46% versus 61% on TerminalBench Hard, and a 49-point gap between Claude Opus 4.6 and GLM-5 in the AI Index ranking as of March 2026. Nevertheless, open-source models are said to deliver near-frontier performance at roughly one-half to one-sixth of the cost, making them especially attractive for commercial deployment. Nomura views China’s open-source ecosystem, domestic-chip advances and lower operating costs as foundations for broader AI penetration. Nomura’s token-economy framework is the report’s central monetization thesis. It assumes total AI compute from major LLM and hyperscale players rises from 13GW in 2025F to 98GW in 2030F, while output per GW improves by 35% annually and blended token prices decline by 20% annually to stimulate usage. On these assumptions, platform token revenue increases from USD136bn in 2025F to USD1.348tn in 2030F. The report expects inference to take a growing share of workloads as consumer and enterprise applications scale, which matters because inference generates token revenue whereas training does not directly do so. Nomura distinguishes asset-light standalone LLM platforms from vertically integrated hyperscalers. Standalone providers rent compute and capture API and token revenue, leaving their margins sensitive to token pricing and compute-rental costs; the report estimates their EBIT margin could reach 29.6% in 2030F with roughly 70.5% gross margin. Hyperscalers fund infrastructure themselves but can absorb compute costs through scale and ecosystem integration; Nomura estimates a 47.1% EBIT margin and pre-tax ROIC of 18% in 2030F. It therefore sees the larger integrated cloud ecosystems as better positioned to earn sustainable returns from AI investment. Infrastructure spending remains substantial in the report’s base case. Bloomberg consensus cited by Nomura puts combined 2026E/2027E capex for Meta, Amazon, Alphabet and Microsoft at USD710bn/USD907bn. Nomura estimates China’s top cloud service providers and telecom operators will spend CNY981.7bn in 2026F and CNY1,106.8bn in 2027F. It believes China remains underinvested because of constraints on advanced chips and uncertain monetization, but expects policy support, domestic-chip upgrades, open-source LLM development and participation by government, SOEs and private companies to accelerate AI data-center investment. The report identifies finance, software, healthcare, autonomous driving and robotics as major application frontiers. Enterprise monetization appears clearer than consumer monetization: membership fees account for 50% and pay-as-you-go models for 30% of enterprise AI-product revenue models cited by the report, while more than 40% of consumer AI products still lack a clear revenue model. In China, Nomura expects business-to-business applications to commercialize faster than consumer applications. It highlights applications including AI agents, AI office software, financial-service automation, medical imaging and drug discovery, end-to-end autonomous driving, and embodied robotics. For China AI infrastructure, Nomura highlights robust IDC demand linked to token consumption. China’s cumulative IDC rack count rose 53% year-on-year to 13.7mn at end-2025, while intelligent computing power reached 1,590 EFLOPS at end-2025 and 2,185 EFLOPS by end-June 2026. Ulanqab is presented as a key AI data-center hub because of natural cooling, proximity to Beijing and renewable-power resources. Nomura estimates utilized IT power there at 0.8GW by July 2026, versus contracted capacity equivalent to 12.5GW, and sees VNET and GDS among participants positioned to benefit from northern-China AI data-center demand. At the stock level, Nomura recommends AI platform, infrastructure and software leaders Alibaba, Kingdee, Kingsoft Cloud, Kingsoft Office, GDS and VNET, while retaining Neutral on iFlytek. Kingdee is viewed as progressing toward an AI-native software model; Kingsoft Office is seen benefiting from enterprise AI agents but facing growing competition; Kingsoft Cloud is expected to benefit from AI-cloud demand despite chip-supply and margin pressure; GDS and VNET are backed by AI-data-center order intake and capacity expansion; and iFlytek is viewed as well placed in domestic LLMs and enterprise verticals but constrained by intense competition and weaker consumer and education trends.

Analysis framework

Nomura combines technology and market-structure analysis of open- versus closed-source LLMs, application adoption and policy conditions with a token-economy model linking compute capacity, token output, pricing and monetization. It then applies company-specific earnings, order-backlog, margin and valuation analysis to selected China AI value-chain companies.

Methodology notes

• Industry AnalysisUpstream-Midstream-Downstream Transmission

  China AI value-chain analysis

  The report connects chips and computing infrastructure with cloud platforms, LLM providers, applications, enterprise software and data centers to explain how AI demand and monetization flow through the value chain.

• Industry AnalysisSupply-demand framework

  Token-economy supply and demand model

  Nomura estimates token supply from inference compute capacity and compares it with projected token demand, using assumptions on hardware efficiency, utilization and token prices.

• Valuation methodsDCF (Discounted Cash Flow)

  Company target-price valuation

  The report uses discounted cash flow valuation for Kingdee, Kingsoft Office, Kingsoft Cloud and VNET, discounting projected cash flows using stated WACC and terminal-growth assumptions.

• Valuation methodsEV/EBITDA valuation

  GDS target-price valuation

  Nomura derives GDS’s target price from an FY27F EV/EBITDA multiple aligned with the company’s historical average.

• Valuation methodsP/E and PEG Valuation

  iFlytek price-to-earnings valuation

  The report values iFlytek using a 65x FY27F EPS multiple, stated to be in line with the A-share software median P/E.

Asset mapping & comparison

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

• Alibaba (9988 HK / BABA US)

  AI platform and cloud leader recommended by Nomura

  Strengths

  Open-source ecosystem and AliCloud exposure

  Comparison

  Part of Nomura’s preferred AI platform cohort

  Risks

  Margin pressure from investment ramp-up and regulatory risks in payments and internet finance

• Kingdee (268 HK)

  AI software leader recommended by Nomura

  Strengths

  AI-native products, subscription growth and enterprise AI operating system Lingee

  Weaknesses

  Cloud-business turnaround remains dependent on execution

  Comparison

  Preferred China software exposure in the report

  Risks

  Slower cloud growth, additional R&D needs and competition from incumbent ERP platforms

• Kingsoft Office (688111 CH)

  AI office software leader recommended by Nomura

  Strengths

  WPS 365, Lingxi and WPS Comate support enterprise AI monetization

  Weaknesses

  Slower-than-expected WPS AI user and ARPU growth led to forecast reductions

  Comparison

  Faces competition from Alibaba and Tencent in AI collaboration software

  Risks

  Weaker demand, slower AI and subscription development, and intensifying AI-office competition

• Kingsoft Cloud (3896 HK)

  AI-cloud infrastructure leader recommended by Nomura

  Strengths

  AI cloud and MaaS demand, including ecosystem customers and LLM customers

  Weaknesses

  Higher hardware and third-party rental costs pressure margins and delay profitability

  Comparison

  An independent China cloud service provider with AI exposure

  Risks

  Slower public- or enterprise-cloud demand and intensified competition

• GDS Holdings (GDS US)

  China IDC leader recommended by Nomura

  Strengths

  Strong 2Q26 order intake, raised booking target and sizeable resource pipeline

  Weaknesses

  Near-term EBITDA growth is moderated by contract renewals and a changing market mix

  Comparison

  A key AI-data-center beneficiary alongside VNET

  Risks

  Weaker China AI-data-center demand, overseas expansion delays, competition and geopolitical supply-chain disruption

• VNET Group (VNET US)

  China IDC leader recommended by Nomura

  Strengths

  Large wholesale bookings, land bank and Ulanqab AI-data-center positioning

  Weaknesses

  New capacity requires utilization ramp-up

  Comparison

  An early mover in Inner Mongolia AI data centers

  Risks

  Slower utilization of new facilities and faster MRR declines amid competition

• iFlytek (002230 CH)

  Domestic LLM and AI-application provider rated Neutral by Nomura

  Strengths

  Domestic-chip-trained Spark LLM and exposure to enterprise verticals including healthcare and autos

  Weaknesses

  Education and consumer segments face competitive pressure

  Comparison

  Nomura states a preference for Kingdee within China software

  Risks

  Rising LLM and voice-recognition competition, slower technology ramp-up and delayed monetization

Key data

• Global token-market revenueUSD136bn in 2025F to USD1.348tn in 2030FNomura token-economy forecast
• Global AI computing power13GW in 2025F to 98GW in 2030FNomura estimate for leading LLM and hyperscale players
• Standalone LLM-platform EBIT margin29.6% in 2030FNomura token-economy model
• Hyperscaler AI-platform EBIT margin47.1% in 2030FNomura token-economy model
• Hyperscaler AI-platform pre-tax ROIC18% in 2030FNomura model; the report cites a 17%-20% range in FY28-30F
• China AI-related CSP and telecom capexCNY981.7bn in 2026F and CNY1,106.8bn in 2027FNomura estimate
• Global AI-agent marketUSD5.29bn in 2024 to USD47.1bn by 2030EMoonfox estimate cited by Nomura
• China cumulative IDC racks13.7mn at end-2025, up 53% year-on-yearNDA and CAICT data cited by Nomura

Impact & implications

Nomura sees token consumption as the mechanism that can convert AI compute investment into recurring revenue. Its preferred exposures are integrated AI platforms, AI-cloud providers, IDCs with scalable capacity and enterprise software firms able to monetize agents and subscriptions; it expects China’s open-source ecosystem and domestic infrastructure buildout to support these areas.

Risks

• Major LLM players may generate lower-than-expected ROI if competition intensifies, compute costs rise or LLM intelligence improves more slowly than expected.
• China AI-application monetization may be slower than expected because of competition, weak end-customer purchasing power and macro conditions.
• China may face IDC oversupply as government, SOEs and private companies simultaneously expand data-center capacity.
• US-China technology restrictions could tighten supply conditions and raise costs across chips, optical components and LLM infrastructure.
• AI-agent deployment carries privacy, data-theft and cybersecurity risks where systems have elevated access to user devices and data.

What to watch

• Token-consumption growth, inference-workload share and the ability of LLM platforms to sustain pricing and utilization.
• China AI-cloud and AI-data-center capex, domestic-chip capacity expansion and the pace at which supply constraints ease.
• Commercial progress of enterprise AI products, particularly agents, MaaS, API usage and subscription adoption.
• IDC order intake, customer move-ins, capacity delivery and utilization rates for GDS and VNET.
• Competitive developments in AI office software, consumer AI applications and China’s open-source LLM ecosystem.
• US-China policy and export-control developments affecting AI chips, model access, data-center components and robotics.