Morgan Stanley sees Rmb8.5tr of China AI capex through 2030 driving IT capacity to 81GW

Overview

This industry deep dive maps China’s AI investment path from capex and chip supply through compute capacity, financing and unit economics. Morgan Stanley forecasts Rmb8.5tr of total capex in 2026-30E and argues that domestic AI-chip ramp-up is the key catalyst for capacity growth and for selected cloud, data-center, model and semiconductor beneficiaries.

Core views

Morgan Stanley forecasts China AI-related capex of Rmb8.5tr (US$1.3tr) in 2026-30E, or about 17% of US-peer spending on average. It expects total China compute capex to rise from Rmb1.3tr in 2026E to Rmb2.0tr by 2030E. Hyperscaler and key-internet capex is forecast at Rmb995bn in 2026E, up 121% year on year, then Rmb1.2tr in 2027E and Rmb1.4tr in 2030E. Neocloud investment is projected at Rmb1.5tr over five years, largely for server procurement, while telco compute capex rises from roughly Rmb81bn in 2026E to Rmb179bn by 2030E. The report expects 70-80% of five-year spending, or Rmb6.4tr, to be domestic. The report’s central supply-side conclusion is that China’s expansion has been constrained more by high-end compute availability than demand. Domestic GPU and ASIC supply is therefore expected to unlock the next capacity leg: China IT power capacity rises from 26GW in 2025 to 81GW in 2030E. Morgan Stanley attributes the 55GW increase to hyperscalers (+34GW), telcos’ self-built external capacity (+9GW), and other players and AI labs (+12GW). Alibaba and Tencent are each expected to add 6-9GW, while Baidu adds about 1GW. Domestic AI-chip shipments are forecast to increase from 1.1mn in 2025 to 2.4mn in 2026E, 4.8mn in 2027E and 11mn in 2030E, expanding domestic-chip TAM from Rmb94bn to Rmb646bn; domestic chips are expected to represent 70-85% of server deployment through 2025-30E. The spending mix is compute-heavy. Morgan Stanley estimates total server spending of Rmb5.3tr, around 88% of compute capex, including Rmb4.9tr of AI-server expenditure. It assumes 30-35% of hyperscaler capex is overseas, 10-15% of domestic capex goes to non-server items, and about 10% of server capex goes to CPUs, with the remainder directed to GPU-related computing. The report expects incremental AI power demand to rise from 5.6GW in 2026E to roughly 13GW in 2030E, and third-party data-center orders to rise from 6.1GW to 11GW. It assumes 20% of hyperscaler and AI-lab capacity is self-built; western regions are expected to account for 33GW of five-year incremental third-party orders versus 14GW in eastern regions. Funding capacity differs by participant. Morgan Stanley considers Alibaba, Tencent and Baidu broadly well funded for domestic capex through operating cash flow, cash balances, leases and prepayments, although Alibaba faces potential cash pressure if food-delivery competition intensifies and Baidu depends on search stabilization. Tencent is described as the most stable of the three. The larger issue is offshore funding: four hyperscalers are projected to spend Rmb2.0tr on offshore expansion in 2026-30E, where capex, debt repayments and shareholder returns compete for limited offshore cash. The report identifies investment-portfolio monetization, debt and equity as potential sources. Neoclouds can combine 30-40% financing leases at around 5% interest, bank loans at roughly 3-4%, customer prepayments and debt or equity issuance, but this leverage makes them sensitive to utilization, delivery schedules, refinancing costs and high-end-server demand. For Kingsoft Cloud, Morgan Stanley models 40% of annual investment financed through five-year leases, while 2Q26 operating cash flow of Rmb2bn was supported by a large customer prepayment. Carrier-neutral data-center vendors are projected to need Rmb569bn of capex over five years; project loans, operating cash flow, REITs, private ABS, private credit and equity are possible funding channels. Telcos, by contrast, are expected to self-fund most compute investment through cash, annual operating cash flow and lower traditional capex. The report emphasizes that returns depend on the monetization layer, not simply capex volume. In its base case, self-owned GPU IaaS produces a 44% operating margin, 13% ROIC and roughly three-year cash payback, based on Rmb250k monthly server rent, 75% utilization and Rmb8mn server capex. Rented/neocloud IaaS produces around a 20% operating margin at Rmb250k customer pricing against Rmb200k monthly rental cost, but needs no upfront server capex and can generate immediate positive cash once the spread is positive. A self-built MaaS model, assuming 4,000 tokens per second per GPU, 50% inference mix and Rmb9.5 per million blended tokens, produces 53% operating margin, 19% ROIC and 2.5-2.6-year payback. Self-built compute hosting a third-party API, under a 10% revenue share, is estimated at 57% operating margin, 29% ROIC and 2.1-year payback; performance discount versus an in-house model is the key uncertainty. Domestic chips broaden capacity but initially lower unit economics relative to foreign servers. Morgan Stanley’s base case for full domestic infrastructure with next-generation HBM3E-equipped chips assumes 30% performance versus foreign servers, Rmb4mn server cost, 31% incremental EBIT margin, 9.1% ROIC and 3.6-year payback. Under 20-40% relative compute efficiency, its sensitivity range is 0.3%-17.9% ROIC and 2.7-5.2 years of payback. Prefill/decode disaggregation can improve ROIC for long-input workloads such as summarization, long-document Q&A and coding assistance, but is less beneficial for balanced 4k/4k input-output workloads. The report believes improved memory bandwidth in next-generation domestic chips could remove a key bottleneck and improve localization economics over time. Morgan Stanley continues to prefer the China cloud supply chain, citing strong compute demand and supply-chain relief. It remains Overweight on Alibaba, Kingsoft Cloud and VNET among infrastructure providers; MiniMax and Zhipu among AI labs; and Cambricon, Illuvatar and Hygon in GPU localization.

Analysis framework

Morgan Stanley builds a top-down five-year capex forecast across hyperscalers, neoclouds and telcos, then translates projected chip shipments and deployment costs into gigawatts of compute capacity. It assesses funding sources and cash-flow constraints by ecosystem participant, and uses per-server unit-economics scenarios to compare IaaS, MaaS, third-party API hosting and domestic-chip deployment returns.

Methodology notes

• Industry AnalysisSupply-demand framework

  AI compute supply-demand and capacity build-out analysis

  The report links projected capex, domestic chip shipments, deployment costs and power demand to compute-capacity additions, identifying chip availability as the binding supply constraint.

• Corporate Fundamentals and FinanceROIC–WACC spread

  ROIC and cash-payback unit-economics analysis

  The report compares operating margins, ROIC and cash payback across different infrastructure ownership and AI-monetization models.

• Industry AnalysisUpstream-Midstream-Downstream Transmission

  AI infrastructure ecosystem financing and demand transmission

  The report traces how hyperscaler capex, prepayments, contracts and investments support neoclouds, data centers, chip vendors and AI labs.

Asset mapping & comparison

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

• Alibaba

  Preferred cloud infrastructure provider and major expected capacity contributor

  Strengths

  Clearer ROIC visibility through AI GPU IaaS and MaaS; expected 6-9GW capacity addition; June-quarter MaaS ARR surpassed Rmb10bn.

  Comparison

  Expected to spend more than Tencent because of clearer AI monetization visibility.

  Risks

  Cash drain could increase if food-delivery competition and subsidies re-emerge.

• Kingsoft Cloud

  Preferred neocloud provider and financing case study

  Strengths

  Large-customer support and customer prepayments support financing; Morgan Stanley forecasts Rmb15bn investment in 2026E.

  Weaknesses

  Capital-intensive expansion requires leases, loans and potential additional debt or equity.

  Comparison

  Its financing practices are considered more durable than those of the broader neocloud industry.

  Risks

  Delivery slowdown, rising interest costs and weaker high-end-server demand.

• VNET

  Preferred data-center beneficiary

  Strengths

  Morgan Stanley highlights its rural-area hub development, including Ulanqab in Inner Mongolia.

  Risks

  Data-center financing and customer ramp-up delays can pressure returns.

• MiniMax and Zhipu

  Preferred AI-lab beneficiaries of compute growth

  Strengths

  Greater compute availability is expected to support model-player growth.

  Weaknesses

  Self-building capacity can add material upfront-capex burden.

  Risks

  Infrastructure self-build and current domestic-chip decoding limits can impair returns.

• Cambricon, Illuvatar and Hygon

  Preferred GPU-localization beneficiaries

  Strengths

  Domestic chip supply is expected to unlock capacity growth and gain a large share of server deployment.

  Weaknesses

  Near-term performance remains below foreign servers in Morgan Stanley’s modeled economics.

  Comparison

  Domestic deployment cost and infrastructure economics differ from US counterparts.

  Risks

  Domestic-chip performance and memory-bandwidth limitations could result in low or negative ROIC in weak cases.

Key data

• Total China AI capexRmb8.5tr (US$1.3tr) in 2026-30EEquivalent to about 17% of US-peer spending on average.
• China IT power capacity26GW in 2025 to 81GW in 2030EA 55GW expansion led by hyperscalers, telcos and other players.
• Domestic AI-chip shipments1.1mn in 2025 to 11mn in 2030EDomestic chip TAM is forecast to rise from Rmb94bn to Rmb646bn.
• Hyperscaler capexRmb995bn in 2026E; Rmb1.4tr in 2030E2026E growth is forecast at 121% year on year.
• Offshore hyperscaler capexRmb2.0tr in 2026-30EMorgan Stanley identifies offshore funding as a key financing consideration.
• Self-owned GPU IaaS base case44% operating margin, 13% ROIC, ~3-year paybackBased on Rmb250k monthly rent, 75% utilization and Rmb8mn server capex.
• Self-built MaaS base case53% operating margin, 19% ROIC, 2.5-2.6-year paybackAssumes 4,000 tokens/sec/GPU, 50% inference mix and Rmb9.5/mn blended token pricing.

Impact & implications

The report argues that domestic-chip ramp-up can convert strong AI demand into faster compute deployment, benefiting China’s cloud and infrastructure ecosystem. It sees the most attractive exposure in selected cloud providers, data centers, AI labs and GPU-localization names, while highlighting that financing structure, utilization and monetization efficiency determine whether capex translates into attractive returns.

Risks

• Offshore capex, debt repayments and shareholder returns may create funding pressure for hyperscalers.
• Neocloud leverage creates sensitivity to utilization, delivery timing, refinancing costs and high-end-server demand.
• Alibaba’s cash position could be pressured by renewed food-delivery subsidies, while Baidu depends on search-revenue stabilization.
• Current domestic chips may be inefficient for LLM decoding, potentially producing low or negative ROIC in unfavorable deployment cases.
• AI-lab self-build plans can increase upfront capital needs, while data-center vendors remain cautious about AI-lab cash-flow stability.

What to watch

• The pace of domestic GPU and ASIC supply ramp-up, including next-generation HBM3E-equipped chips.
• Hyperscaler capex execution, including Alibaba, Tencent and Baidu capacity additions.
• Offshore financing, portfolio monetization, debt issuance and shareholder-return decisions by hyperscalers.
• Neocloud utilization, server procurement prices, lease costs and customer prepayments.
• Inference mix, token throughput, pricing and performance discounts, which drive MaaS and API-hosting ROIC.