Jefferies Raises China's Five-Year AI Capital Expenditure Forecast by 23%

Overview

This report is Jefferies' update on China's AI capital expenditure and power demand. The key change is a 23% increase in cumulative AI capital expenditure for 2026E-2030E to approximately US$1tn, and a 31% increase in cumulative AI power demand over the same period to 34.2GW. The upgrades reflect the potential interim approval of H200 imports, improved domestic AI chip output, a slight increase in the informal import assumption, and growth in general server demand driven by agentic AI.

Core views

The report believes that China's AI industry remains in a high-investment phase. In the short term, external chips such as the H200 will be needed to alleviate inference-side chip shortages, while China will continue to pursue full-chain localization over the long term. Ascend 950 improves output and computing power through a smaller die and higher frequency, but also increases power consumption, making power and IDC demand relatively high-certainty beneficiaries. Semiconductor capital expenditure is expected to remain elevated, domestic WFE share should rise steadily, and AMEC, SMIC, and VNET are identified as the top picks.

Analysis framework

The report updates its five-year forecasts across multiple dimensions, including AI chip supply, advanced-node capacity and yields, GPU power consumption, server demand, IDC capital expenditure, and power demand. It incorporates H200 import assumptions, changes in domestic chip design, the upper limit of 7nm yields, wafer capacity constraints, and expanding application demand into its assessment.

Methodology notes

• Industry Chain AnalysisAI Infrastructure Capital Expenditure Framework

  Linkage among chip supply, wafer capacity, servers, IDC, and power demand

  The report maps growth in AI inference demand to chip procurement, domestic chip output, server deployment, IDC construction, and power demand, thereby assessing capital expenditure and beneficiary stocks.

• Valuation MethodologyP/E, P/B, SOTP

  Using different valuation multiples or segment-based valuation according to company characteristics

  AMEC's target price is based on 2026E/2027E P/E, NVIDIA's target price is based on a P/E multiple of CY28E EPS, SMIC's target price is based on 2026E P/B, and VNET's target price is based on SOTP.

Asset mapping & comparison

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

• Advanced Micro-Fabrication Eqp Inc China

  Beneficiary of domestic wafer fabrication equipment demand

  Strengths

  Benefits from sustained high semiconductor capital expenditure in China and rising domestic WFE share; identified as a top pick, with 688012 CH disclosed as BUY.

  Weaknesses

  Customer expansion, new customer acquisition, and the pace of new tool development may affect earnings realization.

  Comparison

  Compared with international equipment suppliers, the domestic substitution thesis is stronger, but the company is also more exposed to technological iteration and the sanctions environment.

  Risks

  Further sanctions by the US Department of Commerce, customer expansion slower than expected, overseas new customer acquisition slower than expected, and new tool development slower than expected.

• Semiconductor Manufacturing International Corporation

  Beneficiary of China's advanced-node manufacturing and AI chip foundry demand

  Strengths

  Benefits from rising domestic AI chip output and domestic substitution demand; identified as a top pick, with 981 HK disclosed as BUY.

  Weaknesses

  7nm large-chip yields, equipment supply, and wafer capacity remain long-term constraints.

  Comparison

  Compared with mature-node operations, AI GPU-related demand has greater elasticity, but advanced nodes face more pronounced equipment and yield constraints.

  Risks

  Continued US issuance of licenses to US SPE companies for mature-node equipment could reduce supply-chain tightness; stronger-than-expected demand for mature nodes such as 5G, IoT, and CIS could also alter ASPs and capacity allocation.

• VNET Group

  Beneficiary of IDC and power demand

  Strengths

  Rising AI chip power consumption, H200 purchase assumptions, and agentic AI server demand are jointly driving power and IDC demand; identified as a top pick.

  Weaknesses

  IDC expansion depends on power resources, government quotas, and the pricing environment.

  Comparison

  Compared with the chip manufacturing chain, VNET has more direct exposure to data center and power demand following AI compute deployment.

  Risks

  Demand growth slower than expected, insufficient government power quotas, and intensified competition resulting in lower-than-expected pricing.

• NVIDIA Corporation

  Related to short-term H200 supply and China's AI inference demand

  Strengths

  The report assumes that Chinese AI players may continue purchasing H200 chips in 2026 and 2027 to alleviate inference chip shortages; NVDA is disclosed as BUY.

  Weaknesses

  Exports to China are significantly affected by US licensing and technology restrictions, while China will continue to promote localization over the long term.

  Comparison

  The H200 can serve as a short-term supplement, but the domestic Ascend route is China's long-term strategic direction.

  Risks

  Competition from INTC, AMD, and hyperscaler-developed ASICs; slower data center capital expenditure by enterprises and hyperscalers; automotive platform ramp-up slower than expected; and US export restrictions.

Key data

• 2026E-2030E Cumulative AI Capital ExpenditureUS$1tnRaised 23% from the previous forecast.
• 2026E-2030E Cumulative AI Power Demand34.2GWRaised 31% from the previous forecast.
• H200 Purchase Assumption300K in 2026, 100K in 2027A short-term supplemental assumption amid strong growth in China's AI inference demand and chip shortages.
• Ascend 950 Die AreaApproximately 400 sqmmApproximately one-third smaller than the Ascend 910, theoretically increasing the number of chips produced per wafer.
• Ascend 950 Power Consumption and Computing PowerApproximately 600W, approximately 1,000 TFLOPsPower consumption is higher than the estimated 500W for Ascend 910, while computing power is significantly higher than the estimated 375 TFLOPs for 910C.
• 7nm Large-Chip Yield AssumptionUp to approximately 35%The report believes that long-term development remains constrained by 7nm yields, wafer capacity, and equipment.
• SMIC 2030 Wafer Capacity Assumption43kwpmJefferies' estimate, affected by equipment constraints.

Impact & implications

If the report's assumptions are realized, China's AI infrastructure investment will simultaneously benefit the semiconductor equipment, wafer foundry, AI server, IDC, and power-related value chains. H200 imports would ease short-term inference bottlenecks, but localization remains the long-term theme; rising power consumption further increases the importance of IDC power resources, power supply approvals, and rack demand.

Risks

• Further escalation of US export controls or Department of Commerce sanctions could affect the supply of equipment, HBM, and high-end AI chips.
• Uncertainty remains high regarding US licenses for products such as the H200, with resistance potentially greater ahead of the US midterm elections.
• China's domestic 7nm large-chip yields, equipment supply, and wafer capacity could fall below expectations.
• If AI inference demand or agentic AI server demand is lower than expected, capital expenditure and IDC demand would weaken.
• IDC projects may be constrained by government power allocation, electricity-use quotas, and competitive pricing pressure.
• The valuations of related stocks depend on relatively high growth assumptions; slower-than-expected customer expansion or new tool development could hinder target-price realization.

What to watch

• Whether the US further approves or restricts shipments of H200 and related HBM3e-integrated products to China.
• Changes in GPU rental prices among Chinese AI players, inference demand, and the severity of chip shortages.
• The actual mass production, yields, power consumption, and computing performance of Huawei Ascend 950.
• SMIC's 7nm capacity, yields, and equipment availability.
• Power quotas, rack-up rates, pricing, and validation of 2026-2027 demand for Chinese IDC projects.
• The pace at which domestic WFE, materials, and chemicals supply-chain share increases.