Strong AI compute demand; AIDC and switch supply chains emerge as the key beneficiaries

Overview

This report focuses on China's AI data centers (AIDC) and their supply chain. It argues that large-model training and inference demand, rising token consumption, and higher domestic cloud capex will push China's AI compute infrastructure into a high-growth phase. AIDC, as the physical infrastructure connecting compute hardware and software, is a key part of the AI compute supply chain; meanwhile, GPUaaS, data center switches, and SuperPOD-related network architecture upgrades will also benefit.

Core views

The core views are as follows: First, China's AI compute demand is expected to be about 5x 2025 levels by 2028e, with intelligent compute scale reaching about 12,354 EFLOPS (@FP16), and AIDC capacity demand reaching about 5.4GW. Second, limited supply of GPUs and AIDC capacity will support GPUaaS pricing and market share for leaders. Third, the SuperPOD cluster architecture raises the technical barriers for power supply, cooling, load-bearing, networking, and delivery at data centers, benefiting specialized AIDC operators such as Range. Fourth, data center switches will benefit from larger GPU clusters, more network tiers, and higher port speeds, and Ruijie is expected to gain market share. Fifth, China's advantages in model cost efficiency, electricity, and construction costs create opportunities for long-term 'token overseas expansion' and the globalization of AIDC operators.

Analysis framework

The report combines a top-down forecast of compute demand, tracking of cloud providers' capex, estimation of AIDC capacity and market size, China-US comparisons for GPUaaS, analysis of SuperPOD technical architecture, and valuation and financial forecasts for key companies to assess investment opportunities in China's AIDC and supply chain.

Methodology notes

• Demand forecastingMapping token consumption to compute demand

  Derive intelligent compute and AIDC capacity demand from the growth in large-model training, inference, and agent token consumption.

  The report uses China's daily token usage, global agent token usage expectations, and cloud providers' capex as demand-side signals to forecast rapid growth in China's intelligent compute and AIDC capacity.

• Market sizingAIDC TAM model

  Estimate market size by combining intelligent compute capacity, mainstream GPU server or compute tray pricing, and Range's data center operating revenue per MW.

  The report expects China's third-party data center market to reach RMB165bn in 2028e, of which the AIDC market is expected to reach RMB75bn, implying a 2025-2028e CAGR of 78%.

• Technology upgrade analysisSuperPOD architecture value-chain analysis

  Assess the impact of GPU cluster network bottlenecks and upgrades in rack power, liquid cooling, load-bearing capacity, power supply, and interconnect bandwidth on AIDC operator value.

  SuperPOD raises the barriers to data center construction and operations, making leaders with energy quotas, networking, liquid cooling, and delivery capabilities more likely to win orders and gain share.

• Relative comparisonChina-US GPUaaS and data center cost comparison

  Compare China and the US in terms of cloud provider capex, GPU supply, electricity prices, construction costs, gross margins, and EBITDA margins.

  China's disadvantage lies in constrained access to high-end GPUs, while its advantage is lower-cost green electricity in western China and lower construction costs; US neo-cloud providers have a more complete tech stack and higher gross margins and mature-stage EBITDA margins.

Asset mapping & comparison

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

• Range (300442 CH)

  AIDC operator and GPUaaS-related beneficiary

  Strengths

  750MW in service, 6,000MW long-term capacity plan, and 3,200MW approved energy quotas; annualized AIDC revenue per MW is about RMB15m, higher than traditional IDC operating services at about RMB9.6m; its domestic data center market share is expected to rise to 9% by 2028e.

  Weaknesses

  AIDC business has not yet deployed SuperPOD solutions at scale, and the business still depends on GPU supply, capex, and customer demand timing.

  Comparison

  Compared with traditional IDC, AI data centers have higher revenue density; compared with some peers, Range has larger energy quotas and a bigger long-term capacity plan.

  Risks

  AI compute demand falls short of expectations, GPU supply remains constrained, construction and delivery are delayed, pricing weakens, or margins come in below expectations.

• Ruijie (301165 CH)

  Beneficiary in China's data center switch supply chain

  Strengths

  Benefits from larger GPU clusters, more network layers, higher port speeds, and demand for scale-up switches driven by SuperPOD; its China data center switch share is expected to rise to 24% in 2028e, while net margin is expected to increase from 5% in 2025 to 10% in 2028e.

  Weaknesses

  Growth depends on the pace of data center switch upgrades and cloud providers' capex, while competition may pressure pricing and margins.

  Comparison

  The report names it the preferred name in the switch segment, together with AIDC operator Range forming an AI compute infrastructure basket.

  Risks

  Slower-than-expected network upgrades, customer concentration, price competition, and slower-than-expected realization of scale economies.

• AIDC operators

  Core beneficiary segment of AI compute demand and SuperPOD architecture upgrades

  Strengths

  AIDC is the physical infrastructure for compute, responsible for data processing and connecting hardware and software; leaders with energy quotas, liquid cooling, high-bandwidth low-latency networks, and delivery capabilities have an edge.

  Weaknesses

  Construction capex is high, and requirements for power supply, cooling, load-bearing capacity, and networking continue to rise.

  Comparison

  Compared with traditional IDC, AIDC requires higher technical and resource intensity, while also offering higher revenue density and stronger customer-demand elasticity.

  Risks

  Capacity expands too quickly, utilization is insufficient, policy and energy-consumption constraints, and changes in technology pathways.

• GPUaaS

  A high-growth business model under tight GPU supply

  Strengths

  Enterprises can access high-end GPUs faster through leasing and reduce self-built capex; price increases indicate strong demand.

  Weaknesses

  Gross margins for China's GPU leasing providers are about 20%-30% and net margins about 10%-15%, below the mature US neo-cloud model; high-end GPU supply is constrained by export controls.

  Comparison

  US neo-cloud providers can reach about 70% gross margin and over 60% EBITDA margin at maturity, though depreciation and amortization pressure is also significant.

  Risks

  GPU prices decline, supply improves and rental prices fall back, uncertainty over a revenue-sharing model transition, and changes in export controls.

• China AIDC overseas expansion and token export

  Long-term incremental opportunity

  Strengths

  China's models have high cost efficiency, and western China benefits from low-cost green electricity and construction costs; growth in overseas token demand may drive Chinese AIDC services for global inference demand.

  Weaknesses

  Cross-border network latency and data compliance are more complex; overseas deployment requires local resources, customers, and regulatory adaptation.

  Comparison

  Model overseas expansion uses overseas compute to reduce latency and compliance friction, while token export trades China's low-cost compute for a cost advantage.

  Risks

  Overseas regulation, restrictions on cross-border data flows, geopolitics, customer demand shifts, and changes in model competition.

Key data

• China intelligent compute data center demand2025-2028e CAGR 74%, about 5.4GW in 2028eDriven by growth in large-model training, inference, and token usage.
• China intelligent compute scaleAbout 12,354 EFLOPS in 2028e (@FP16)Close to 8x the 2025 level.
• China AIDC market sizeAbout RMB75bn in 2028e, 2025-2028e CAGR 78%Estimated based on capacity, compute equipment pricing, and operating revenue per MW.
• China third-party data center market sizeAbout RMB165bn in 2028eIncludes AIDC and related third-party data center demand.
• China cloud provider capexByteDance, Alibaba, and Tencent combined RMB346bn in 2025; over RMB700bn in 2026eMore than double 2025, but only about 16% of capex by US hyperscale cloud providers.
• Data center switch TAMAbout RMB118bn in 2030e, 2025-2030e CAGR 28%Driven by larger GPU clusters, more network layers, higher port speeds, and scale-up switch penetration.
• H100 rental priceRises from RMB12/GPU/hour to RMB17/GPU/hour, about +40%Reflects tight supply-demand conditions and upward pricing in the GPUaaS market.
• Range rating and target priceBuy, target price 111.00 LC, current price 85.18 LC, upside 30%The report expects high growth in its AIDC revenue, with 2025-2028e EBITDA CAGR of 43%.
• Ruijie rating and target priceBuy, target price 89.00 LC, current price 67.80 LC, upside 31%The report expects its China data center switch share to rise from 21% in 2025 to 24% in 2028e.
• Western China green power priceAbout 50% lower than US industrial power pricesOne of the long-term cost advantages for China's AIDC industry.

Impact & implications

The investment implication is that as AI compute demand expands from model training to inference and agent token consumption, the beneficiaries in the industry chain are not limited to GPUs; they also include AIDC operators, GPUaaS providers, switch suppliers, liquid cooling, and power infrastructure. Supply constraints make leaders with capacity reserves, energy quotas, technical delivery capabilities, and customer resources more likely to gain share; meanwhile, if Chinese models continue to win token share overseas, domestic AIDC operators may further expand long-term room by deploying overseas compute or supporting 'token overseas expansion' with China's low-cost compute.

Risks

• Growth in AI training, inference, and agent token usage falls short of expectations, leading to weaker-than-expected AIDC capacity demand and GPUaaS pricing.
• China's access to high-end GPUs is affected by US export controls, which may raise procurement costs and limit the expansion of some AIDC or GPUaaS businesses.
• SuperPOD places high demands on power supply, liquid cooling, load-bearing capacity, networking, and delivery, so construction progress or technical implementation may fall short of expectations.
• If cloud providers' capex slows, it will affect the order pace for AIDC, switches, and GPUaaS.
• If GPUaaS price increases prove unsustainable, leasing providers' revenue growth and margins may be compressed.
• Token overseas expansion faces cross-border network latency, data compliance, overseas regulation, and geopolitical uncertainty.
• Excessively rapid data center supply expansion may lead to localized underutilization or price competition.

What to watch

• Capex execution by Chinese cloud providers such as ByteDance, Alibaba, and Tencent in 2026e and beyond.
• China's daily token usage, agent token usage, and changes in overseas token share for domestic large models.
• H100 and domestic GPU leasing prices, GPUaaS contract durations, and progress in revenue-sharing models.
• Range's AIDC capacity ramp-up, energy quota realization, SuperPOD orders, and changes in revenue per MW.
• Ruijie's data center switch share, port speed upgrades, scale-up switch penetration, and net margin improvement.
• Western China green power prices, data center energy-consumption policy, and progress in the 'east data, west compute / compute coordination' policy.
• US export controls on high-end GPUs and progress in domestic ASIC and high-performance computing chip substitution.
• Progress in overseas projects and customer signings by Chinese AIDC operators.