Goldman Sachs: China's AI Capex to Reach RMB 8.5 Trillion Over Five Years, Computing Capacity to Triple
nashnova research
Goldman Sachs projects China's AI-related capital spending at RMB 8.5 trillion (~$1.3 trillion) from 2026 to 2030, driving national IT power capacity from 26GW to 81GW — a signal that China is shifting AI from a chip-scramble phase into a full-scale infrastructure buildout that will reprice the entire supply chain.
Where does the RMB 8.5 trillion go?
Hyperscale cloud operators — Alibaba, Tencent, Baidu and peers — account for roughly RMB 6.3 trillion over five years, nearly three-quarters of the total. Their 2026 year-on-year growth rate is forecast at 121%, far above the 41% Goldman expects from US peers.
Next-gen cloud operators — newer platforms led by ByteDance — add about RMB 1.5 trillion, mostly spent on server procurement.
Telecom operators' computing capex rises from RMB 81 billion in 2026 to RMB 179 billion by 2030, a five-year CAGR of roughly 10%.
This means → the bulk of China's AI spending is a bet on the top-tier cloud platforms; telecoms play a steady but secondary role. Total Chinese capex averages about 17% of US peers — the gap remains wide.
Is the bottleneck shifting from demand to supply?
Goldman notes that China's AI computing expansion has so far been constrained by high-end chip supply, not by demand. In plain terms = it's not that nobody wants computing power — chips and servers simply can't keep up.
As domestic GPUs — graphics processors, the core chips for AI training — and ASICs — purpose-built accelerator chips — begin to ship at scale, this bottleneck should gradually ease.
Hyperscalers are expected to add 34GW from 2025 to 2030; telecoms add 9GW of externally available capacity; other enterprises and AI labs add roughly 12GW.
This reflects a narrative shift: China's AI story is moving from "explosive demand" to "can supply keep up?" — whoever solves the supply side first captures the incremental growth.
Where does the money come from? Offshore financing is the key variable
Onshore funding is relatively comfortable. Hyperscalers lean on operating cash flow, cash reserves, and leasing arrangements.
The bigger constraint is offshore. Overseas capex totals roughly RMB 2.0 trillion over five years, driven largely by ByteDance. Potential funding sources include portfolio monetization, debt issuance, and equity financing.
Next-gen cloud operators use a more fragmented mix: finance leases at ~5% covering 30–40% of investment + bank loans at ~3–4% + customer prepayments + new debt or equity.
This means → such structures support rapid scaling but make these companies highly sensitive to utilization rates, delivery timelines, and refinancing costs. If utilization falls short, financial pressure escalates fast.
Does return on investment rise the closer you get to the application layer?
Goldman's ROIC framework shows: GPU-based IaaS — infrastructure-as-a-service — carries a base-case operating margin of ~44%, ROIC of ~13%, and a cash payback period of about three years.
Self-built compute + in-house model: margin rises to ~53%, ROIC to ~19%, payback shortens to ~2.5 years.
Self-built compute + third-party API: margin reaches ~57%, ROIC hits ~29%, payback drops to ~2 years.
In plain terms = selling raw computing power earns the least. Stack a proprietary model or application on top and margins climb sharply — the closer to the end user, the faster the money comes back.
Can domestic chips handle full inference?
Goldman estimates that next-generation Chinese chips paired with HBM3E — third-generation high-bandwidth memory — can run a complete inference pipeline. Base-case operating margin: 31%; ROIC: ~9%; payback: ~3.5 years.
This means → the fully domestic stack still trails overseas server configurations, but it can already run the full workflow. As localization advances, margins have room to improve.
On the GPU localization theme, Goldman favors Cambricon, Moore Threads (沐曦), and Hygon (海光信息).
Which names is Goldman backing?
Cloud operators: overweight on Alibaba, Kingsoft Cloud, and VNET Group (世纪互联), citing strong computing demand and supply-chain easing.
AI model companies: overweight on MiniMax and Zhipu (智谱).
Data centers: bullish on VNET Group's first-mover positioning in remote hub regions such as Ulanqab.
Goldman flags two key checkpoints: whether domestic chips ship at expected volumes + whether offshore financing for overseas capex materializes. These two items determine whether the investment thesis plays out.
市场有风险，内容仅供研究参考，不构成投资建议。