The $295 Billion Bet: How China's National AI Compute Plan Reshapes Global Capex
China's $295B, 5-year plan for a unified AI network—operated by state telecoms and 80% supplied by domestic chips—locks Nvidia/AMD out, signaling global AI capex is entering its heavy-infrastructure phase. What A-share and US semiconductor investors need to know now.
A structural reading of Bloomberg's June 9 scoop — and what it means for A-shares, US semiconductors, and the next phase of the AI infrastructure cycle.
Lead
On June 9, 2026, Bloomberg reported that China is preparing to spend approximately RMB 2 trillion ($295 billion) over the next five years on building a nationwide network of AI data centers — drafted by the National Development and Reform Commission (NDRC), operated primarily by state telecoms (China Mobile, China Telecom), and supplied at least 80% by domestic technology firms led by Huawei.
The total investment, when integrated with power grid upgrades, could reach RMB 5 trillion (~$735 billion).
This is not a routine policy announcement. It is the most explicit signal yet that China has formally entered the heavy-infrastructure phase of the AI capex cycle — the same phase the US entered with OpenAI's Stargate, xAI's Colossus, and Meta's Hyperion.
Why It Matters
Three structural shifts are embedded in this announcement:
1. The Sino-US AI compute decoupling is now state-directed, not market-led. Until June 2026, Chinese AI infrastructure was primarily built by private hyperscalers (Alibaba, Tencent, ByteDance) using a mix of imported (Nvidia H100/H200) and domestic (Huawei Ascend) chips. The NDRC blueprint formalizes a different model: state-owned operators, sovereign-debt funding, and an 80% domestic chip mandate. This is no longer a private capex story — it is industrial policy at the scale of the original Made in China 2025 program.
2. Nvidia's China optionality is structurally lower than the market currently prices. Washington's recent decision to allow H200 sales to China was widely interpreted as a thaw. But the NDRC's 80% domestic mandate effectively caps the addressable market for Nvidia and AMD at ~20% of Chinese AI compute. Even if export controls fully relax, Nvidia's China revenue ceiling has been redrawn. Jensen Huang's May 21 acknowledgment that Nvidia has "largely conceded" the China market now reads as a leading indicator, not a temporary setback.
3. The "scattered to unified" architecture is the underappreciated story. The plan's stated goal is to interconnect China's currently fragmented regional compute hubs (Wulanchabu, Guizhou, Ningxia, Hangzhou) into a single national network by 2028. This is functionally analogous to building a national power grid for compute — and historically, when China has executed this kind of unified infrastructure (national high-speed rail, national 5G), the resulting cost curves and deployment speeds outperformed Western analyst expectations by 2-5x.
The Numbers in Context
To calibrate the scale of this commitment:
Two observations stand out:
- Per-year US private AI capex ($725B) still dwarfs China's state-directed plan ($59B/year average). The real fight is between US private capital + sovereign infra (Stargate) versus China private + state. On a per-year basis, US still spends 4-5x more.
- But China's plan is integrated and unified. A single $295B program coordinated by NDRC + China Mobile + China Telecom + Huawei has lower coordination costs than fragmented $725B across competing US hyperscalers.
The comparison is not "who spends more," but "who deploys more compute per dollar."
Investment Implications
This announcement creates three layers of investable consequences:
Layer 1 — A-share infrastructure suppliers (the clearest beneficiaries)
Domestic firms in the AI infrastructure value chain face a multi-year, state-backed demand cycle:
- AI chips: Cambricon (688256.SH), Hygon (688041.SH) — alongside non-listed Huawei
- Server / system integration: Inspur (000977.SZ), Sugon (603019.SH), H3C (via Unisplendour 000938.SZ)
- Liquid cooling: Envicool (002837.SZ), Yinlun Machinery (002126.SZ)
- Power transformers and HVDC: TBEA (600089.SH), XJ Electric (000400.SZ), NR Electric (300831.SZ)
- Data center construction: Aofei Data (300738.SZ), GDS Holdings (GDS)
- State telecoms (operators): China Mobile (600941.SH / 0941.HK), China Telecom (601728.SH / 0728.HK)
- Power generation (the bottleneck): China National Nuclear Power (601985.SH), Yangtze Power (600900.SH)
Layer 2 — US semiconductor exposure (re-rating risk)
- Nvidia (NVDA): China revenue ceiling re-priced lower; thesis shifts from "China growth optionality" to "China is structurally capped"
- AMD (AMD): Same dynamic, lower exposure
- TSMC (TSM): Less direct impact (still benefits from Huawei via SMIC, even if at lower nodes)
- Broadcom (AVGO): Networking/switching demand from Stargate-equivalents partially offsets China loss
Layer 3 — Global power and grid plays
The bottleneck is global. Whether AI compute is being built in Wulanchabu or Memphis, the binding constraint is the same:
- Nuclear / SMR: Cameco (CCJ), Constellation Energy (CEG), Vistra (VST)
- Grid equipment: GE Vernova (GEV)
- Specialty natural gas: Williams Companies (WMB)
What to Watch
Three signals to monitor over the next 6-12 months:
- NDRC's formal blueprint release. Bloomberg's reporting indicates the plan is in "early discussion stage." A formal NDRC document — likely tied to the next Five-Year Plan ratification — will provide concrete provincial allocation, funding sequence, and procurement specifications.
- Provincial power quota allocations. Inner Mongolia, Ningxia, Guizhou, and Sichuan are the most likely hosts for GW-scale AI compute. Watch for NDRC bulletins on AI data center power priority.
- Huawei Ascend chip production ramp. The 80% domestic mandate is feasible only if Huawei (and SMIC as its foundry) can meaningfully scale Ascend 910C / 910D production. SMIC's quarterly capex disclosures and Huawei's annual report (typically March) will be the cleanest signals.
Closing
The first phase of the AI capex cycle (2022-2025) was characterized by a single trade: buy Nvidia. That trade enriched a generation of investors and concentrated AI exposure in one stock.
The second phase (2026 onward) is geographically and structurally bifurcated:
- US side: private-led, hyperscaler-dominated, Nvidia-centered, energy-bottlenecked
- China side: state-directed, telecom-operated, Huawei-supplied, grid-integrated
Investors who continue to hold AI exposure exclusively through US hyperscalers and Nvidia are running a structurally concentrated bet. The China side of the trade — A-share infrastructure suppliers, state telecoms, domestic chip leaders — is now publicly anchored at a $295 billion floor.
The trade is no longer "buy Nvidia." It is "build the substations, regardless of which flag flies above them."
