China's $295 Billion AI Infrastructure Plan

TL;DR

• China is preparing a CNY2 trillion ($295B) plan to build a nationwide network of interconnected AI data centres over five years, with state-owned telcos China Mobile and China Telecom as operators.
• At least 80% of the technology — including AI chips — must come from domestic suppliers, principally Huawei, effectively locking Nvidia and AMD out of the largest AI infrastructure buildout in history.
• The plan is part of Beijing's "six-network" national infrastructure strategy, placing AI compute alongside water, electricity, and telecoms as state-guaranteed utilities.
• Nvidia's Jensen Huang has already conceded China's AI chip market to Huawei. This plan makes that concession structural and permanent.
• The blueprint remains in early discussions and details could change. But the direction is set.

What happened

On 9 June, Bloomberg reported that China is drafting a plan to spend approximately CNY2 trillion ($295 billion) over the next five years on a nationwide network of AI-focused data centres and interconnected computing hubs. The National Development and Reform Commission (NDRC) — China's top economic planning agency — is leading the blueprint, according to people familiar with the matter. ¹

The architecture is specific. State-owned telecom operators China Mobile and China Telecom would operate the bulk of the infrastructure and ensure the computing hubs are connected. At least 80% of the technology deployed — AI chips, networking equipment, storage — would come from domestic suppliers, with Huawei Technologies positioned as the primary beneficiary. ²

This is not a standalone initiative. Omdia analyst Guang Yang confirmed to RCR Wireless on 17 June that the plan is part of China's "six-network" national infrastructure strategy, announced by the country's top leadership in late April. The six networks include water supply, electricity, next-generation telecoms, and — now — a national computing network. AI compute is being treated as a state-guaranteed utility. ³

The plan's goal, Yang said, "is not only to build AI data centres but also to achieve efficient coordination among computing, networks, and electricity to match AI demand with computing resources and cost-effective energy supply." That means routing cost-sensitive AI workloads to data centres with cheap renewable power and latency-sensitive workloads to edge facilities — a national load-balancing architecture for AI.

What it actually means

This is a decoupling architecture, not a spending programme

The $295 billion figure grabs attention. But the number is not the story. The story is the architecture of exclusion embedded in the plan.

Three mechanisms lock foreign suppliers out:

1. The 80% domestic procurement rule. Huawei and domestic chip designers get a guaranteed market. Nvidia and AMD do not. This formalises what export controls began — Jensen Huang told CNBC in May that Nvidia has "largely conceded" China's AI chip market to Huawei. ⁴ This plan makes that concession permanent.

2. State-owned operators as gatekeepers. China Mobile and China Telecom are not commercial hyperscalers making build-vs-buy decisions. They are instruments of state policy. When Beijing tells them to buy domestic, they buy domestic. The plan shifts AI infrastructure from a market governed by price and performance to one governed by procurement directives.

3. The "six-network" framing. By placing AI compute alongside water and electricity, Beijing is signalling that AI infrastructure is critical national infrastructure — not a commercial sector where foreign competition is welcome. The regulatory and security apparatus that applies to power grids and water systems will apply to AI data centres.

The telco angle is the underappreciated story

The plan positions state-owned telcos to take a central role in China's AI ecosystem — a role currently dominated by private hyperscalers Alibaba, ByteDance, and Tencent. ³

This is a structural shift. China's AI market has been shaped by Alibaba Cloud, Tencent Cloud, and ByteDance's internal infrastructure. The state-owned telcos have been infrastructure providers, not AI platform operators. This plan changes that.

Yang noted that Chinese telcos have already been deploying AI infrastructure based on domestic chips, but utilisation has lagged behind competitors using Nvidia hardware. "It is complex and time-consuming for customers to migrate their AI workloads from the CUDA ecosystem to homegrown Chinese systems," he said. ³

The plan's solution: use government-funded sectors — healthcare, transportation, smart cities — as anchor tenants. "The first step should be to leverage demand from these sectors to fill in telcos' AI data centres, creating a closed loop from investment to revenue." ³

This is state-directed demand creation. The government funds the infrastructure, mandates domestic chips, and then directs state-funded sectors to use it. The market doesn't need to choose Huawei over Nvidia — the choice has already been made.

The token economy layer

Chinese telcos are already building the commercial layer on top of this infrastructure. China Mobile, China Telecom, and China Unicom have all launched AI token plans — billing models that shift from mobile data quotas to AI usage credits. ⁵

China Telecom's chairman Ke Ruiwen has said the company plans to put tokens at the centre of its entire business. Daily AI token usage in China reached approximately 140 trillion tokens in March 2026, according to China's National Bureau of Statistics. ⁵

The infrastructure plan and the token economy are two halves of the same strategy: build the compute, control the access, bill the usage. It is a vertically integrated AI utility model.

Hype deconstruction

This is not a done deal. The Bloomberg report explicitly notes the blueprint "remains in early discussions and details could change." ¹ China Mobile, China Telecom, and the NDRC did not respond to Reuters' requests for comment. ² The plan has not been formally announced or funded.

The $295 billion is aspirational, not appropriated. China's five-year plans set direction; they do not guarantee execution. The actual capital deployed will depend on fiscal conditions, technology readiness, and the political calendar.

The CUDA migration problem is real. Yang's point about the difficulty of migrating workloads from Nvidia's CUDA ecosystem to domestic alternatives is not a footnote — it is the central technical challenge. Huawei's Ascend chips have demonstrated inference capability and even some training capability (a joint Huawei-DeepSeek team successfully used Ascend 910C chips for post-training of DeepSeek-V4-Pro ⁶), but the software ecosystem gap remains substantial.

This is not a China-only phenomenon. The US is spending comparable amounts. Big Tech companies are expected to spend more than $700 billion on AI infrastructure this year alone. ² JPMorgan estimates the total debt financing component of the global AI capex buildout at $4.1 trillion. ⁷ China's plan is large, but it is not uniquely large — it is the mirror image of what is happening in the US, with the critical difference that China's buildout is state-directed and domestically sourced.

Stakeholder landscape

Cross-layer implications

Semiconductor layer: The 80% domestic rule, combined with the scale of the plan, creates a demand signal that will accelerate Huawei's Ascend roadmap. Huawei has already announced the Ascend 960 (2027) and Ascend 970 (2028), and introduced a new "Tau Scaling Law" for chip development targeting 1.4nm-equivalent transistor density by 2031. ⁸ Guaranteed demand at this scale shortens the feedback loop between deployment and iteration.

Energy layer: Yang's emphasis on matching AI workloads to renewable energy supply is not incidental. China's western provinces have abundant renewable energy and cheap land but are far from population centres. The plan's "computing, networks, and electricity" coordination framework is effectively a national scheme to locate AI data centres where power is cheapest and route workloads over high-bandwidth networks.

Geopolitical layer: The plan lands in the same month that the White House forced Anthropic to shut down Fable 5 and Mythos 5 on national security grounds, and as G7 leaders negotiate "trusted partner" access to US AI models. The world is bifurcating into two AI infrastructure stacks — one built on Nvidia/CUDA with US export controls, one built on Huawei/Ascend with Chinese procurement mandates. The $295 billion plan is China's answer to the question: "What if we can't buy their chips?"

Commercial layer: The token economy model — AI usage credits sold by telcos — is a billing innovation that has no direct Western equivalent. If it works at scale, it creates a new commercial pathway for AI infrastructure that doesn't depend on hyperscaler cloud margins.

What this means for you

For AI practitioners and engineers: If you build or deploy AI systems that might ever touch the Chinese market, the CUDA-to-Ascend migration is no longer a theoretical exercise. Start understanding the Huawei Ascend software stack (CANN, MindSpore). The migration tax is real, but the direction of travel is clear.

For infrastructure investors and strategists: The plan signals that AI compute is being reclassified as critical national infrastructure. This has implications for regulation, procurement, and competition policy far beyond China. Watch for similar "AI as utility" framings in the EU, India, and Japan.

For enterprise technology buyers: If your supply chain includes AI hardware or cloud services that depend on Nvidia GPUs manufactured in or routed through Asia, map your exposure. The bifurcation of the global AI chip supply chain is accelerating.

For policymakers and analysts: The plan is evidence that export controls are a double-edged instrument. They deny China access to cutting-edge US chips, but they also guarantee demand for China's domestic alternatives, accelerating their development. The question is not whether export controls "work" — it is what kind of AI ecosystem they produce on the other side.

For everyone else: There is nothing actionable in this story for the general public today. But the direction it points to — AI compute treated as a state-guaranteed utility, with all the regulatory and geopolitical baggage that implies — will shape the AI products and services available to everyone within five years.

Uncertainty ledger

• Plan status: The blueprint is in early discussions. No formal announcement, no appropriated funding. Details could change materially.
• CUDA migration feasibility: Whether Huawei's software ecosystem can close the gap with CUDA at scale remains unproven. The DeepSeek post-training success is a data point, not proof of ecosystem maturity.
• Telco execution capability: State-owned telcos have not historically been AI platform operators. Their ability to execute at the scale and speed required is uncertain.
• Utilisation risk: Yang flagged that Chinese telcos' AI data centres already show lower utilisation than competitors using Nvidia chips. The plan's anchor-tenant strategy (government healthcare, transport, smart cities) may not generate sufficient demand to fill capacity.
• US policy response: The plan could trigger further US export controls on semiconductor manufacturing equipment, which would constrain Huawei's ability to produce Ascend chips at scale.

Bottom Line

China's $295 billion AI infrastructure plan is not primarily about spending. It is about architecture. By placing AI compute inside a state-guaranteed utility framework, mandating domestic chips, and routing demand through state-owned telcos, Beijing is building a parallel AI infrastructure stack that does not depend on — and structurally excludes — US technology. The plan is early-stage and aspirational, but its direction is unambiguous. The global AI supply chain is bifurcating, and this is the most concrete evidence yet of what the other side looks like.

Sources: