- China plans a massive AI computing grid backed by domestic chips
- High-bandwidth memory shortages restrict advanced AI accelerator production in China
- Domestic chipmakers still lag behind global leaders by several years
China is drafting a plan which could direct roughly 2 trillion yuan (about $295 billion) toward a nationwide AI computing network.
The proposal would connect data centers across the country into a unified computing grid operated largely by state-backed telecommunications companies.
Officials reportedly want at least 80% of the underlying technology, including AI chips and related infrastructure, to come from Chinese suppliers.
A massive buildout centered on local technology
The blueprint is being developed by the National Development and Reform Commission, while major carriers including China Mobile and China Telecom would oversee the operation.
According to reports, the network would be linked into a single national computing platform by 2028 through extensive infrastructure deployment.
Financing would rely heavily on sovereign borrowing and ultra-long government bonds, while associated power grid upgrades could increase costs substantially.
The total capital requirement could rise beyond 5 trillion yuan (about $738 billion) when energy infrastructure is included in broader estimates tied to the rollout.
The plan arrives as Beijing continues tightening restrictions on foreign semiconductor products used in data centers and AI facilities.
In 2025, authorities introduced requirements that data centers obtain at least 50% of their chips from domestic manufacturers – and by November that year, state-funded projects reportedly faced additional restrictions that excluded foreign accelerators from facilities still under construction.
Officials also pushed compliance measures that required the removal of Nvidia, AMD, and Intel components in projects below 30% completion.
Those measures have increased opportunities for Chinese chip companies, including Huawei, while reducing dependence on suppliers such as Nvidia, AMD, and Intel.
The policy ensures that critical AI tools and LLMs operate on hardware developed within China, but replacing imported processors remains a difficult task.
Domestic chip supply remains a major challenge
Chinese semiconductor manufacturing capabilities mostly come from SMIC and a small group of state-approved foundries.
SMIC co-chief executive Zhao Haijun has warned that excessive infrastructure expansion could leave facilities underused.
Reports indicate that SMIC’s most advanced stable manufacturing process remains roughly comparable to 7nm technology and is already operating above 93% utilization levels.
With numerous domestic chip designers competing for the same production resources, expanding output quickly may prove difficult under current wafer allocation limits.
High-bandwidth memory also remains a major constraint, limiting how many advanced accelerators can be assembled for AI workloads and AI tools deployment.
Industry estimates suggest domestic suppliers may only cover around 76% of Chinese AI chip demand by 2030, even as demand expands toward a $67 billion market size.
Huawei has increased shipments, including an estimated 812,000 chips last year, but supply chain limits continue affecting production scaling.
Chinese industry executives have acknowledged that domestic AI data center chips remain 5 to 10 years behind leading international competitors in some categories.
Reports also indicate that DeepSeek returned to Nvidia hardware for certain training tasks after experimenting with Huawei alternatives in heavy workloads.
This suggests Chinese processors may still struggle with the most demanding AI training environments despite progress in inference performance.
Via Tom’s Hardware
Follow TechRadar on Google News and add us as a preferred source to get our expert news, reviews, and opinion in your feeds.
https://cdn.mos.cms.futurecdn.net/hQtTkoYUP8T6i7tWxMcj58-1920-80.png
Source link
