China is drafting a plan to spend roughly 2 trillion yuan, about $295 billion, over five years on a nationwide web of AI data centres, according to a Bloomberg report picked up this week by Tom's Hardware and others. The National Development and Reform Commission would design the blueprint, state carriers China Mobile and China Telecom would run most of the facilities, and everything would be stitched into a single national computing grid by 2028. The defining requirement sits in one number: at least 80 percent of the underlying technology, chips included, must come from domestic suppliers such as Huawei. That threshold effectively writes Nvidia and AMD out of the world's second-largest economy's flagship infrastructure project.
The reports agree the money is the easy part. The plan leans on sovereign debt and ultra-long special bonds, and folding in the necessary power grid upgrades could push total capital needs past 5 trillion yuan. The harder constraint is physics and fabrication. SMIC, China's leading foundry, tops out at its N+2 process, roughly equivalent to 7 nanometres, and is already running above 93 percent utilisation while every government-certified Chinese chipmaker competes for the same wafer slots. High-bandwidth memory is scarcer still, capping how many Ascend-class accelerators Huawei can actually assemble. One industry estimate cited in the coverage suggests domestic suppliers will meet only about 76 percent of Chinese AI chip demand by 2030. Even Chinese executives are publicly hedging: SMIC's co-CEO has compared the buildout to constructing highways ahead of the traffic, and DeepSeek, after being steered toward Huawei hardware for training, reportedly drifted back to Nvidia.
What makes the story bigger than an infrastructure budget is the squeeze happening from the other side at the same moment. Digitimes reports that Taiwan is weighing significantly tougher restrictions on advanced AI chip exports to China, aligning its rules with Washington's and targeting the smuggling networks that keep restricted silicon flowing. And in Foreign Affairs, Ely Ratner and co-authors argue that export controls are working exactly as competitive strategy should: not stopping China's chip drive, but forcing Beijing to spend more money, time and effort for worse results. Their piece urges the US to go further, closing loopholes around shell companies, cloud access and equipment servicing, and to stop approving sales like Nvidia's H200 licences that hand back the advantage.
Read together, the two sides describe a feedback loop. Western controls make Chinese self-sufficiency a necessity; Beijing's 80 percent mandate makes the decoupling formal and permanent, locking out foreign vendors even where domestic parts lag. The Foreign Affairs authors raise what they see as the deeper stake: if AI systems eventually become capable of improving themselves, the size of a compute lead could compound rather than erode, which is why both capitals are treating wafer capacity as a strategic resource rather than a market.
The caveat worth keeping is that Bloomberg's sources describe early-stage discussions, not settled policy. The supplier rules, funding and timeline could all shift. But the direction has been consistent for a year now, from a 50 percent local content rule last August to a full bar on foreign accelerators in state-funded projects by November. The question is no longer whether China builds a parallel AI stack. It is whether the parallel stack can keep up, and what a world with two of them looks like.