Beijing is drawing up plans for something no other country has tried at this scale: a single, nationwide computing grid for artificial intelligence, knitting data centres across the country into one platform run largely by state-backed telecom carriers. The price, according to reports on the draft, is roughly 2 trillion yuan, about 295 billion dollars, and that figure could swell past 738 billion once the power-grid upgrades needed to feed it are counted. The National Development and Reform Commission is shaping the blueprint, with China Mobile and China Telecom slated to operate it, and a target of linking everything into one national system by 2028.
The defining condition is self-reliance. Officials reportedly want at least 80 percent of the underlying technology, the AI chips above all, to come from Chinese suppliers. That is the logical endpoint of a policy China has been tightening for more than a year. In 2025 authorities began requiring data centres to source at least half their chips domestically, then moved to exclude foreign accelerators from state-funded projects still under construction, in some cases ordering Nvidia, AMD and Intel parts pulled from facilities less than a third of the way built. The grid would lock that principle into national infrastructure: critical AI systems running on hardware made at home.
The ambition collides with an awkward fact. China cannot yet make enough of the chips this plan assumes. Domestic production leans on SMIC and a handful of approved foundries, and SMIC's most reliable advanced process is still roughly equivalent to 7-nanometre technology, a step or two behind the leading edge, with its lines already running above 93 percent utilisation. There is little slack to absorb a national buildout. High-bandwidth memory, the specialised chips that sit beside an accelerator and largely determine how fast it can train a model, is an even tighter bottleneck. One industry estimate suggests domestic suppliers might cover only about 76 percent of Chinese AI-chip demand by 2030.
The candour from inside the industry is striking. SMIC's co-chief executive has warned that overbuilding could leave expensive facilities sitting idle, the classic risk of a state-driven infrastructure rush. Chinese executives have acknowledged that domestic AI data-centre chips trail the best international parts by five to ten years in some categories. Most telling, DeepSeek, one of the country's marquee AI labs, reportedly went back to Nvidia hardware for certain training runs after trying Huawei's alternatives on the heaviest workloads, a quiet admission that home-grown silicon can still struggle with the most demanding end of the job.
So the plan reads as two stories at once. One is a genuine show of strategic will: a country prepared to spend hundreds of billions to wall off a core technology from foreign leverage, and to absorb the inefficiency that self-sufficiency costs. The other is a bet placed before the hardware exists to win it. Inference, the running of finished models, is closer to within reach on domestic chips than training, the far hungrier task of building them. The open question is whether Chinese fabs can close that gap on anything like the 2028 timeline, or whether the grid gets built faster than the silicon meant to fill it. Either way, the direction is unmistakable. The contest over AI is settling into the ground, into fabs, foundries and the electricity that runs them.