A draft five year plan from China's top economic planner would federate thousands of data centers into a state run AI compute network, but the real fight is over the meaning of "domestic" and whether the grid can carry the load.
China's draft five-year plan to build a 2-trillion-yuan (about US$295 billion) national AI compute grid is being read in Western chip markets as a verdict on Nvidia and AMD. That reading is half right, and half wrong. The procurement headline, which says at least 80% of the underlying technology in the network, including AI accelerator chips, must come from domestic suppliers, is the visible part of the policy. The operative part is a set of words that have not been written yet: what counts as "domestic," what the 80% is measured against, and whether advanced memory and packaging are inside the denominator at all.
The plan, attributed to China's National Development and Reform Commission (NDRC, the cabinet-level economic planning agency), was first reported by Bloomberg on June 9, 2026 and summarized the same day in Tech Times. It envisions stitching together thousands of Chinese data centers into a single state-controlled AI compute network, with the 80% mandate applying to that network's underlying technology stack. On the day of the Bloomberg report, Nvidia shares fell roughly 2.4% and AMD dropped about 4% in U.S. trading, a market reaction to a single procurement wave, not a declaration that either company is exiting China.
Whether the mandate functions as a foundry-localization subsidy or a true lockout depends entirely on the implementing rules. If "domestic" is defined as "assembled or packaged in China," Nvidia could route shipments through SMIC or a downstream partner and remain inside the 80%. If "domestic" is defined as "Chinese-owned IP from the transistor up," Nvidia is functionally excluded regardless of where the wafer is polished. The same definitional fight governs the 80/20 denominator. If the percentage is measured in chip count, a single Nvidia flagship GPU and a hundred domestic inference accelerators land at very different weights than if the same rule is denominated in dollar value or in delivered FLOPS (floating-point operations per second, the standard measure of AI compute throughput). None of those denominator choices are in the public draft.
The mandate also stops at chips, but AI training workloads do not. A modern AI accelerator is incomplete without high-bandwidth memory (HBM, the stacked DRAM chips that sit next to the processor and feed it data at terabytes per second) and advanced packaging (the techniques that bind multiple chip dies into one package, including TSMC's CoWoS and similar domestic efforts). Whether HBM and packaging are inside the 80% threshold is the single largest open question in the policy, because China's HBM capacity, led by CXMT, is years behind SK Hynix and Samsung, and its advanced packaging roadmap is not yet at parity with TSMC's CoWoS lines. A mandate that defines "underlying technology" to include HBM and packaging effectively raises the bar from a chip-level localization to a stack-level one, and changes the timeline.
The second constraint is electrical. A June 22, 2026 analysis from The Next Web, referenced by Tech Times, projects Chinese data-center power demand rising by 300 to 500 billion kilowatt-hours between 2026 and 2030, roughly a fifth of China's total electricity-demand growth over that period. Chinese data-center operators are already struggling to reconcile AI training load spikes with the country's clean-energy mandates and grid reliability targets. The five-year horizon of the NDRC plan assumes that 300 to 500 billion kWh of new supply, plus the transmission upgrades to deliver it, can be built out on a 2026-to-2030 schedule. Whether that assumption holds is the load-bearing variable the chip-procurement headline does not mention.
The scale is genuinely historic. A 2-trillion-yuan central envelope is the largest single-nation AI infrastructure commitment publicly on record, and analysts cited in the Tech Times summary put the supporting electrical infrastructure at roughly 3 trillion yuan more, which would put the all-in total near 5 trillion yuan (about US$740 billion). That total is a planned figure, not an executed spend, and it depends on financing from China's policy banks, with China Development Bank (CDB) and the Industrial and Commercial Bank of China (ICBC) among them, whose credit covenants will determine which operators can actually draw on the pool. The mandate and the money are two different policy instruments, and the gap between them is where the next 36 months of Chinese AI capacity will actually be decided.
For readers tracking this from outside China, the checklist is short and concrete. Watch for the NDRC implementing glossary: does "domestic" mean assembly, equity, or IP. Watch for the 80/20 denominator: chip count, dollar value, or delivered FLOPS. Watch for the HBM and packaging scope: are they inside the mandate or carved out. Watch for the credit covenants on CDB and ICBC financing: do they require Chinese-sourced silicon at the wafer level. And watch the grid: whether 300 to 500 billion kWh of new supply actually arrives by 2030, and whether the operators running training workloads can keep their loads inside the clean-energy envelope. None of those five answers is in the chip-procurement headline. All of them are the policy.