On June 9, 2026, Bloomberg reported that China's National Development and Reform Commission (NDRC) and other central agencies are drafting a blueprint to spend roughly 2 trillion yuan ($295 billion) over five years on a nationwide grid of state-operated AI data centres. The hubs would be run mostly by China Mobile and China Telecom, knit into a single interconnected network by 2028, and — most consequentially — would source at least 80% of their technology, including AI chips, from domestic suppliers like Huawei, effectively squeezing out Nvidia and AMD. It is billed as a prong of Beijing's new "Six Networks" infrastructure programme, and with power-grid integration the total could reach 5 trillion yuan.
The headline is the money. The more important story is the doctrine. This plan is not a break from China's regulatory tradition; it is its logical endpoint. The same localisation logic that the 2017 Cybersecurity Law applied to data is now being applied to the machines that process it.
From data localisation to compute localisation
Article 37 of the Cybersecurity Law, effective June 1, 2017, requires that critical information infrastructure operators "store within mainland China" the personal information and important data they gather domestically, permitting cross-border transfer only after a state security assessment. The premise was that where data physically sits is a national-security variable, not a commercial one.
The 2022 "Eastern Data, Western Computing" project — approved by the NDRC and three other central departments on February 18, 2022, with eight national computing hubs and ten data-centre clusters — extended that premise from storage to location of computation, routing eastern data to power-rich western provinces. The 2026 plan takes the final step: it nationalises the supply chain of compute itself, mandating that the hardware be overwhelmingly domestic and state-operated. Localisation has travelled from the bytes, to the buildings, to the chips.
The steelman
The case for this is genuinely strong, and worth stating plainly before contesting it. China has watched successive U.S. export-control rounds restrict Nvidia's H20 and H200 shipments, leaving its frontier-AI ambitions hostage to another government's licensing decisions. Even Nvidia concedes the ground is shifting: CEO Jensen Huang said in May 2026 that the company has "largely conceded" China's AI-chip market to Huawei. For a state that treats compute as critical infrastructure, building a supply chain no foreign regulator can throttle is a rational resilience play, not mere protectionism. If your adversary can switch off your factories, vertical integration is self-defence.
Where proportionality breaks
The problem is that an 80% domestic-content floor is a blunt instrument that optimises for control at a measurable cost to capability. Huawei's Ascend accelerators are competitive at inference but trail on training: reporting indicates the 910C delivers roughly 80% of the H20's memory bandwidth on an HBM2E standard two generations behind the frontier. A buildout legally barred from buying the best available silicon is, by construction, a buildout that runs slower per watt and per dollar than its global competitors. Beijing is partly betting it can compensate through scale — lashing weaker chips into vast clusters — but that converts an efficiency gap into a much larger energy bill, which is precisely why the power grid is folded into the same plan.
There is also a state-capacity question. Routing $295 billion through China Mobile, China Telecom, and ultra-long-term sovereign bonds concentrates allocation decisions in administrative hands rather than competitive ones. China's own private AI sector — over 6,200 firms and a core industry valued near $174 billion in 2025 — grew on a far more pluralistic model, from DeepSeek's efficiency breakthroughs to a dense start-up ecosystem. State-operated capacity risks crowding that out, subsidising incumbents and standardising on hardware chosen for its passport rather than its performance.
The proportionate path
None of this argues against China building sovereign compute. Resilience is a legitimate policy goal, and a country facing export controls is entitled to reduce single-points-of-failure. The objection is to the blunt 80% mandate. A proportionate design would ring-fence genuinely security-sensitive workloads — military, critical-infrastructure, state-secret processing — for verified domestic stacks, while leaving commercial AI free to buy the best chips available. That preserves the security interest the Cybersecurity Law was written to protect without taxing the entire economy's compute with a generational performance penalty.
The broader lesson reaches well past Beijing. Washington's export controls and China's localisation mandate are now a feedback loop: each restriction justifies the next wall. The predictable result is two compute blocs that are individually more "sovereign" and collectively less efficient, with Nvidia losing a market it spent two decades building and Chinese developers paying a hidden tax in foregone performance. Computing sovereignty, pursued by mandate rather than by competitiveness, is a costly insurance policy — and the more rigid the content floor, the higher the premium everyone pays.