> ## Content Index > Fetch the complete content index at: https://bytevyte.com/llms.txt > Use this file to discover other available public pages before exploring further. # China's AI Infrastructure Plan Commits $532 Billion to 9,800 EFLOPS by 2030 - URL: https://bytevyte.com/chinas-ai-infrastructure-plan-commits-532-billion-to-9-800-eflops-by-2030/ - Published: 2026-09-11T06:40:22.000Z - Updated: 2026-09-11T06:40:22.000Z - Description: China's AI infrastructure plan commits $532 billion to reach 9,800 EFLOPS by 2030, with 100,000-card clusters built on domestic chips. - Author: Bytevyte Editorial - Tags: ai-beats China's AI infrastructure plan commits CNY3.8 trillion, roughly $532 billion, to information infrastructure between 2026 and 2030, and sets a national target of 9,800 EFLOPS of intelligent computing capacity by the end of the decade. The **Ministry of Industry and Information Technology** published the 15th Five-Year Plan for the information and communications sector on September 7, placing AI compute at the center of Chinese industrial policy for the next five years. The capacity goal is about six times the 1,590 EFLOPS baseline MIIT records for 2025. The plan calls for the orderly deployment of AI computing clusters built around 100,000 or more accelerator cards each, and asks for greater efforts to adapt infrastructure to home-grown chips. It names six priority tasks, among them building information infrastructure with moderate overcapacity, strengthening network and data security, and widening integrated applications across traditional industries. The build-out runs through China's East Data, West Computing strategy, which routes compute to resource-rich western provinces while keeping eastern demand centers connected to the capacity. ## What China's AI Infrastructure Plan Commits The document covers connectivity as well as compute. By 2030 MIIT expects 50 5G base stations per 10,000 people and a 5G user penetration rate of 95%, targets that set the transport layer for distributed inference and for the industrial deployments the plan wants to push into manufacturing. The base-station goal counts 5G-Advanced sites alongside standard 5G, tying the connectivity target to the factory and logistics applications the ministry wants to expand. | Metric | 2025 baseline | 2030 target | | -------------------------------------------------------------- | -------------- | ------------------------------------ | | Intelligent computing capacity | 1,590 EFLOPS | 9,800 EFLOPS | | AI cluster scale | Not specified | 100,000+ accelerator cards | | 5G base stations per 10,000 people | Not specified | 50 | | 5G user penetration | Not specified | 95% | | Cumulative information infrastructure investment, 2026 to 2030 | Not applicable | CNY3.8 trillion (about $532 billion) | Spread across five years, the CNY3.8 trillion commitment averages about CNY760 billion a year, and MIIT expects state and market participants to fund the build jointly rather than through a single budget line. Clusters at the 100,000-card scale that China's AI infrastructure plan specifies change the physical shape of the project. A site of that size needs dedicated substations, liquid cooling at scale, and high-bandwidth interconnect across tens of thousands of nodes, which is why the East Data, West Computing routing matters: the plan separates compute from demand so the power-hungry part of the stack can sit where electricity is cheapest. The trajectory behind the headline is steep. RAND's assessment of Chinese AI industrial policy puts total national compute capacity at 246 EFLOP/s as of June 2024 under the earlier action plan, with a 300 EFLOP/s goal for 2025\. The 1,590 EFLOPS baseline MIIT now cites is several times larger, and the plan does not break out how much of that installed capacity suits modern training and inference workloads. RAND has also noted that a substantial share of Chinese compute is not well matched to AI work, so the EFLOPS figure is best read as installed capacity rather than usable AI throughput. Exaflop targets carry a unit caveat. Numbers of this kind are usually quoted at reduced precision, so 9,800 EFLOPS of intelligent compute is not directly comparable to training throughput measured in BF16 or FP8 on a single model. Treat it as a planning target across the country's public and commercial data centers. Depending on the baseline used, the expansion is described as anywhere from roughly four and a half times to six times current capacity, which is why the plan's own 2025 reference point matters more than the round numbers attached to the announcement. ## The Silicon Problem the Plan Names The most consequential language in the plan concerns chips. MIIT asks for greater efforts to adapt infrastructure to home-grown computing chips, a position shaped by US export controls that restrict Chinese access to the most advanced Nvidia accelerators. Huawei's Ascend line is expected to take a rising share of the workload. Huawei shipped around 812,000 AI chips last year and projects about $12 billion in processor revenue for 2026, a pace its own supply chain has struggled to sustain. Supply projections point to a persistent gap. Domestic suppliers are projected to cover around 76% of Chinese AI chip demand by 2030, against a market growing toward $67 billion. Even on that assumption, close to a quarter of demand would have to be met from elsewhere or left unmet, and the shortfall falls hardest on the largest training clusters, which need the highest-performance parts and the most advanced high-bandwidth memory. That makes fabrication capacity, yield, and memory supply the binding constraints on the 9,800 EFLOPS target, not the capital commitment. Money is the part of the equation Beijing has already solved. Wafer starts and advanced packaging are not. Cluster specifications also assume a supply chain that can deliver at volume. A 100,000-card installation built on domestic accelerators needs matching domestic switches, optics, and high-bandwidth memory, so a bottleneck in any single link holds up an entire site rather than one rack. The instruction to adapt infrastructure to home-grown chips acknowledges that the integration work is not finished. ## The Trade-offs of Building Ahead of Demand Listing moderate overcapacity as a priority task is a deliberate choice to build ahead of demand. That reduces the risk of shortages when a single training run needs tens of thousands of accelerators at once. It creates the opposite risk in a market where model development consolidates around fewer, larger labs: data centers that sit underused while their power contracts and depreciation schedules keep running. Power is the second constraint. Routing clusters to western provinces solves land and electricity price problems, but 9,800 EFLOPS of installed capacity needs generation, transmission, and cooling at a scale that provincial grids have to absorb. No energy budget is attached to the compute target in the document, which leaves that question to provincial planning. The plan pairs the capacity target with network and data security tasks, which indicates governance requirements will be attached to the same infrastructure as it comes online. The funding model spreads both risks. State telecom carriers, domestic chip vendors such as Huawei, and the server, cooling, and interconnect suppliers that build the halls are the direct beneficiaries of a program averaging CNY760 billion a year. Nvidia's addressable market inside China narrows further with each cluster specified for domestic silicon. For Chinese AI developers, abundant domestic compute lowers cost per token. The trade-off is software lock-in, because code tuned for CUDA does not port to Ascend and other domestic accelerators without real engineering work. Policy follow-through matters as much as the headline investment. MIIT Minister Li Lecheng has signalled that the ministry will issue an AI Plus Manufacturing plan under its wider AI Plus program, which would push the same compute base into factory floors, inspection systems, and supply-chain planning. That application layer is where the spending has to earn a return. Capacity built for its own sake carries no revenue. The near-term signals are concrete. Whether MIIT publishes annual milestones against the 1,590 EFLOPS baseline, how much of the CNY3.8 trillion lands on state carrier balance sheets rather than private budgets, and whether Ascend-class accelerators ship at the volume Huawei's revenue guidance implies will decide whether 9,800 EFLOPS is a build schedule or an aspiration. ## Why this matters China's AI infrastructure plan moves the competitive question from which lab has the best model to who can build, power, and populate the most compute. If even part of the 9,800 EFLOPS target is delivered, Chinese labs gain cheaper domestic capacity and less exposure to export controls, while the pressure shifts to advanced memory and packaging supply. For buyers of AI capacity outside China, the signal is that hardware is being added faster than the porting and tooling work needed to use it well. *AI-generated image.* ## Related Articles - [ByteDance AI Infrastructure Budget Hits $30 Billion for 2026](https://bytevyte.com/bytedance-ai-infrastructure-budget-hits-30-billion-for-2026/) - [Nvidia Dismisses China AI Chip Plans Even as H200 Deliveries Land at ByteDance and Tencent](https://bytevyte.com/nvidia-dismisses-china-ai-chip-plans-even-as-h200-deliveries-land-at-bytedance-and-tencent/) - [EU AI Gigafactories: Europe's €30B Sovereign Compute Bet](https://bytevyte.com/eu-ai-gigafactories-europes-eu30b-sovereign-compute-bet/) ✔Human Verified --- *Researched and cross-referenced against primary sources by the Bytevyte editorial team. This article was generated with the assistance of artificial intelligence and reviewed by the Bytevyte editorial team.*