China has drafted a 2 trillion yuan plan—$295 billion at current exchange rates—to build a unified national AI computing grid by 2028. The mandate requires that at least 80% of the AI accelerator chips powering the infrastructure come from domestic suppliers. That requirement, if enforced, effectively terminates Nvidia and AMD's access to China's AI build-out market and positions Huawei, Alibaba Cloud, Biren Technology, and Moore Threads as the primary hardware beneficiaries of the largest state-directed AI infrastructure investment in history.

For context: the United States' Stargate initiative, announced in January 2026, committed $500 billion over four years in private investment. China's plan is $295 billion over five years in government-directed spending. The scale is comparable. The architecture is different. And the constraint it places on domestic chip producers may be the biggest challenge the plan faces.

The plan, drafted by China's National Development and Reform Commission, calls for connecting thousands of data centers into a single unified computing grid. The facilities will be operated primarily by state carriers China Mobile and China Telecom, which provide the network fabric to stitch regional data centers into a nationally accessible compute resource. The target is 2028 for full interconnection. Individual data centers are already under construction across Beijing, Shanghai, Chengdu, and the designated western computing zones that China has been developing since 2021 under the "East Data West Computing" initiative.

The grid is designed to provide AI compute as a national resource, accessible to domestic companies, research institutions, and government agencies through a unified API layer—essentially a national-scale cloud owned by the state, built on domestic silicon, and interoperable across carriers. Funding flows primarily through sovereign debt and ultra-long special government bonds, a financing mechanism China has used for infrastructure at scale in previous programs. When power-grid integration is included—the electrical infrastructure required to run facilities at this density—the total projected investment reaches at least 5 trillion yuan, approximately $740 billion.

The most consequential element is the chip sourcing requirement. At least 80% of AI accelerator hardware in the national grid must come from domestic suppliers. The remaining 20% can come from international sources, but the mandate effectively structures the entire build-out around Chinese silicon. This is not new policy direction; it represents the acceleration of a framework that has been building since 2022. In May 2026, Beijing formally approved nine categories of domestically developed AI chips for deployment across government and security-sensitive sectors—effectively clearing Huawei's Ascend series, Alibaba's Hanguang 800, Biren Technology's BR100 and BR104, and Moore Threads' MTT S80 for procurement under government frameworks.

The mandate accomplishes three things simultaneously. First, it eliminates Nvidia and AMD from the primary tender process for the national grid. Both companies are US-headquartered, and their flagship AI accelerators—Nvidia H100, H200, B100; AMD MI300X—remain subject to US export controls that prohibit their sale into China for AI applications above specific capability thresholds. Even without this mandate, those export controls already restrict the highest-performing Nvidia and AMD chips from the Chinese market. Second, it creates guaranteed procurement volume for domestic chip producers at a scale they have never experienced. Biren Technology, founded in 2019, and Moore Threads, founded in 2020, are pre-revenue at meaningful enterprise volumes. A national grid mandate changes their market structure overnight. Third, it accelerates the timeline for Chinese chips to close the performance gap through volume production experience. Chip performance is partly a function of manufacturing volume—more production means better yields, better quality control, and faster iteration. State procurement at this scale functions as both an infrastructure program and as industrial policy for semiconductor maturation.

Huawei is the most advanced domestic AI chip producer. Its Ascend 910B chip, produced on SMIC's 7nm process, is the current benchmark for domestically produced AI accelerators. Benchmarks published by domestic researchers in 2025 showed the Ascend 910B at roughly 60–70% of H100 performance on standard transformer training workloads—a meaningful gap, but narrower than it was two years earlier. The production constraint is HBM (High Bandwidth Memory), the stacked memory that sits alongside the AI accelerator die and is critical to performance on attention-heavy workloads. SK Hynix, Micron, and Samsung produce the world's HBM supply—all US-allied companies subject to export control frameworks. China's domestic HBM capacity is minimal. Huawei can design accelerators; it cannot easily fill the memory requirement without imported HBM, and that dependency constrains how many Ascend chips it can assemble.

Biren Technology produces the BR100, a 7nm chip targeting HPC and AI workloads. Early benchmarks suggest competitive performance with older Nvidia A100-class hardware. The company received government security clearance in May 2026 and is now eligible for national grid procurement. Moore Threads focuses on graphics and AI inference rather than training. Its MTT S80 chips target primarily inference deployment, which matters because China's build-out requires both training infrastructure (building models) and inference infrastructure (serving them). Alibaba Cloud produces the Hanguang 800, optimized for inference on Alibaba's specific workload profile. It is less a general-purpose AI accelerator and more a tailored inference chip; its relevance to the national grid depends on whether inference workloads are contracted to Alibaba's cloud infrastructure.

Chinese chip executives have publicly conceded that the country trails the leading edge in AI data center silicon by five to ten years. That gap is not primarily in chip design—it is in the manufacturing process and in the memory stack. TSMC at 3nm and Samsung and SK Hynix at advanced HBM3/HBM3e represent production capabilities China cannot currently replicate domestically. SMIC, the leading Chinese foundry, produces at 7nm and is expanding to 5nm-equivalent through process refinement. That is meaningful progress, but not catch-up.

The $295 billion mandate accelerates procurement of what China can make now. It does not eliminate the underlying technology gap. A national grid built on 7nm-class silicon and constrained HBM supply will provide substantially less compute per dollar than equivalent US infrastructure running Nvidia Blackwell chips with full HBM3e memory stacks.

China has demonstrated a workaround at the model level: inference efficiency. DeepSeek's R2 series achieved competitive performance on reasoning benchmarks against GPT-5 while using substantially less compute per inference call. If Chinese model development continues to optimize for hardware-constrained inference, the gap in compute capacity matters less for the applications the national grid serves.

The comparison in dollar terms is straightforward; the structural comparison is complex. US AI investment is private capital chasing returns. China's investment is state capital pursuing strategic capacity. Both are building compute at scale. Only one is doing it with hardware constraints built into the mandate from day one.

For developers and companies building AI applications, the China national grid plan formalizes something that has been informally true since 2022: there are now two separate global AI infrastructure stacks, and they are diverging. Applications built for US and allied-market deployment run on Nvidia hardware, use US-headquartered API providers, and operate under export control frameworks that assume chip access. Applications built for deployment in China run on domestic silicon, connect to state-operated compute resources, and operate under data localization requirements that assume state access. These two stacks are not interoperable in any meaningful way. A model trained on Nvidia hardware and served through OpenAI's API cannot simply be ported to the national grid. A model trained on Ascend hardware and optimized for domestic inference chips will require re-work to run at equivalent performance on Nvidia infrastructure.