Beijing Sets Sights on 9,800 Eflops: Inside China's Five-Year AI Infrastructure Sprint
A new government blueprint targets quadrupling compute capacity by 2030, with clusters of 100,000 accelerator cards and $532 billion in investment riding on infrastructure that still faces chip constraints.
The Scale of the Ambition
China's Ministry of Industry and Information Technology unveiled a five-year blueprint on Monday that puts hard numbers to Beijing's AI infrastructure ambitions. The target: 9,800 eflops of intelligent computing capacity by 2030, roughly four times the current installed base. To get there, the plan calls for deploying clusters that house 100,000 accelerator cards apiece, a scale that would rank among the largest training environments globally, and mobilising 3.8 trillion yuan (532 billion US dollars) in investment across the period.
At DailyTechWire, we've tracked the region's compute build-out for the past two years, and this is the first time Beijing has published a capacity target in exaflops alongside cluster specifications. The figure itself is revealing less for its absolute size than for what it signals about sequencing. A fourfold expansion in five years implies annual growth rates north of 30 per cent, sustained across a period when export controls on advanced chips remain in force and domestic alternatives are still climbing the performance curve.
Cluster Architecture and Card Economics
The 100,000-card cluster target matters because it speaks to training workloads, not just inference at the edge. Clusters of that size require low-latency interconnect, synchronised memory hierarchies, and power infrastructure that can deliver tens of megawatts to a single site. According to the ministry's plan, these facilities will serve as anchors for regional compute hubs, with priority given to provinces that already host semiconductor fabs or hyperscale data centres.
Card economics are the other half of the equation. At current prices for domestically produced AI accelerators, outfitting a single 100,000-card cluster runs into the hundreds of millions of dollars for silicon alone, before accounting for networking, cooling, or real estate. The 3.8 trillion yuan investment envelope suggests Beijing expects a mix of state-owned enterprises, provincial governments, and private capital to shoulder the build-out, with the ministry providing coordination and, likely, preferential loans through policy banks.
The Eflops Metric and What It Hides
Eflops, or exaflops, measure floating-point operations per second at the ten-to-the-eighteen scale. It is a clean headline number, but it aggregates across chip generations, precisions (FP16, INT8, and lower), and utilisation rates that vary widely in practice. A system rated at 1,000 eflops running inference at INT8 precision is not directly comparable to another running FP32 training workloads, yet both contribute to the national tally.
The ministry's plan does not break out how much of the 9,800 eflops will be dedicated to training versus inference, nor does it specify which chip architectures, domestic or imported, will make up the majority. That ambiguity is strategic. It allows Beijing to claim progress on the headline figure while navigating the reality that the highest-performance training chips remain subject to US and allied export restrictions, and domestic substitutes, while improving, have not yet closed the gap in memory bandwidth or energy efficiency.
Investment Flows and Regional Distribution
The 3.8 trillion yuan commitment dwarfs previous rounds of AI infrastructure spending and positions compute as a priority on par with 5G network roll-outs or electric-vehicle charging grids. The plan directs funds toward three categories: building new clusters, upgrading existing data centres with AI accelerators, and developing the software stack (compilers, scheduling tools, model libraries) needed to keep utilisation rates high.
Regional competition will shape where the clusters land. Provinces with existing semiconductor ecosystems (Jiangsu, Shanghai, Guangdong) have natural advantages, but the ministry has signalled that inland regions with surplus renewable energy (Sichuan, Inner Mongolia) may also win allocations if they can guarantee power supply and connectivity to backbone networks. For provincial officials, hosting a 100,000-card cluster brings prestige, tax revenue, and a ticket to participate in the national AI conversation.
The Chip Constraint That Dare Not Speak Its Name
None of this happens in a vacuum. The US Bureau of Industry and Security tightened export controls on AI chips in late 2023, and subsequent updates in 2024 and early 2025 closed loopholes around performance-per-watt thresholds and third-country re-export. Chinese firms responded by stockpiling older-generation GPUs, designing custom ASICs, and pushing domestic foundries to produce AI accelerators using 14-nanometre and 7-nanometre processes.
The result is a bifurcated landscape. Tier-one labs and well-funded startups have access to pre-ban inventory or smuggled chips, keeping them competitive on frontier model training. Everyone else relies on domestic silicon that trades performance for availability. The ministry's plan tacitly acknowledges this by emphasising "independent and controllable" supply chains, a euphemism for chips made in China, even if they lag by a generation or two.
What the plan does not say is whether 9,800 eflops is enough. Frontier labs in the US are already training on clusters that exceed 50,000 GPUs, with roadmaps pointing toward 200,000-GPU systems by 2027. If China's 100,000-card clusters are built with chips that deliver half the throughput per card, the effective training capacity may not keep pace with the headline number suggests.
Software, Utilisation, and the Last Mile
Hardware is necessary but not sufficient. The ministry's plan allocates a portion of the 3.8 trillion yuan to software infrastructure: frameworks that abstract across heterogeneous accelerators, schedulers that maximise cluster utilisation, and libraries of pre-trained models that lower the barrier for enterprises without in-house AI teams.
Utilisation rates are where many national compute initiatives stumble. A cluster that sits idle 40 per cent of the time because of scheduling bottlenecks or incompatible software delivers far less value than its peak spec suggests. The plan calls for "intelligent scheduling platforms" that can allocate resources dynamically, but building those platforms requires expertise in distributed systems, not just procurement budgets. Whether the ministry can attract that talent, or whether it ends up concentrated in a handful of hyperscalers, will determine how much of the 9,800 eflops translates into actual model training and inference.
What This Means for the Region's Compute Race
China's five-year target lands in the middle of a broader Asian compute arms race. South Korea announced a 2.1 trillion won AI infrastructure fund in early 2026; Singapore is expanding its national AI compute cluster; Japan has earmarked budget for sovereign AI development. Each country faces different constraints (Korea has Samsung and SK Hynix but limited hyperscale cloud; Singapore has capital but relies on imported chips; Japan has research strength but fragmented commercial deployment), and none can match China's absolute scale.
For the rest of Asia, China's build-out has two implications. First, it sets a floor for what "serious" AI infrastructure investment looks like, pressuring governments to commit multi-year, multi-billion-dollar budgets or risk falling behind in the capability race. Second, it reshapes the economics of cloud AI services. If Chinese hyperscalers can offer inference at lower cost because of subsidised domestic compute, that puts pressure on regional cloud providers and accelerates the bifurcation of the AI stack along geopolitical lines.
The 2030 target is ambitious, and the path to 9,800 eflops is littered with technical and political obstacles. But the plan's existence, and the investment commitment behind it, signals that Beijing views AI compute as infrastructure in the same category as roads, ports, and power grids: essential, strategic, and worth building even if the return on investment takes a decade to materialise. For anyone tracking the region's technology trajectory, this is the baseline to measure against.
