US-China AI Competition Daily Briefing.
1. The Domestic Milestone: 100,000-Chip Domestically-Produced Supercluster Debut.
In a major hardware milestone reported by state broadcaster CCTV, China’s first domestically-produced 100,000-chip AI supercluster made its official public debut at the 2026 National High-Performance Computing Conference:
- Sovereign Scaling Milestone: The supercluster integrates 100,000 domestic AI accelerators into a unified logical fabric, marking a major milestone in China's strategy to compensate for single-chip manufacturing limits via massive system-level integration.
- Optical Interconnect & Fabric Integration: The architecture relies on domestic optical switch fabrics and high-speed inter-chip buses to maintain model FLOPS utilization (MFU) across multi-thousand-node pools, providing sovereign compute capacity designed for multi-trillion parameter model training runs.
2. Export Control Policy: MOFCOM Reaffirms SaaS & API Exemption Details
Official follow-ups from China's Ministry of Commerce (MOFCOM) regarding its October 16 draft catalogue revision—which proposes controls on 40 technologies (including 23 AI domains)—provided key operational guidance for global tech firms:
- Clear Exemption for API/SaaS Distribution: MOFCOM explicitly clarified that overseas access to Chinese foundation models via application programming interfaces (APIs), software-as-a-service (SaaS) platforms, web browsers, or lightweight local client apps does not constitute a technology export. Global developers can continue using models like Alibaba’s Qwen series and DeepSeek without triggering export licensing.
- Focus on Transfer of Core IP: The restricted controls target the permanent outbound transfer or licensing of underlying model weights, training code, advanced EDA/chiplet IP, and cluster management know-how.
- Streamlined Cloud Operations: Chinese tech firms hosting proprietary models on their own overseas cloud data center infrastructure qualify for simplified or automatic export licensing, minimizing operational disruption for global expansion.
3. Supply & Demand Gap: China's 2026 Compute Deficit
Financial and industry reports evaluating hardware supply metrics across mainland China illustrate the persistent gap between domestic chip production and enterprise demand:
- 1.5 Million Unit Supply Deficit: Industry estimates project total Chinese demand for AI chips to exceed 5 million units in 2026, against a total domestic supply (including Huawei Ascend, Alibaba T-Head, Cambricon, and Baidu Kunlunxin) reaching approximately 3.5 million units.
- Gray Market Dynamics: To bridge the compute gap, Chinese tech firms continue utilizing secondary markets and workstation hardware alongside sovereign clusters, maintaining upward pressure on gray-market hardware prices across Asian trade hubs.
4. Software & Compiler Layer: TileLang Community Adoption
Developer metrics following the open-source release of TileLang—developed jointly by DeepSeek and Huawei—highlight expanding adoption across domestic research labs:
- Bypassing CUDA in Production: TileLang provides a native, high-efficiency compilation pipeline tailored directly for Huawei Ascend 950 silicon. By providing pre-compiled ports for custom operators (FlashMLA, DeepEP, and DeepGEMM), the language allows Chinese developers to write high-performance kernels natively without depending on Nvidia's CUDA.
- Testbed Megaclusters: The compilation layer is actively driving model training across gigawatt-scale data center facilities, such as DeepSeek’s 160,000-unit Huawei Ascend deployment in Ulanqab, Inner Mongolia.
Strategic Takeaway
The US-China AI dynamic has shifted into an era of reciprocal policy frameworks and system-level scaling. While Washington enforces chip transshipment controls and capital restrictions, Beijing is deploying its own export control catalogue to protect sovereign model IP—even as domestic tech leaders debut 100,000-chip sovereign superclusters to secure multi-trillion parameter compute capacity within national borders.
At the 2026 National High-Performance Computing Conference, official state media (including CCTV) highlighted the public debut of China's first fully domestically produced 100,000-chip AI supercluster (built on the Sugon/Dawning 8000 architecture).
Rather than relying on single-die transistor scaling, the platform demonstrates how Chinese hyperscalers and national computing hubs are using system-level super-fusion, optical fabrics, and liquid cooling to achieve cluster-scale parity with Western AI factories.
1. Architectural & Hardware Specs
The 100,000-card platform transitions domestic computing infrastructure from isolated 10,000-card pods into a single unified logical supercomputer capable of training multi-trillion parameter foundation models.
- Core Accelerators: Integrates 100,000 domestic AI accelerators (utilizing customized x86-compatible Haiguang DCU / Hygon and specialized domestic ASIC/NPU matrix engines).
- FP16 / BF16 Compute Performance: Delivers over 20 ExaFLOPS (EFLOPS) of FP16/BF16 mixed-precision compute capacity in a single physical deployment.
- Low-Precision Execution (FP8/FP4): Provides native hardware acceleration for FP8 and FP4 formats, boosting effective token generation throughput for Mixture-of-Experts (MoE) architectures by 2.8x to 3.2x.
- Memory Subsystem: Combines local high-bandwidth memory (HBM/HiBL) stacks into a shared, cluster-wide addressable memory pool exceeding 8 to 10 Petabytes (PB), allowing models with 5 to 10 trillion parameters to sit directly in high-speed system memory.
2. Interconnect Architecture & Optical Fabric
Scaling from 10,000 to 100,000 cards introduces massive network contention and packet-loss risks. To maintain a high Model FLOPS Utilization (MFU) across 100,000 cards, the system implements a unified optical switch fabric:
100,000-Card Interconnect Topology
ACCELERATOR NODES OPTICAL INTERCONNECT FABRIC SUPER POD LEVEL
┌───────────────────────┐ ┌────────────────────────────┐ ┌─────────────────────┐
│ 100,000 Domestic DCUs │ ──► │ scaleFabric / Lingqu Bus │ ──► │ Single Logical │
│ (128-card Supernodes) │ │ Non-Blocking RDMA Routing │ │ 100K-Card Engine │
└───────────────────────┘ └────────────────────────────┘ └─────────────────────┘
- Non-Blocking scaleFabric RDMA Interconnect: Employs an IB-native (InfiniBand-compatible) non-blocking RDMA network layer (scaleFabric) supporting 800 Gbps to 1.6 Tbps bidirectional bandwidth per node.
- Near-Packaged Optics (NPO) & Lingqu UnifiedBus: Integrates Near-Packaged Optics (NPO) modules directly alongside chiplet substrates. By converting electrical signals to optical pulses adjacent to the processor, the architecture reduces optical transceiver power consumption by 55% and cuts inter-chip latency across multi-thousand-card groups to sub-microsecond levels.
- 100,000-Node All-to-All Dispatch Bandwidth: Across Mixture-of-Experts (MoE) all-to-all token dispatch workloads, the interconnect delivers 312–320 GB/s per-channel aggregation bandwidth, preventing communication bottlenecks during large-scale model training.
3. Fault Tolerance, Liquid Cooling & System Efficiency
Operating 100,000 accelerators continuously introduces high hardware fault rates and massive heat output, requiring novel power and software-checkpointing mechanisms:
- Sub-Second Checkpointing & Automated Fault Recovery: Large-scale training runs on 100,000 cards normally experience node failures every few hours. The supercluster utilizes an automated in-memory checkpointing scheduler that detects NPU/NPO failures, isolates faulty nodes within milliseconds, and resumes training without rolling back global model states.
- Direct-to-Chip Warm-Water Liquid Cooling: Employs a full-cabinet direct-to-chip liquid cooling infrastructure, maintaining a Power Usage Effectiveness (PUE) below 1.10 even under peak 100% compute loads.
- Integration into National Supercomputing Internet: The 100,000-chip cluster functions as a primary core node connected to China's National Supercomputing Internet, making massive parallel compute accessible via standardized cloud APIs to over 1.4 million researchers, state enterprises, and domestic AI labs (such as DeepSeek and Baidu).
Strategic Impact
The public debut of the 100,000-chip supercluster confirms that Chinese infrastructure strategy has moved past single-chip EUV lithography limitations. By combining domestic x86/DCU silicon, optical switch fabrics (NPO), and open compiler stacks (such as TileLang), China has established a sovereign, 100,000-node hardware foundation capable of sustaining frontier AGI research independent of Western hardware supply chains.
China’s National Supercomputing Internet (Guojia Chaoji Jisuan Wulianwang) represents Beijing’s macro-level initiative to unify geographically fragmented computing centers into a single, national-scale, cloud-native computational utility. Initiated under the guidance of the Ministry of Science and Technology (MOST) and officially launching its central core nodes in mid-2026, the network pools heterogenous compute resources—ranging from traditional exascale supercomputers (Sunway, Tianhe) to domestic AI accelerator clusters (Huawei Ascend, Haiguang DCU)—into an integrated fabric.
1. System Architecture & Network Topology
The architecture solves the geographic decoupling of China's energy supply (concentrated in the west) and digital compute demand (concentrated in eastern coastal cities), operating on a three-tier hub-and-spoke model.
┌─────────────────────────────────────────────────────────────────────────┐
│ EASTERN DEMAND & ORCHESTRATION │
│ Zhengzhou Core Node (Central Routing / Scheduling / API Gateway) │
└────────────────────────────────────┬────────────────────────────────────┘
│
FNTF High-Speed Ultra-Low-Latency Optical Backbone
(34,000+ Miles / Deterministic RDMA Routing / <10ms Latency)
│
┌────────────────────────────────────┴────────────────────────────────────┐
│ WESTERN GREEN COMPUTE FOUNDATION │
│ Ulanqab Megacluster (Gigawatt-Scale Energy / Sovereign AI Clusters) │
└─────────────────────────────────────────────────────────────────────────┘
- The Virtualized Resource Layer: The platform aggregates CPU cores and specialized AI accelerators into a unified, software-defined memory and scheduling environment. Using standardized abstraction layers, it translates heterogeneous hardware stacks (such as Huawei’s CANN/TileLang, Sugon’s scaleFabric, and Baidu’s Kunlunxin XPU) into unified execution environments.
- Low-Latency Optical Backbone (FNTF Protocol): Built atop the Future Network Test Facility (FNTF) backbone—a dedicated, multi-thousand-mile fiber optic network connecting over 40 major cities—the architecture utilizes deterministic routing and RDMA over Converged Ethernet (RoCE). This enables remote clusters to achieve up to 98% of the operational parallel efficiency of a single centralized data center.
- Algorithmic Load Balancing ("Eastern Data, Western Computing"): The network dynamically routes compute tasks based on latency tolerance. Real-time inference and latency-sensitive industrial control are executed at regional edge nodes, while massive parameter pre-training and offline data crunching are dispatched automatically to renewable-powered western hubs.
2. Strategic Computing Hubs: Zhengzhou & Ulanqab
While the network spans dozens of regional nodes (including Tianjin, Guangzhou, Chengdu, and Shanghai), Zhengzhou and Ulanqab serve as its two structural pillars:
Zhengzhou: The Central Coordination & Control Node
Located in Henan Province, Zhengzhou functions as the operational brain and central dispatching core of the National Supercomputing Internet.
- Orchestration & Dispatch: Houses the core routing engines that manage multi-tenant job scheduling, user authentication, cross-provincial bandwidth allocation, and software container distribution.
- Massive Aggregation Capacity: Zhengzhou coordinates an integrated resource pool that links over 3.5 million CPU cores and 250,000+ AI accelerator cards across national research centers, serving as the primary interconnection hub for central government initiatives, university research, and enterprise API routing.
Ulanqab: The Gigawatt-Scale Green Compute Hub
Situated in Inner Mongolia, Ulanqab ("China’s Digital Grassland") serves as the physical muscle for massive model training and raw compute scaling.
- Energy-to-Compute Parity: Leverages abundant wind and solar resources alongside a cold climate to achieve a Power Usage Effectiveness (PUE) below 1.10. Facilities like Envision’s Galaxy Campus target up to 2 Gigawatts (2GW) of dedicated compute capacity.
- Sovereign Megaclusters: Ulanqab hosts nearly 90 major data center projects, serving as the main deployment bed for giant sovereign clusters—including DeepSeek’s 160,000-unit Huawei Ascend 950DT data center and gigawatt-scale domestic non-GPU clusters designed to bypass foreign lithography embargoes.
3. Commercial Access Models & Marketplace Economics
To transition supercomputing from an exclusive academic resource to a commercial public utility, the platform operates a unified commercial platform designed like an "App Store for Compute."
Commercial Access Pipeline
┌───────────────────────────┐ ┌───────────────────────────┐ ┌───────────────────────────┐
│ Token/Credit Marketplace │ ──► │ Uniform API Gateway │ ──► │ Decentralized Execution │
│ (0.1 RMB / 1M Tokens) │ │ (PyTorch, MindSpore, REST)│ │ (Zhengzhou Router ➔ Hubs) │
└───────────────────────────┘ └───────────────────────────┘ └───────────────────────────┘
- Standardized Token-Based Pricing: Compute is monetized using standardized API token metrics rather than legacy raw server-hour rentals. Users purchase standardized compute units (e.g., promotional baseline rates of 0.1 RMB per 1 million tokens for foundation model inference), making costs predictable for startups and academic institutions.
- Software-as-a-Service (SaaS) & Agent Marketplaces: Rather than manually managing bare-metal clusters, developers access pre-configured execution environments for leading open-weights models (DeepSeek V4, Alibaba Qwen series, Z.ai) directly via web portals or REST APIs. The platform also hosts domain-specific scientific research agents (such as SClaw) for automated drug discovery, climate modeling, and materials science.
- Enterprise Hardware Subsidies ("Compute Vouchers"): To encourage domestic enterprise adoption of non-CUDA sovereign hardware, municipal governments and the NDRC issue regional Compute Vouchers (Jisuanquan). Small-to-medium enterprises (SMEs) can redeem these digital vouchers on the National Supercomputing Internet platform to offset up to 30–50% of their training and inference costs when utilizing domestic chip clusters (Ascend, Zhenwu, DCU).
China’s "Eastern Data, Western Computing" (Dongshu Xisuan) strategy is Beijing’s macro-economic and industrial infrastructure framework designed to solve a fundamental geographic imbalance: the concentration of digital demand and economic activity in eastern coastal cities, contrasted with the abundance of land, cold climate, and renewable energy resources in the western provinces.
Officially launched in early 2022 by the National Development and Reform Commission (NDRC), Cyberspace Administration of China (CAC), and Ministry of Industry and Information Technology (MIIT), the initiative reached full operational execution by 2026 as gigawatt-scale data center hubs came online.
1. Primary Policy Goals
The strategy serves three primary state objectives:
Core Policy Pillars of Dongshu Xisuan
ENERGY-COMPUTE BALANCING DIGITAL SOVEREIGNTY REGIONAL EQUALIZATION
┌──────────────────────────────┐ ┌────────────────────────────┐ ┌────────────────────────────┐
│ Re-route energy-intensive │ │ Build sovereign, scale-out │ │ Channel eastern capital and│
│ compute workloads to western │ │ datacenter infrastructure │ │ high-tech investment into │
│ wind, solar, & hydro hubs. │ │ safe from external shocks. │ │ less-developed western regions│
└──────────────────────────────┘ └────────────────────────────┘ └────────────────────────────┘
- Carbon Neutrality & Energy Efficiency: Traditional data centers in Shanghai, Beijing, and Shenzhen faced severe power grid constraints and high cooling costs. By channeling compute workloads westward, China leverages cheap renewable energy (wind and solar in Inner Mongolia and Gansu; hydroelectricity in Guizhou and Sichuan) to achieve strict Power Usage Effectiveness (PUE) targets (PUE < 1.15 for new western facilities).
- Industrial Scale & Sovereign Resilience: The initiative establishes a unified, national-scale computational power grid that connects non-GPU domestic chip clusters (Huawei Ascend, Alibaba T-Head, Sugon/Haiguang) into massive, resilient megaclusters (such as the 100,000-chip superclusters in Ulanqab).
- Closing the East-West Economic Divide: Actively redirects capital expenditures from affluent coastal provinces into inland infrastructure, transforming resource-rich western regions into high-tech digital nodes.
2. Regional Hub Divisions: The "8+10" National Network
The strategy organizes the nation into 8 National Computing Hubs containing 10 Designated Data Center Clusters, creating a structured division of labor based on network latency tolerance.
National Computing Hub Divisions ("Dongshu Xisuan")
EASTERN HUB (Latency-Sensitive) WESTERN HUBS (Capacity & Power)
┌───────────────────────────────────┐ ┌───────────────────────────────────┐
│ • Beijing-Tianjin-Hebei (Zhangjiakou) │ │ • Inner Mongolia (Ulanqab / Hohhot)│
│ • Yangtze River Delta (Wuhu/Yangpu)│ ──► │ • Guizhou (Gui'an) │
│ • Guangdong-HK-Macao (Shaoguan) │ │ • Gansu (Qingyang) │
│ • Chengdu-Chongqing │ │ • Ningxia (Zhongwei) │
└───────────────────────────────────┘ └───────────────────────────────────┘
Eastern Hubs (Real-Time & Low-Latency Execution)
- Hubs: Beijing-Tianjin-Hebei, Yangtze River Delta, Guangdong-Hong Kong-Macao Greater Bay Area, and Chengdu-Chongqing.
- Functional Focus: Reserved strictly for latency-sensitive, real-time workloads requiring sub-10ms response times. This includes financial algorithmic trading, industrial IoT control, autonomous driving inference, video streaming, and real-time user-facing API queries.
Western Hubs (Exascale Pre-Training & Offline Storage)
- Hubs: Inner Mongolia, Guizhou, Gansu, and Ningxia.
- Functional Focus: Dedicated to capacity-heavy, latency-tolerant workloads where execution speed is dictated by compute throughput rather than network ping. This includes:
- Multi-trillion parameter foundation model pre-training (e.g., DeepSeek and Qwen runs).
- Synthetic data generation and model distillation.
- Scientific supercomputing, climate modeling, and offline video/data archiving.
3. Economic Impacts & Market Dynamics
The deployment of Dongshu Xisuan has reshaped China's technology capex, energy consumption, and corporate AI strategies:
Economic Flow & Incentive Dynamics
EASTERN ENTERPRISES COMPUTE VOUCHER SYSTEM WESTERN INFRASTRUCTURE
┌──────────────────────────┐ ┌──────────────────────────┐ ┌──────────────────────────┐
│ Startups, Cloud Platforms│ ──► │ Municipal "Compute │ ──► │ Gigawatt-Scale Green │
│ & AI Research Labs │ │ Vouchers" (30-50% Subsidy)│ │ Datacenters (Ulanqab) │
└──────────────────────────┘ └──────────────────────────┘ └──────────────────────────┘
- The $560 Billion Infrastructure Multiplier: The national initiative has catalyzed over $560 billion (RMB 4 trillion+) in cumulative investment across data center construction, high-voltage power transmission, ultra-low latency fiber-optic backbones (FNTF), and domestic chip procurement.
- Lowering the Cost of Frontier AI Training: By positioning training clusters directly adjacent to gigawatt-scale wind and solar farms in places like Ulanqab and Zhongwei, local operational expenditure (OpEx) for power drops by 30% to 50%. This cost reduction directly underpins the ability of Chinese labs (such as DeepSeek and Alibaba) to offer inference and API tokens at fractions of Western prices (~$0.10–$0.87 per million tokens).
- The "Compute Voucher" (Jisuanquan) Economy: To force domestic adoption of western clusters, municipal governments issue digital Compute Vouchers to local startups and university labs. These vouchers subsidize 30% to 50% of the cost when an enterprise dispatches its training workloads to non-CUDA sovereign clusters located in western hubs.
- Energy Transition Alignment: Data centers now serve as dynamic load balancers for western renewable energy grids. During peak solar and wind production hours, Western megaclusters scale up background batch-training jobs, absorbing excess electricity that would otherwise be wasted due to power grid curtailment.
Strategic Takeaway
Dongshu Xisuan is not merely a datacenter construction plan—it is the industrial engine enabling China's AI self-reliance. By converting western renewable energy into sovereign compute capacity, Beijing has built a physical infrastructure layer capable of powering multi-trillion parameter model training on domestic non-GPU chips, completely insulated from foreign hardware embargoes.
By Jugo Ikeji-Uju.
admin@ubuntusafa.com
sappertekinc@gmail.com
+++++++++++++++++++
Sponsored by. vesselping.com
Real-Time Global Tracking — Monitor 50,000+ commercial vessels live across 160+ maritime nations. (CTA: Learn More)
Predictive AI Analytics — Stay ahead of supply chain delays with automated route prediction & risk scoring. (CTA: Try Pro Free)
Port Congestion & Trade Lane Insights — Get deep actionable data on key shipping routes and port turnaround times. (CTA: View Plans)
Setup in 60 Seconds — From $0/mo free plans to dedicated business support. Join 500+ maritime professionals today at vesselping.com. (CTA: Sign Up)
No comments:
Post a Comment