Why AI Supernodes and 10,000‑GPU Clusters Will Dominate 2025

AI Supernodes: The Performance Ceiling

AI supernodes aim to break both compute density and communication efficiency limits, becoming the "super‑engine" for large‑model training. By 2025 two parallel tracks are expected: Huawei’s Ascend 384 supernode (NPU‑centric) and Shanghai‑based LightSphere X (GPU‑centric) optical‑interconnect architecture. The Ascend 384 links 384 NPUs with 192 CPUs via MatrixLink, achieving 300 PFLOPS—surpassing Nvidia’s GB200 NVL72 system (288 PFLOPS). LightSphere X deploys a distributed optical‑switch with 1 P liquid‑cooled GPU modules, supporting a 2‑k‑card scale, sub‑microsecond latency, and delivering >2.5× performance over traditional clusters, with multimodal training efficiency up to 3×.

Ten‑Thousand‑Card GPU Clusters: The Practical Era

2025 marks three key transitions for "ten‑thousand‑card" clusters: from laboratory validation to massive deployment, from foreign‑dependency to domestic substitution, and from single‑compute to heterogeneous GPU scheduling. Globally, Meta has built a 24,576‑GPU cluster, xAI is deploying 100 k H100 GPUs, and Microsoft plans to supply 300 k GB200 GPUs to OpenAI. Domestically, Sugon’s ScaleX series integrates 10,240 AI acceleration cards, delivering >5 EFLOPS and supporting mixed‑brand GPU deployments for multiple full‑parameter large‑model trainings.

Chinese innovators introduce notable advances: ScaleX’s single‑cabinet 640‑card design uses immersion liquid‑cooling to achieve a PUE of 1.04; its proprietary ScaleFabric network provides 400 Gb/s bandwidth, raising GPU utilization by 55%. A partnership between Muxi and the Chinese Academy of Sciences built a 1‑k‑card GPU cluster based on the fully‑domestic Xiyun C600, proving that domestic GPU clusters can handle large‑model pre‑training and boosting material‑R&D efficiency by 3‑6×, with breakthrough results in protein research.

GPU: The Core Compute Engine

In 2025 the GPU market remains dominated by Nvidia, whose Q2 PC‑GPU AIB shipments reach 11.6 million units (94% market share) and the RTX 5000 series occupies half of the performance leaderboard. Meanwhile, China’s "four little dragons"—Moore Threads, Muxi, Biren Tech, and Suiyuan Tech—have all gone public, delivering fully domestic IP, wafer, and packaging. Muxi’s C600 matches flagship international performance and supports ten‑thousand‑card scaling.

Technical differentiation includes Biren’s 1 P liquid‑cooled module as the core of optical‑interconnect supernodes, Suiyuan’s ASIC‑focused S60 (over 100 k orders) for high‑frequency scenarios, and Muxi’s "1+6+X" strategy covering finance, healthcare, and five other verticals. Market reaction was strong: Muxi’s IPO share price surged 700% and its valuation exceeded ¥300 bn, confirming strong expectations for domestic GPUs.

Knowledge Bases: The Core of Enterprise Intelligence

AI hallucination errors can reach 38%, prompting a shift to Retrieval‑Augmented Generation (RAG). By 2025, 75% of enterprise AI applications will be built on RAG architectures, and GPU acceleration triples document‑processing speed. A case study of an appliance manufacturer shows a 50% improvement in maintenance efficiency after adopting RAG for technical‑drawing retrieval.

Multimodal fusion now supports 15 data formats; Shanghai Fourth People’s Hospital integrated 30 k clinical cases into a GPU cluster, generating five treatment plans in 15 seconds and reducing misdiagnosis by 41%.

Security‑focused deployment combines local GPU servers with government‑cloud services, meeting ISO‑27001 audit requirements with millisecond‑level logging.

Embodied Intelligence: From Demo to Industry

In 2025 embodied intelligence is featured in the national government work report and the 15‑year plan, elevating it to a strategic industry. GPU miniaturization and cost reduction, together with domestic component supply, enable humanoid robots to transition from technology showcases to commercial products.

The industry logic shifts from demo‑centric to value‑centric B‑side scenarios; GPU energy‑efficiency and cost become decisive competitive factors. Analysts predict a "shuffling year" in 2026 where firms that tightly integrate GPUs with robotic hardware and ensure long‑term stability will lead the market.

Optical Interconnect: The Transmission Revolution

AI compute demand is projected to grow 300% YoY in 2025, pushing GPU cluster sizes from ten‑thousand to hundred‑thousand cards. Optical interconnect evolves from an auxiliary link to a core infrastructure, moving the focus from CPO to "optical‑in‑cabinet + GPU direct" solutions.

Scenario upgrades include "optical‑in‑cabinet" designs that boost compute density beyond 100 PFLOPS per rack, addressing a projected $12 bn annual market demand. LightSphere X’s distributed optical switching enables dynamic GPU topology reconfiguration, breaking the power‑wall limitation of single‑rack deployments.

Architectural redesigns reduce transmission loss from 30% (copper) to under 5% (full‑optical). IBM’s test data center reports a 40% latency reduction for GPU‑to‑GPU communication, and Google plans to deploy 100 flexible‑optic GPU cabinets by 2026.

The supply chain benefits: optical‑module leaders such as Zhongji Xuchuang capture 42% of revenue from GPU‑cluster interconnects, while XinYi Sheng’s 112 Gbps products achieve a 48% gross margin. Domestic firms are expected to exceed 60% of the global optical‑module market share.

OpenClaw: From Single Agent to "Personal Silicon Company"

OpenClaw is an open‑source AI agent that runs locally, can invoke large models, operating‑system calls, and external tools to automate data, file, and code tasks. Its default configuration is insecure, granting high privileges and exposing many vulnerabilities; it can be hijacked to leak data or take over systems. The Ministry of Industry and Information Technology has issued a security warning about these risks.

OpenClaw is also employed for AI‑generated rumor automation, becoming a key tool for industrial‑scale misinformation production.

Further Reading

https://mp.weixin.qq.com/s?__biz=MzAxNzU3NjcxOA==∣=2650766668&idx=1&sn=494a8b3cb79e23571b0512a043f08ee3

https://mp.weixin.qq.com/s?__biz=MzAxNzU3NjcxOA==∣=2650766542&idx=1&sn=702c77e25ef4548494e5bf84deeced66

https://mp.weixin.qq.com/s?__biz=MzAxNzU3NjcxOA==∣=2650764412&idx=1&sn=ec008a870210cef5299f5d74949205b8

https://mp.weixin.qq.com/s?__biz=MzAxNzU3NjcxOA==∣=2650764388&idx=1&sn=bba55c51af7fd86dcc2a23734331bc33

https://mp.weixin.qq.com/s?__biz=MzAxNzU3NjcxOA==∣=2650763920&idx=1&sn=9162cdb14ae8ca04b6ca019c789bdfc5

https://mp.weixin.qq.com/s?__biz=MzAxNzU3NjcxOA==∣=2650763456&idx=1&sn=1950ccf8ca69046c6c18dba668ce6402

https://mp.weixin.qq.com/s?__biz=MzAxNzU3NjcxOA==∣=2650763419&idx=1&sn=8f5d17e4df3b074d3cc3eac3bd0ab06a

https://mp.weixin.qq.com/s?__biz=MzAxNzU3NjcxOA==∣=2650763153&idx=1&sn=152cc7d19dfd4e52c9b5729654b1c594

https://mp.weixin.qq.com/s?__biz=MzAxNzU3NjcxOA==∣=2650762620&idx=1&sn=example