Muxi (沐曦)는 AI 추론 칩과 슈퍼노드 통합 시스템을 공급하며, 칩 공급에서 엔드투엔드 컴퓨팅 플랫폼으로 확장하는 중국의 경로를 선도하고 있습니다.
By July 2026, capital market attention toward MetaX had reached a peak. Fewer than seven months after its IPO, the domestic GPU maker saw multiple brokerages issue dense research reports, expressing strong confidence in its strategic push into the super-node sector.
On July 17, the World Artificial Intelligence Conference (WAIC) 2026 opened, drawing all eyes to MetaX’s booth in Hall H2 of the Expo Exhibition and Convention Center. The “Xi Jing” S-series super-node made its public debut. A display wall showcased the product lineup extending from individual boards to full-cabinet super-nodes, with numerous attendees lingering at the exhibit. This marked MetaX’s first official unveiling of its new super-node product line. Meanwhile, just over ten kilometers away at the Zhangjiang Science Hall, MetaX hosted an AI4S (AI for Science) exhibition, where the newly launched Xi Suo X series demonstrated real-world applications in materials science, life sciences, and other fields.
Capital markets are closely watching MetaX because it has tapped into a critical inflection point in domestic computing power: the shift from merely stacking chips to architecting integrated systems. Understanding the company requires stepping beyond a single-product perspective and viewing it against the broader backdrop of China’s GPU industry evolving from isolated breakthroughs to systemic synergy.
**01 The Limits of Chip Stacking and the Rise of Super-Nodes**
To grasp why super-nodes became the undisputed focus of WAIC 2026, consider the data. Single-chip performance is approaching physical limits. During large model training, cross-node communication overhead now accounts for over 30 percent of total training time: for every 100 cards purchased, roughly 30 spend their time waiting for data. The traditional path of simply adding more cards is rapidly narrowing.
Super-nodes were born to address this bottleneck. By leveraging high-speed interconnect technologies, they integrate dozens or even hundreds of GPUs into a single, low-latency, high-bandwidth collaborative computing system. Compared to traditional GPU clusters, super-nodes deliver significant advantages in communication latency, compute density, and total cost of ownership (TCO). They preserve near-single-card performance during training while boosting inference efficiency by 20 to 50 percent.
Industry observers now regard 2026 as the “Year of Domestic Super-Nodes.” The domestic super-node market is projected to reach approximately 341.4 billion RMB by 2028, representing a compound annual growth rate (CAGR) of 194 percent between 2026 and 2028. At WAIC 2026, Huawei publicly displayed the Atlas 950 SuperPoD, starting at 64 cards per cabinet and scalable to 8,192 NPUs. Sugon introduced its scaleX cluster system capable of supporting 100,000 cards. The competition in domestic computing power has unequivocally escalated from single-card comparisons to system-level capabilities.
According to MetaX, 64 cards represent the most concentrated market demand for super-nodes today. This scale can support large-scale deployments exceeding 10,000 cards and aligns with the requirements of most mainstream large language models. However, larger is not inherently better. Super-nodes only offer a cost advantage when model scales exceed the capacity of traditional 8-card nodes. This is especially true in inference scenarios, where selection must be matched to model volume and precision requirements rather than blindly pursuing larger node sizes.
**02 A Compute Network: MetaX’s Super-Node Product Matrix**
The launch of the Xi Jing S600 fundamentally addresses a long-standing delivery gap in the domestic GPU industry: companies used to ship only chips, leaving server integration, cluster networking, and compute scheduling entirely to the customer. But as AI compute scales to thousands and tens of thousands of cards, clients no longer want just a chip—they require a turnkey compute system. The Xi Jing S600 super-node marks MetaX’s pivotal transition from a chip supplier to a compute system operator.
On the hardware front, the Xi Jing S600 pairs MetaX’s fully self-developed GPGPU with the complete MXMACA software stack. It supports high-density 64-card deployments per cabinet, horizontally scalable to 10,000-card clusters. Several technical details stand out: full-switch interconnect within the cabinet ensures low-latency data communication and flexibly supports multi-dimensional parallel strategies like EP and TP. An orthogonal backplane replaces traditional external cabling, achieving a zero-cable design for critical intra-cabinet data transmission. All GPU units connect directly via the backplane, eliminating additional plugging operations. Delivering the entire cabinet with zero-cable compute and switch nodes significantly compresses construction cycles and maintenance costs. Built on a proprietary full-precision GPGPU architecture, it covers both intelligent and general-purpose compute scenarios. Furthermore, the balanced routing of the orthogonal backplane ensures consistent latency from each GPU unit to the switching circuitry, further enhancing interconnect performance while drastically reducing failure-prone components like cables and optical modules. In large-scale clusters, this zero-cable design yields exponential reliability gains through reduced failure rates.
Equally important is MetaX’s system-level standardization efforts. In March, MetaX, alongside the Shanghai Artificial Intelligence Laboratory, 16 industry partners, and eight universities, published the *White Paper on Super-Node Technology Systems*. The document spans the entire industry chain, covering chip R&D, compute deployment, software adaptation, academic research, and industry applications. Two MetaX researchers played deep roles in drafting the core chapter on scale-up interconnect protocols. From products to standards, MetaX’s super-node strategy has clearly transcended single-product development.
**03 The Moat: Beyond the Chip**
Super-node competition tests comprehensive system-level capabilities. While chip design relies on talent breakthroughs, fusing chips, interconnects, cooling, power delivery, and software into a highly efficient system demands rigorous engineering prowess and accumulated experience. MetaX’s moat rests on four pillars.
First, product layout. The company has established four series—Xi Yun, Xi Si, Xi Cai, and Xi Suo—covering AI training, inference, graphics rendering, and scientific computing. Its entire GPU architecture, core GPU IPs, instruction set architecture (ISA), and accompanying software stack are independently developed. Launched in July 2025, the Xi Yun C600 is the first training-and-inference unified GPU built on fully domestic manufacturing processes. The next-generation Xi Yun C700 benchmarks against international mainstream models and has completed most core designs and functional validations. During an interview with *Semiconductor Industry Focus*, MetaX Senior Vice President Sun Guoliang specifically addressed a market misconception: MetaX’s chips feature native training-and-inference unified designs, with no strict separation between training and inference products. “At every WAIC we attend, we emphasize that our products and chips are unified for both training and inference,” he stated. This clarification directly counters prior market skepticism regarding domestic compute’s alleged inability to handle training workloads.
Second, software ecosystem. The self-developed MXMACA software stack is compatible with over 40 mainstream AI frameworks, stably supports more than 500 AI models, and boasts an open-source developer community exceeding 500,000 members. MetaX has become the officially supported Chinese GPU vendor for vLLM and jointly launched the MXAIE optimization solution with Red Hat, while maintaining deep involvement in upstream projects like PyTorch. At WAIC 2026, ten leading open-source communities—including PyTorch, vLLM, Red Hat, SGLang, and OpenAtom XuanYuan (龙蜥)—co-exhibited at MetaX’s “Open Source Workshop” zone, covering the full software stack from operating systems and cloud-native platforms to inference acceleration and model services. MetaX views rapid software adaptation as a core long-term advantage, possessing strong compatibility with existing mainstream open-source models and emerging models alike, enabling swift responses to iterative demands across different scenarios. High compatibility across its full-stack software allows “Day 0” adaptation for mainstream open-source models—meaning immediate compatibility upon release. This capability stems from a key industry insight: chip iteration cycles typically span two years, whereas model updates occur on a weekly basis. “If your chip lacks versatility, it becomes a disaster.” Recognizing this disparity in iteration cycles, MetaX adheres to a general-purpose GPU roadmap, utilizing fine-grained instruction sets to achieve hardware-software decoupling and ensure seamless adaptation to both current and future models.
Third, system integration. Its self-developed MetaXLink high-speed interconnect technology supports various topologies ranging from 2 to 128 cards. The release of the Xi Jing S600 signals that MetaX now possesses the engineering capability to seamlessly integrate chips, interconnects, cooling, power delivery, and software into ready-to-deploy compute systems.
Fourth, commercial deployment. Guided by its “1+6+X” strategy, MetaX has achieved commercialized applications across six major sectors: finance, healthcare, energy, education and research, transportation, and entertainment. The exhibition area featured real-world deployment cases from leading enterprises in these industries, replacing theoretical demos with proven commercial closed-loop pathways. Beyond smart transportation and embodied AI, MetaX also unveiled its latest initiatives in “space compute.”
MetaX has clearly defined its core positioning: rather than relying on any single major client, it operates as a supplier of general-purpose compute infrastructure. The company is already collaborating with, shipping to, and jointly optimizing systems alongside multiple top-tier internet firms. On ecosystem strategy, MetaX maintains an open, win-win philosophy: “Whether in software or hardware, we choose openness.” Compared to fully vertical integration models, an open architecture can aggregate the best suppliers across each ecosystem niche, making it easier to assemble competitively positioned product portfolios.
Advancing simultaneously across chips, software, systems, and industry applications constructs a deeply systemic moat.
**04 From Laboratory to Exhibition Floor: A Twenty-Year Foundation**
MetaX’s rapid ascent is underpinned by a core team with profound industry pedigree.
Founder Chen Weiliang brings over two decades of GPU expertise, having previously served as Global Head of GPU SoC Design at AMD. Co-founder and CTO Peng Li holds the distinction of being AMD’s first Chinese female scientist. Chief Technical Officer Yang Jian was AMD’s first Chinese Fellow. All three core members possess more than 20 years of high-performance GPU R&D experience and have successfully led teams to tape out and mass-produce fifteen high-performance GPU projects. Over its five-year history, MetaX has achieved one-shot tapeout success for three flagship chips.
On December 17, 2025, MetaX listed on the STAR Market. Between 2023 and 2025, its revenue surged more than thirtyfold. In 2025, its net profit loss narrowed by 43.97 percent year-over-year, while its gross margin expanded to 56.51 percent. Nine financial institutions project that MetaX will return to profitability in 2026.
**05 Conclusion: MetaX and the Dawn of Domestic Compute**
WAIC 2026’s decision to establish intelligent computing as an independent track was no accident. The localization of compute infrastructure is transitioning from a phase of “availability” to one of “excellence.”
MetaX’s moves at WAIC 2026—from debuting the Xi Jing S-series super-nodes to showcasing a full-stack product matrix spanning chips to systems—reflect a continuous release of value. This momentum is driven by two primary forces: the catalytic effect of its WAIC super-node debut, and the sustained explosion of domestic AI compute demand amid constrained overseas high-end chip supplies. This environment has opened a substantial window for domestically autonomous GPUs. The market for large language models, intelligent compute centers, and cloud-based autonomous driving is vast and urgently demanding general-purpose compute.
As the theme banner at MetaX’s exhibition read, “沐晨曦兮，芯力绽放” (As dawn breaks, chip strength blooms). Domestic computing power is entering its own dawn moment. And super-nodes may well be that first ray of light.