ILUVATAR COREX(09903) releases the fourth-generation architecture "overtaking" roadmap, Tangyang edge performance surpasses NVIDIA, landing more than 300 benchmark customers, trillion-dollar track pattern changes.

AI computing power domestic alternative ushered in a milestone breakthrough! On January 26, ILUVATAR COREX (stock code: 09903), which had just been listed on the Hong Kong Stock Exchange for less than a month, made a high-profile disclosure of its four-generation architecture technology roadmap. It clearly stated that within three years, it will gradually surpass the three major core architectures of NVIDIA Hopper, Blackwell, and Rubin. At the same time, it released the "Tongyang" series of edge products, whose performance in practical testing surpassed the industry benchmark NVIDIA AGX Orin. For the first time, it also revealed a list of over 300 benchmark customers and over 1000 deployment achievements covering Internet, finance, medical, and transportation fields. This announcement signifies that domestically produced general-purpose GPUs have officially entered a new phase of "performance surpassing, and scale ramping up" in the process of domestic substitution in the race for trillion-level computing power. As the core engine of new productivity, general-purpose GPUs are considered the "crown jewel" of the AI industry, which has long been monopolized by overseas giants. The dual breakthrough in technology and business by ILUVATAR COREX not only broke the bottleneck in the high-end computing power field but also provided a highly valuable core target for technology industry investments. The technological barriers, commercial logic, and industry impact behind it are worthy of in-depth analysis. Aim at NVIDIA with the four-generation architecture: achieve comprehensive surpassing by 2027, "high-quality computing power" redefines industry standards In the chip industry, architecture roadmap directly determines a company's long-term competitive strength. The four-generation architecture roadmap unveiled by ILUVATAR COREX's AI and acceleration computing technology leader, Dan Tianyi, can be described as a "declaration of overtaking" for domestic chip architecture: In 2025, the Dayuan Tianchu architecture was launched, which had already surpassed NVIDIA's current main Hopper architecture. In 2026, the Dayuan Tianxuan and Dayuan Tianji dual architecture were released, with the former benchmarking and the latter comprehensively surpassing NVIDIA's latest Blackwell architecture. In 2027, the Dayuan Tianquan architecture was introduced, achieving surpassing NVIDIA's Rubin architecture. After that, ILUVATAR COREX will focus on breakthrough architecture design and completely break away from a follow-the-leader approach to R&D. Over the next three years, the company will launch multiple products based on this roadmap, continuously raising the ceiling of computing power performance. Addressing the current pain points in the industry such as low computing power efficiency, high deployment costs, and difficult algorithm adaptation, ILUVATAR COREX has put forward three core standards for "high-quality computing power", directly addressing customers' core demands: Efficiency focuses on optimizing the best Total Cost of Ownership (TCO), allowing customers to achieve the optimal balance of computing power cost and performance in complex scenarios by hardware optimization and software coordination. It is expected that precise simulation technology can achieve "what you see is what you get" performance before deployment, solving the industry's long-standing problem of computing power products looking good on paper but underperforming in practice. Sustainability emphasizes full algorithm compatibility, seamlessly adapting from traditional models to future unknown algorithm evolution, ensuring long-term investment value for customers. To support this goal, ILUVATAR COREX has created a new paradigm in AI++ computing power systems, building a "hardware base + software stack + application ecology" comprehensive chain collaborative system. This model is based on a foundation of lower-level computing libraries, with an intermediate layer connecting models to computing scheduling, maximizing hardware potential through software optimization, fundamentally changing the inefficient model of traditional computing power systems that are "heavy on hardware and light on software". In terms of technical details, the innovations of the four-generation architecture are highly disruptive: The Dayuan Tianchu architecture achieves high-precision scientific and AI computing coverage, with an attention mechanism computing efficiency utilization rate exceeding 90%. Through the core technologies of TPC BroadCast (Compute Group Broadcast Mechanism), Instruction Co-Exec (Multiple Instruction Parallel Processing System), and Dynamic Warp Scheduling (Dynamic Thread Group Scheduling System), the computing efficiency is improved by 60% compared to the industry average, and in large model scenarios such as DeepSeek V3, the performance exceeds NVIDIA's Hopper architecture by 20%. The Dayuan Tianxuan architecture adds support for ixFP4 precision, further improving low-precision computing efficiency. The Dayuan Tianji architecture achieves full-scenario AI and acceleration computing coverage, breaking through the barriers of cloud, edge, and terminal computing power. The Dayuan Tianquan architecture incorporates more innovative designs, laying the foundation for breakthrough computing. The Tongyang series "strongly competes" with NVIDIA: 300TOPs computing power + crushing practical performance, domestically reaching the peak of the edge market If the architecture roadmap is the guarantee of long-term competitiveness, then the market performance of the product matrix is the key to the current breakthrough. ILUVATAR COREX Vice President Guo Wei officially launched the "Tongyang" series of edge computing power products at the conference, completing a full-scene computing power layout of "cloud + edge + terminal". Its core performance directly competes with NVIDIA AGX Orin, and the practical testing data shows comprehensive overtaking, making it the "performance benchmark" for domestic edge computing power. The four products released this time accurately cover different application scenarios, forming a differentiated competitive advantage: The Tongyang TY1000 computing power module uses a 699-pin interface, integrating industry-level computing power and open ecology in a "pocket-sized" device, leading in deployment flexibility in the industry. The Tongyang TY1100 computing power module integrates ARM v9 12-core CPU and self-developed GPU modules, with ample computing power and strong compatibility, meeting the needs of diverse scenarios. The Tongyang TY1100_NX terminal computing power, with its large memory design, becomes a high-cost-effective option, described as the "small cannon" of edge computing power. The flagship product, Tongyang TY1200 terminal computing power, provides extreme practical dense computing power of 300TOPs, combined with a compact body design, offering core computing power support for frontier scenarios such as Autonomous Intelligent Point Cloud (AIPC) and embodied intelligence. It is worth noting that all products in the Tongyang series claim practical dense computing power without any exaggeration, covering a range of 100T-300T of computing power. In practical testing in core application scenarios such as computer vision, natural language processing, and the DeepSeek 32B large language model, Tongyang TY1000 performs better than NVIDIA AGX Orin in terms of inference latency, efficiency ratio, and other key metrics, creating a significant advantage. "Our goal is to become the number one in edge computing power domestically," Guo Wei stated clearly. The Tongyang series will become the optimal carrier for physical AI connecting the physical world, transforming AI from "being able to speak" to "being able to do things" and opening up new space for edge computing applications. Industry analysts point out that the edge computing power market is expanding at a rapid pace with a compound annual growth rate exceeding 40%, with NVIDIA AGX Orin long holding a dominant position in the high-end market. The performance overtaking and large-scale deployment of the ILUVATAR COREX Tongyang series not only break the monopoly of overseas giants but also increase the domestication rate of edge computing power to new heights. Its high cost-effectiveness advantage will accelerate the penetration of AI technology in end-terminal scenarios such as industrial control, smart transportation, and consumer electronics, creating new opportunities in the trillion-level edge intelligence market. 300+ customers, 1000+ deployments: commercial closed-loop landing, domestic computing power enters the growth realization period The value of hard technology ultimately relies on commercial validation. At this conference, ILUVATAR COREX systematically revealed its achievements in large-scale deployment for the first time, providing real customer and delivery data to prove that the domestically produced general-purpose GPUs have transitioned from the "laboratory to the production line" and entered a period of commercial ramp-up. ILUVATAR COREX Vice President Zou Jian revealed that the company's core product, the Tianyi general-purpose GPU, with its three major advantages of "high computing power, high efficiency, and high compatibility," has achieved significant results in landing in various industries: Compatible with over 400 mainstream models, thousands of general operators, and over 100 custom operators, new large models released domestically could run on the same day, and compatibility has reached the industry's top level. In specific industry applications, the following results have been achieved: In the Internet AI field, single-machine performance doubled, and token costs halved, with a model adaptation operator reuse rate of 95% for large models. In the field of scientific exploration, a cluster of a thousand calculations ran steadily for over a thousand days, and has been deployed in research laboratories at several top domestic universities. In the financial sector, report generation efficiency increased by 70%, and quantitative analysis response speed increased by 30%, becoming the core computing power support for leading securities firms and financial technology companies. In the medical field, the time for generating structured medical records was reduced to 30 seconds per record, significantly improving clinical diagnosis and treatment efficiency. The just-released Tongyang series of edge products also achieved "release to landing": In the field of embodied intelligence, it provided low-latency, high-reliability computing power core to Siasun Robot & Automation; in the field of industrial intelligence, it landed in multiple intelligent parks and manufacturing lines, promoting the automation upgrade of production processes; in the field of business intelligence, the Tongyang solution was fully deployed in thousands of Luckin Coffee stores, optimizing operational efficiency through video stream analysis and consumer data mining; in the field of traffic intelligence, it cooperated with 20 top pilot cities in the "car-road-cloud integration", verifying edge computing power solutions in car-road collaborative scenarios. Up to now, ILUVATAR COREX has served more than 300 paying customers and completed over 1000 commercial deployments, covering more than 20 core industries such as the Internet, finance, medical, research, industry, and transportation. From a financial perspective, the company has formed a commercial closed-loop of "technology R&D - product landing - scalable profitability." With the release of production capacity and optimization of customer structure, performance is expected to enter a high-speed growth channel, becoming the benchmark of the domestic computing power sector. Ecological synergy + policy support: the rise of domestic computing power, highlighting the investment opportunities in the trillion-dollar track The breakthrough progress of ILUVATAR COREX is the result of multiple resonances of technological accumulation, market demand, and policy support. As a leading company in the domestic AI computing field, the company has always adhered to the path of independent research and development of general-purpose GPUs since its establishment, giving up the short-term but restrictively effective ASIC chip path in favor of the GPGPU general path, which has higher technical barriers but greater flexibility. This strategic choice, in today's rapidly evolving AI algorithm landscape, demonstrates strong foresight. ILUVATAR COREX has maintained high-intensity R&D investments for many years, with accumulated R&D expenditures exceeding 2 billion yuan, providing a solid guarantee for architectural innovation and product iteration. In terms of ecosystem construction, ILUVATAR COREX has signed strategic cooperation agreements with several hardware manufacturers and solution providers to further improve the domestic computing power ecosystem of "chip - hardware - software - application" and form collaborative industry strength. At the policy level, computing power as a core productivity of the digital economy has been included in the national strategic layout, and various regions have introduced support policies for computing power infrastructure construction and domestic chip substitution, providing a good development environment for domestic general-purpose GPU enterprises. For investors, with the dual tailwinds of domestic substitution and the explosion of AI computing power demand, ILUVATAR COREX, with its three core advantages of fully self-developed technological barriers, full-scene product layout, and large-scale commercial deployment, has become the core target for seizing the investment opportunities in the computing power track. Looking ahead to the industry's development, with the rapid growth of emerging scenarios such as generative AI, embodied intelligence, and car-road collaboration, according to public data from Frost & Sullivan, the market value of AI computing power hardware (dominated by GPUs) will reach 3,611.974 billion yuan by 2029, with a vast space for domestic substitution. The four-generation architecture roadmap and large-scale deployment achievements presented by ILUVATAR COREX not only signify the company's entry into the fast lane of development but will also drive the coordinated development of the domestic computing power industry chain, furthering China's competitive position in the global computing power competition. From technological breakthroughs to commercial realization, the rise of domestic computing power is currently unfolding its success story.