A landmark breakthrough has been achieved in the domestic substitution of AI computing power! On January 26, ILUVATAR COREX (09903), which had been listed on the Hong Kong stock market for less than a month, publicly disclosed a four-generation architecture technology roadmap. It clearly outlined a plan to progressively surpass NVIDIA's three core architectures—Hopper, Blackwell, and Rubin—over the next three years. Concurrently released, the "Tongyang" series of edge products demonstrated measured performance that exceeds the industry benchmark, NVIDIA's AGX Orin. For the first time, the company also revealed a list of over 300 benchmark clients across sectors like internet, finance, healthcare, and transportation, along with results from more than 1,000 deployments. These commercialization metrics signal that domestic general-purpose GPUs have officially entered a new phase characterized by "performance surpassing and scaling volume," accelerating the domestic substitution process in the hundred-billion-yuan computing power sector.
As the core engine of new quality productive forces, general-purpose GPUs are regarded as the "crown jewel" of the AI industry, a domain long monopolized by overseas giants. ILUVATAR COREX's dual breakthrough in both technology and commerce not only alleviates the "bottleneck" constraints in the high-end computing power field but also presents a highly valuable core investment target for the technology industry. The underlying technical barriers, commercialization logic, and industry impact warrant in-depth analysis.
The four-generation architecture explicitly targets NVIDIA, aiming for comprehensive surpassing by 2027, with "high-quality computing power" set to redefine industry standards. In the chip industry, an architecture roadmap directly determines a company's long-term competitiveness. The four-generation architecture roadmap presented by Shan Tianyi, Head of AI and Accelerated Computing Technology at ILUVATAR COREX, can be termed a "declaration of overtaking" for domestic computing power: the Tianshu architecture was launched in 2025, already surpassing NVIDIA's current workhorse, Hopper; the Tianxuan and Tianji architectures will be released successively in 2026, with the former matching and the latter fully surpassing NVIDIA's latest Blackwell architecture; the Tianquan architecture will debut in 2027, achieving superiority over NVIDIA's Rubin architecture, after which ILUVATAR COREX will pivot to breakthrough architectural designs, completely abandoning a follow-the-leader R&D approach. Over the next three years, the company will intensively launch multiple products based on this roadmap, continuously raising the ceiling of computing performance.
Addressing industry pain points such as low computing power energy efficiency, high deployment costs, and difficult algorithm adaptation, ILUVATAR COREX proposed three core standards for "high-quality computing power," directly targeting key customer demands: High Efficiency focuses on optimal TCO (Total Cost of Ownership), achieving the best balance between computing cost and performance in complex scenarios through hardware optimization and software synergy; Predictability utilizes precise simulation technology to realize "what-you-see-is-what-you-get" performance before deployment, solving the chronic industry issue of impressive "on-paper specs" versus disappointing real-world performance; Sustainability emphasizes full algorithm compatibility, seamlessly adapting from traditional models to future, unknown algorithm evolutions to protect customers' long-term investment value.
To support this goal, ILUVATAR COREX has created a new paradigm called the AI++ computing power system, building a fully synergistic chain encompassing "hardware foundation + software stack + application ecosystem." This model, based on underlying computational libraries, bridges models and computational scheduling in the middle layer, maximizing hardware potential through software optimization and fundamentally changing the inefficient traditional model of "hardware-heavy, software-light" computing systems.
From a technical perspective, the innovations across the four generations are highly disruptive: The Tianshu architecture achieves dual coverage for high-precision scientific computing and AI computing, with an effective utilization rate of attention mechanism computing power exceeding 90%. Through three core technologies—TPC BroadCast, Instruction Co-Exec, and Dynamic Warp Scheduling—computational efficiency is increased by 60% compared to the industry average, delivering 20% higher performance than NVIDIA's Hopper architecture in scenarios involving the DeepSeek V3 large model. The Tianxuan architecture adds support for ixFP4 precision, further enhancing low-precision computational efficiency; the Tianji architecture achieves full-scenario AI and accelerated computing coverage, breaking down barriers between cloud, edge, and endpoint computing power; the Tianquan architecture incorporates more innovative designs, laying the foundation for breakthrough computing capabilities.
The Tongyang series directly challenges NVIDIA with 300 TOPS computing power and benchmark-crushing real-world tests, positioning domestic solutions to lead the edge market. If the architecture roadmap ensures long-term competitiveness, then the market performance of the product portfolio is key to breaking through in the present. Guo Wei, Vice President of ILUVATAR COREX, officially launched the "Tongyang" series of edge computing products at the event, completing the company's "cloud + edge + endpoint" full-scenario computing power layout. Its core performance directly benchmarks against NVIDIA's AGX Orin, with comprehensive surpassing achieved in actual tests, establishing it as the new "performance benchmark" for domestic edge computing.
The four newly released products precisely cover different application scenarios, forming a differentiated competitive advantage: The Tongyang TY1000 computing module, using a 699-pin interface, integrates industry-grade computing power and an open ecosystem into a "pocket-sized" form factor, leading the industry in deployment flexibility; The Tongyang TY1100 computing module integrates an ARM v9 12-core CPU with a self-developed GPU module, offering ample computing power and strong compatibility to meet diverse scenario needs; The Tongyang TY1100_NX computing terminal, with its large memory design, becomes a high-cost-performance choice, akin to a "pocket powerhouse" for edge computing; The flagship Tongyang TY1200 computing terminal delivers an extreme measured dense computing power of 300 TOPs paired with a compact design, providing core computing support for cutting-edge scenarios like AI PCs and embodied intelligence.
Notably, the claimed computing power for all Tongyang series products refers to measured dense computing power, avoiding any inflation from "paper specs," and covers the 100T-300T computing power range. In benchmark tests across core application scenarios like computer vision, natural language processing, and the DeepSeek 32B large language model, the Tongyang TY1000 demonstrated overall superior performance compared to NVIDIA's AGX Orin, establishing significant advantages in key metrics such as inference latency and energy efficiency ratio. "Our goal is to become number one in domestic high-performance edge computing," Guo Wei stated clearly, indicating that the Tongyang series will serve as the optimal carrier for physical AI to connect the physical world, promoting the transition of generative AI from "articulate" to "capable" physical AI, thereby unlocking new application spaces for edge computing.
Industry analysts point out that the edge computing market is expanding at a compound annual growth rate exceeding 40%, with NVIDIA's AGX Orin long dominating the high-end segment. The performance superiority and scaled deployment of ILUVATAR COREX's Tongyang series not only break the monopoly of overseas giants but also elevate the domestic substitution rate for edge computing to new heights. Its high cost-performance advantage is expected to accelerate the penetration of AI technology into terminal scenarios like industrial control, intelligent transportation, and consumer electronics, catalyzing new opportunities in the hundred-billion-yuan edge intelligence market.
Over 300 clients and 1,000+ deployments mark a closed commercialization loop, signaling that domestic computing power has entered a phase of growth realization. The ultimate value of hard technology must be validated through commercialization. At this event, ILUVATAR COREX systematically disclosed its scaled deployment achievements for the first time, using solid client and delivery data to prove that domestic general-purpose GPUs have moved "from the laboratory to the production line," entering a period of commercial scaling and growth.
Zou Xuan, Vice President of ILUVATAR COREX, revealed that the company's core product, the Tiange general-purpose GPU, has achieved significant results across multiple industries thanks to its three advantages of "high computing power, high efficiency, and high compatibility." It is compatible with over 400 mainstream models, thousands of general operators, and more than 100 custom operators, achieving "same-day operational readiness" for newly released domestic large models, reaching top-tier industry compatibility levels.
In specific industry applications: In internet AI, single-machine performance has doubled while token costs were halved, with a large model adaptation operator reuse rate reaching 95%; In scientific exploration, thousand-card clusters have stably operated for over 1,000 days and are already deployed in research labs at several top domestic universities; In finance, report generation efficiency has increased by 70%, and quantitative analysis response speed accelerated by 30%, making it the core computing support for leading brokerages and fintech companies; In healthcare, structured medical record generation time has been compressed to 30 seconds per record, significantly enhancing clinical diagnosis efficiency.
The newly released Tongyang series edge products have also achieved "immediate deployment upon release": In embodied intelligence, it provides low-latency, highly reliable computing cores for Gelanruo Robots; In industrial intelligence, it has been deployed in multiple smart parks and manufacturing production lines, driving automation upgrades in production processes; In business intelligence, thousands of Luckin Coffee stores have fully deployed the Tongyang solution, optimizing operational efficiency through video stream analysis and consumer data mining; In transportation intelligence, collaborations have been established with 20 leading pilot cities nationwide for "vehicle-road-cloud integration," validating the edge computing solution in vehicle-infrastructure cooperative scenarios.
To date, ILUVATAR COREX has served over 300 paying clients and completed more than 1,000 commercial deployments, covering more than 20 core industries including internet, finance, healthcare, scientific research, industry, and transportation. From a financial perspective, the company has formed a commercial closed loop of "technology R&D - product deployment - scaled profitability." With capacity release and client structure optimization, its performance is expected to enter a high-growth trajectory, potentially becoming an earnings benchmark within the domestic computing power sector.
Ecological synergy and policy support underpin the rise of domestic computing power, highlighting investment opportunities in the trillion-yuan赛道. The breakthrough progress by ILUVATAR COREX results from the resonance of technological accumulation, market demand, and policy support. As a leading enterprise in domestic AI computing, the company has consistently adhered to the path of independent R&D for general-purpose GPUs since its inception, forgoing the short-term but limiting ASIC-specific chip route in favor of the more technically challenging yet flexible GPGPU general route. This strategic choice demonstrates strong foresight in today's era of rapid AI algorithm iteration.
ILUVATAR COREX has maintained high-intensity R&D investment over the years, with cumulative R&D expenditure exceeding 2 billion yuan, providing solid assurance for architectural innovation and product iteration. In terms of ecosystem development, the company signed strategic cooperation agreements with multiple hardware manufacturers and solution providers at the event, further完善ing the domestic computing power ecological closed loop of "chip - hardware - software - application," forming a synergistic industrial合力.
At the policy level, computing power, as a core productive force of the digital economy, has been incorporated into national strategic planning. Various regions have successively introduced support policies related to computing infrastructure construction and domestic chip substitution, providing a favorable development environment for domestic general-purpose GPU enterprises.
For investors, under the dual tailwinds of domestic substitution and exploding AI computing power demand, ILUVATAR COREX, with its three core advantages—fully self-developed technical barriers, full-scenario product layout, and scaled commercial deployment—has become a key target for capturing investment opportunities in the computing power赛道. Looking at industry development, with the rapid growth of emerging scenarios like generative AI, embodied intelligence, and vehicle-road coordination, and according to Frost & Sullivan's public data, the market value of AI computing hardware (primarily GPUs) is projected to reach 3,611.974 billion yuan by 2029, indicating vast space for domestic substitution.
The four-generation architecture roadmap and scaled deployment results unveiled by ILUVATAR COREX not only mark the company's own entry into a fast-development lane but are also set to drive the synergistic development of the entire domestic computing power industry chain, promoting China's position to become more advantageous in the global computing power competition. The story of the rise of domestic computing power, from technological breakthrough to commercial realization, is now unfolding.