On August 9, domestic GPU leader Moore Threads Technology Co.,Ltd. (688795.SH) released its 2026 semi-annual report. In the first half of 2026, the company achieved revenue of 1.736 billion yuan, a massive year-on-year increase of 147.42%, already surpassing its total revenue for the full year of 2025. Gross profit reached 989 million yuan, up 103.78% year-on-year, with significant improvements in profitability. Net profit attributable to the parent company and net profit after deducting non-recurring gains and losses narrowed losses by 95.73% and 52.37% respectively compared to the same period last year, indicating a comprehensive acceleration in commercialization.
The strong performance is underpinned by Moore Threads' sustained high-intensity R&D investment, which forms its technological foundation. The report shows that the company's R&D expenditure in the first half of 2026 was 769 million yuan, a year-on-year increase of 38.16%. Since 2022, the company's cumulative R&D investment has approached 5.9 billion yuan. As of June 30, 2026, the company had filed a total of 2,167 patent applications, including 2,049 invention patents, and had been granted 788 patents, with 725 being invention patents, placing it among the top domestic GPU companies. Supported by high R&D investment and a deep patent moat, the company's performance continues its high-growth trajectory.
In its announcement, Moore Threads stated that this growth was primarily driven by strong market demand for full-function GPUs, coupled with the accelerated deployment of its Kua'e Intelligent Computing Cluster and stable supply. This has enhanced its market competitive advantage and commercialization capabilities. The company is reportedly accelerating its penetration into key customer segments such as the internet and telecom operators, leveraging more diverse application scenarios to sustain rapid revenue growth.
As Moore Threads' core growth engine, its flagship AI training and inference computing card, the MTT S5000, continues to scale in mass production. Leveraging its high computing power and full-precision capabilities, the MTT S5000 benchmarks against mainstream international products in large model training and inference. It was among the first to pass the national 'Security and Reliability Assessment,' meeting high industry standards in technical maturity, security, and stability. This provides secure and reliable computing power support for critical sectors such as government, finance, and energy, opening up more market growth opportunities.
The MTT S5000 intelligent computing cluster has achieved large-scale sales, with deployments and deliveries completed in multiple locations including Beijing, Wuxi, and Hangzhou. Key technological dimensions have all achieved breakthroughs, and product performance has received high recognition from customers. In terms of high scalability, as the cluster training scale continues to increase, it can maintain a linear expansion efficiency of up to 95%. In high precision, its training accuracy is on par with mainstream international computing cards, with highly consistent training loss curves. In terms of high stability, the cluster supports checkpoint resumption, with effective training time accounting for over 90% of total training time.
Large model training is the 'ultimate exam' for GPUs. Based on the high-end training capabilities of the Kua'e Intelligent Computing Cluster, Moore Threads has achieved several 'national chip training national models' results. For example, it successfully trained a complete MoE-236B foundational large model from scratch, using a corpus of over 25 trillion tokens. It also achieved the world's first full-stack native training of a 5D world model, 'Peking University EvoPhys-World', which topped the Stanford WorldScore global ranking in the 'World Generation' dimension for 37 consecutive days after its release. Additionally, the first code large model, MusaCoder, trained entirely on the S5000, has been open-sourced, with kernel operator generation performance on par with international SOTA levels.
This series of practices validates that the Kua'e Intelligent Computing Cluster can not only efficiently train large language models but also support cutting-edge fields like embodied intelligence, world models, and physical AI, demonstrating strong general-purpose capabilities. The general computing power advantage of Moore Threads' full-function GPUs extends far beyond cloud-based intelligent computing. As AI moves from the digital world to the physical world, computing power demand is rapidly expanding to the edge and terminal devices.
At the edge, the AI module MTT E300 provides efficient, low-latency, and highly reliable edge AI capabilities for areas such as industrial quality inspection, energy inspection, embodied intelligence, smart cars, and the low-altitude economy. At the terminal, the AI computing power device MTT AIBOOK and the home AI hub MTT AICUBE bring domestic computing power into households. From cloud intelligent computing to edge reasoning and terminal accessibility, Moore Threads has built a general-purpose computing landscape covering the full 'cloud-edge-terminal' scenario.
Simultaneously, leveraging its integrated 'compute, render, and simulate' general-purpose computing base, Moore Threads is carving out a unique ecological niche in frontier tracks like embodied intelligence. As a rare domestic GPU company that has closed the ecosystem loop of 'large model training - simulation - end-side deployment', Moore Threads significantly lowers the development and deployment threshold through its MT Lambda simulation platform and embodied intelligence community. On the industrial implementation front, the company has established an Industrial Embodied Intelligence Innovation Center in Wuxi. It also serves as a co-constructor and industry committee member of the National AI Application Pilot Base (Embodied Intelligence), and has established the 'Embodied Intelligence Computing Power and Simulation Joint Laboratory', deeply participating in the national embodied intelligence industry ecosystem.
As a long-term 'moat' for its technological competitiveness and commercial viability, Moore Threads' proprietary MUSA ecosystem is maturing. Currently, MUSA has achieved comprehensive deep compatibility with the internationally mainstream CUDA ecosystem. It not only achieves 100% compatibility with core math libraries and full compatibility with over 3,000 PyTorch operators but also covers 55 categories of core AI operators. Furthermore, it has received official support from the world's top inference frameworks vLLM and SGLang, enabling Day-0 support for mainstream models, with the developer base now exceeding 800,000 people.
On the same day, Moore Threads announced the initiation of preparations for issuing H-shares and listing on the Hong Kong Stock Exchange, becoming another hard-tech company pursuing an 'A+H' dual listing strategy.
Comments