Moore Threads Technology Co.,Ltd. is actively working to reshape its corporate identity.
For the past five years, the company has been widely recognized as a Chinese GPU manufacturer. However, it is now deliberately moving away from this label. A clear demonstration of this shift was its annual product launch event last month. Over nearly two hours, the company unveiled six major products and initiatives.
These included the Kuae supercomputing cluster capable of scaling to ten thousand GPUs, the MTT AICUBE and MTT AIBOOK smart terminals powered by its in-house "Yangtze" SoC, the "Xiaomai" digital world agent, the MT Lambda embodied AI simulation platform, and continued advancements in its MUSA ecosystem. This comprehensive showcase presented a full-stack intelligent computing matrix covering cloud, edge, and device applications.
Shifting Focus from Core GPU Business
It raises a question: as a company traditionally seen as a GPU specialist, why did this event feature no new chip announcements, instead expanding into areas like embodied AI, home hubs, and hardware?
From Ten-Thousand to Hundred-Thousand GPU Clusters
A year and a half ago, the company's primary narrative was straightforward: GPUs. When it went public on the STAR Market in late 2025, its valuation of 300 billion yuan was anchored by the Kuae ten-thousand-card intelligent computing cluster.
Subsequent performance data for this system has been strong, demonstrating high model training efficiency and scalability. A significant milestone was a 660 million yuan contract signed on March 30, 2026, for a Kuae cluster, representing 44% of the company's total 2025 revenue and proving its capability in large-scale deployments.
Financial results for 2025 and Q1 2026 further solidified the strength of its cloud business. Revenue for 2025 grew 243.37% year-over-year to 1.506 billion yuan, while Q1 2026 revenue reached 738 million yuan, a 155.35% increase, with net profit attributable to shareholders turning positive for the first time since listing.
Buoyed by these results, company leadership has indicated plans for a next-generation hundred-thousand-card cluster. This expansion reflects an optimistic outlook on the exponential growth in computational demand driven by agentic AI applications.
However, this growth path requires massive capital investment. In 2025, R&D expenditure accounted for 86.68% of revenue, an exceptionally high ratio even within the listed chip sector. The question remains whether the GPU-centric model alone can sustain its valuation ceiling.
Furthermore, the company operates in a fiercely competitive global landscape for computing power. Geopolitical factors have made domestic GPUs a strategic necessity in China, creating an opportunity but also intense rivalry among local firms like Moore Threads. Relying solely on the GPU narrative is no longer seen as a sufficient strategy.
Expanding from GPU to Intelligent SoC
While cloud computing acts as a "super brain," device-side products are the "nerve endings" that deliver AI to end-users. Typically, GPU companies' edge strategy stops at providing AI inference chips. Moore Threads Technology Co.,Ltd. is taking a different approach.
Its "Yangtze" SoC is not a simple GPU or NPU but a complete edge AI computing chip integrating CPU, GPU, NPU, and VPU, with heterogeneous AI computing power of 50 TOPS. It positions itself against players like Qualcomm, MediaTek, and Intel, rather than Nvidia.
Based on this SoC, the company has built a product matrix including the MTT AIBOOK (personal AI computing notebook), MTT AICUBE (home AI hub), and MTT E300 (industrial edge AI module). The AICUBE combines an agent, AI PC, and AI NAS, featuring the "Xiaomai" agent and extensive control capabilities.
This is supported by a full software stack, including a native Linux OS, the MTClaw agent framework, an application market, and AI programming tools. This "chip-OS-device-agent" chain covers roles traditionally held by different companies, from Nvidia and Qualcomm to Lenovo and Microsoft, packaged into a unified story of infrastructure for the agent era.
The strategic bet is that the primary workload for future PCs will be AI agents, potentially creating an opening for a Linux-native ecosystem where Windows has long dominated. This is a forward-looking gamble with inherent risks, dependent on whether conversational agents become mainstream and whether Moore Threads' products can gain traction in time.
This move into the device side is not based on existing market success—edge and terminal products contributed a negligible 1.70% to 2025 revenue. Instead, it represents an effort to create a second growth curve for valuation, hedging risks in the highly competitive pure-play GPU arena.
Building an Independent Developer Ecosystem with MUSA
While hardware can be caught up with, building a software ecosystem is a greater challenge. Following a logic similar to Nvidia's CUDA moat, Moore Threads Technology Co.,Ltd. has developed the MUSA ecosystem, a comprehensive GPU computing architecture.
A key achievement is the high compatibility of MUSA SDK 5.1.0 with CUDA 12.8, covering 100% of core math libraries and PyTorch operators. This allows millions of PyTorch developers to migrate models to MUSA with minimal code changes.
Furthermore, MUSA has been integrated as an official backend into major inference frameworks like SGLang and vLLM, providing a more direct compatibility path. The company's developer platform, Moore Academy, has attracted over 450,000 developers and learners, a rapid growth for a six-year-old company, though still an order of magnitude smaller than CUDA's community.
The real challenge for MUSA is moving beyond being a CUDA follower. The company needs to establish unique technical advantages that make developers choose MUSA because it offers a better experience for specific tasks, not just because it is compatible.
The Other Side of High Growth
The shift to a quarterly net profit was seen as a key inflection point. However, this positive result requires scrutiny. The company attributed its 2025 revenue surge to the AI boom and the domestic substitution window created by export restrictions on high-end GPUs.
Yet, the Q1 2026 report shows that non-recurring gains totaled 83.64 million yuan, with government subsidies accounting for 70.06 million yuan (about 84%). Excluding these items, the company would have reported a core operating loss of 54.28 million yuan, albeit a 60.1% year-over-year improvement.
In essence, the reported profitability was achieved with the help of government support; the core business itself has not yet reached self-sufficiency. The true breakthrough will depend on expanding its commercial customer base, continuously improving product performance and stability, and growing the MUSA developer community.
Connecting the Strategic Dots
Looking at Moore Threads' strategic choices holistically, this is not a typical case of a GPU company diversifying horizontally. It is a hardware-tech firm leveraging its core competency in computing to expand downwards into SoCs and upwards into operating systems and agents.
The Kuae cluster in the cloud addresses "where computing power comes from." The "Yangtze" SoC and AICUBE address "where computing power goes." The MUSA ecosystem addresses "why developers stay." The "Xiaomai" agent and MT Lambda platform address "in what form AI ultimately reaches the physical world."
These four layers are interconnected, forming a complete blueprint from silicon to application scenarios, from training to inference, and from the virtual to the physical world. In the long run, Moore Threads is betting on a fundamental premise: AI computing will not remain confined to "training large models" but will permeate every device, physical space, and human-computer interaction.
From this perspective, the company's expansion from the cloud to the edge is not a "crossover" for a GPU firm but a strategic elevation of its ecosystem, aligning with the broader evolution of AI. To ultimately validate this vision, Moore Threads Technology Co.,Ltd. will need not just time, but also a sustainable, self-reinforcing commercial cycle.
Comments