[Four Questions]
I. What Does the 147% Growth First Prove?
II. As Hardware and Super Nodes Gradually Homogenize, What Still Bottlenecks Computing Power Deployment?
III. How Does MUSA Reduce Software Migration Costs for Existing CUDA Projects?
IV. How Does PD Heterogeneity Enable Synergy Between New GPUs and Existing Cards in Data Centers?
Figure 1 | Overview of Current Period Revenue and the Delivery of the 660 Million RMB Kuae Contract. Data compiled from Moore Threads' 2026 Semi-Annual Report.
I. What Does the 147% Growth First Prove?
The answer lies primarily in cloud products and Kuae delivery. In the first half of the year, revenue from cloud intelligent computing products reached 1.693 billion RMB, accounting for 97.5% of total revenue; the 660 million RMB Kuae contract signed in March was delivered and recognized as revenue within the reporting period.
Since the semi-annual report does not separately list the accounting revenue and gross profit corresponding to this contract, the 660 million RMB can only be regarded as the contract scale and cannot be directly converted into the current period's revenue proportion. What can be confirmed is that the 147% growth stems from the release of computing power demand for large models, the stable supply of S5000 and Kuae, and the smooth conversion of orders into revenue.
II. As Hardware and Super Nodes Gradually Homogenize, What Still Bottlenecks Computing Power Deployment?
Let us first take a quick look at the hardware foundation: the S5000 single card delivers up to 1000 TFLOPS of AI dense computing power, equipped with 80GB of VRAM and 1.6TB/s memory bandwidth, supporting FP8 to FP64. Eight cards can form a fully intra-node interconnect via MTLink. According to the metrics published by Moore Threads, it achieves over 74% of the benchmark products in selected CV and LLM training scenarios, with Llama 3 70B training MFU exceeding 60%, and a linear scaling efficiency of 95% for 10,000-card level training.
These metrics indicate that the S5000 has the foundation to enter large model training and large-scale networking. However, forming super nodes with GPUs, interconnects, collective communications, scheduling, training software, and RAS is already the common route for mainstream manufacturers. What Kuae truly proves is that Moore Threads can combine these capabilities into a complete system that is acceptable, schedulable, and maintainable by customers, and complete large-scale deliveries according to the contract.
However, cluster deployment does not mean the end of the transition: customers still need to migrate CUDA projects that have been running for years, and also handle the existing GPUs in the data center that still have usage value. The former is a software sunk cost, and the latter is a hardware sunk cost.
Figure 2 | Kuae Delivery, MUSA Basic Capabilities, and PD Heterogeneity Solutions Are at Different Evidence Levels.
III. How Does MUSA Reduce Software Migration Costs for Existing CUDA Projects?
After the hardware enters the data center, the original CUDA projects still need to be connected to the new software stack.
On the surface, an AI project uses PyTorch or vLLM, but underneath, it is often connected to CUDA operators, communication libraries, compilation scripts, and custom extensions; changing cards also means redirecting these calls to MUSA. The semi-annual report explicitly lists torch_musa, MUSA SDK, and acceleration libraries as mature capabilities supporting Kuae's 1,000-card and 10,000-card cluster training and inference, and regards reducing migration costs as a critical barrier to GPU commercialization. This shows that MUSA is already part of the current cluster delivery.
On this basis, Moore Threads has also prepared migration entry points of varying depths: MUSIFY modifies source code, MUSA Mapping maps APIs at compile time, torchada transfers from the Python call layer, and vLLM-MUSA accesses from the inference framework. They target different project conditions, breaking down a comprehensive rewrite into multiple optional entry points. Among them, MUSIFY and torch_musa can be found in the semi-annual report, while other entry points come from developer documentation and code repositories; the semi-annual report has not yet quantified the orders, revenue, or maintenance cost reductions brought independently by these migration tools.
Figure 3 | Depending on Project Conditions, CUDA Projects Can Access MUSA at Different Levels.
IV. How Does PD Heterogeneity Enable Synergy Between New GPUs and Existing Cards in Data Centers?
Intelligent computing centers are usually built in phases, and GPUs from different generations and manufacturers coexist in the data center; the input and output lengths of online requests are also constantly changing, and the proportion of resources required for Prefill and Decode fluctuates with traffic.
Traditional deployment allows compute-intensive Prefill and memory-intensive Decode to share a set of cards, which is prone to queuing on one side and idling on the other. PD separation turns the two stages into independent resource pools, scaling up separately according to time to first token (TTFT) and time per output token (TPOT); cross-vendor heterogeneity then allows different card types to handle the stages they are best at.
Figure 4 | PD Heterogeneity Deploys Prefill and Decode in Separate Pools and Connects the Same Inference Request via KV Cache.
Therefore, for intelligent computing centers that bear online inference, long contexts, and mixed card types with SLA requirements, PD is more like a general scheduling capability. The semi-annual report only disclosed that the company launched a PD heterogeneous separation solution to improve resource utilization and provide a hybrid deployment path that balances cost, performance, and flexibility, without disclosing customers, contracts, or revenue.
More specific cross-vendor Prefill and Decode division of labor comes from the joint white paper released by Moore Threads and SiliconFlow in August.
The white paper provides an engineering example: 4 S5000s are responsible for Prefill, and 2 HXXX are responsible for Decode. Both ends uniformly adopt FP8 and pass KV Cache via Mooncake RDMA. Under the testing metrics of both parties, the conventional concurrent time to first token is 1 to 6 seconds, and the time per output token is 0.011 to 0.020 seconds, overall benchmarking against a homogeneous cluster composed of 4 HXXX.
This set of results proves the engineering feasibility under joint testing conditions, but it cannot be concluded that PD heterogeneity contributed to the current period's revenue; its long-term stability and large-scale commercial effects still need to be verified by the production data of named projects.
Conclusion
The 147% growth directly corresponds to the demand for full-function GPUs, the acceleration of Kuae commercialization, and stable supply. Although MUSA has not formed a separate revenue item, it has already supported Kuae clusters and model adaptation, belonging to the basic capabilities of current product commercialization; whether multi-entry migration can further shorten the customer switching cycle still lacks quantitative cases. PD heterogeneity, on the other hand, remains at the level of solution disclosure and joint testing, with no customer and revenue evidence yet.
By separating the three layers of relationships, we can not only explain where the current period's revenue comes from but also see clearly how Moore Threads plans to expand the scope of deployment in the next stage.
References:
[1] Moore Threads: 2026 Semi-Annual Report. [2] Moore Threads: Announcement on Signing a Major Contract for Daily Operations (2026-009). [3] Moore Threads: MTT S5000 Product Page and Kuae Intelligent Computing Cluster Technical Materials. [4] Moore Threads Developer Documentation: MUSIFY and MUSA Mapping. [5] Moore Threads: torchada, torch_musa, and vLLM-MUSA Code Repositories. [6] Moore Threads, SiliconFlow: "Technical White Paper on the Practice of Mixed Deployment of PD Separation and Heterogeneous Computing Power", August 2026. [7] Moore Threads: Mooncake Transfer Engine MUSA and PD Separation Access Instructions.
#DeepChip Alliance #GBA Chip Expo #Semiconductor Industry Research #Moore Threads #Kuae #MUSA #PD #GPU