Overview
The latest TOP500 list released at SC26 kept the Lingsheng system at the National Supercomputing Center in Shenzhen at world number one, clocking 2.198 exaflops on the Linpack double-precision benchmark — about 21% ahead of the United States’ El Capitan — while retaining the top spot on HPCG. Built on the domestic Lingkun platform and LX2 processors, the machine features a fully homegrown stack spanning chips, interconnects, storage and system software.
Domestic computing power is now being validated beyond benchmarks. On November 17, Moore Threads announced it had completed training adaptation for Moonshot AI’s Kimi K3.1 Pro using its MTT S5000 training-and-inference card and MUSA software stack, building a native FSDP2 multimodal training path covering model loading, sharding and gradient synchronization. Kimi K3.1 Pro is a 2.8-trillion-parameter open-source flagship model.
Chip architecture also advanced: Li Auto’s M100 architecture paper was accepted into the industry track of ISCA 2026. The 5nm automotive-grade chip delivers 1,280 TOPS with over 82% utilization and entered mass production in vehicles in May. The company said it will gradually open the chip to external customers.
Background & Interpretation
1. Technological Background and Evolution
Supercomputer evaluation is broadening. Linpack measures peak floating-point throughput, HPCG better reflects memory-intensive real-world loads, and AI workloads further stress interconnect bandwidth, software maturity and scaling efficiency. Holding both top spots suggests domestic systems have built coordinated hardware and systems-software capability rather than relying on isolated component breakthroughs.
2. Core Drivers and Underlying Mechanisms
Training adaptation is fundamentally software engineering. Training a 2.8-trillion-parameter model requires combined tensor, pipeline and expert parallelism, and any communication overhead is amplified at scale. A native FSDP2 path on the MTT S5000 moves domestic GPUs from “capable of inference” to “capable of training.” Whether domestic AI chips achieve scale depends less on paper TFLOPS than on compilers, communication libraries and framework compatibility.
3. Market Landscape and Value-Chain Impact
The three developments point in one direction: a layered domestic supply chain spanning supercomputers, training cards and automotive chips. Supercomputers serve research and foundation-model pretraining, training-inference cards serve commercial iteration, and automotive silicon pushes compute to the edge. Competition is shifting from single-card peak performance to system capability across chips, interconnects, software stacks and model adaptation.
Implications & Outlook
Near term, more model-chip adaptation cases are likely, driven by cost and supply security. Over the medium term, whether externally supplied automotive chips attract third-party customers will test commercialization. Watch the progress of domestic training cards in mainstream frameworks, the opening of supercomputing resources to industrial users, and the actual share of model training running on domestic hardware.
Takeaways for Industry Participants
Chip vendors should treat software stacks and toolchains as first-class investments, prioritizing native paths in mainstream frameworks. Model companies can build abstraction layers for multi-chip compatibility to reduce single-supplier dependence. Enterprise buyers should evaluate real-task performance per watt and total cost of ownership rather than peak specifications.
This column compiles industry information and shares technical perspectives; it does not constitute investment advice.