In the AI computing power circle, NVIDIA GPUs have long been the go-to choice for training large domestic models. However, concerns over potential 'chokeholds' have kept industry players on edge. Recently, domestic computing power has achieved a remarkable breakthrough: large model companies are quietly initiating a 'chip swap,' replacing overseas chips with domestic AI chips for the pre-training cores of trillion-parameter large models. This marks a significant step forward for domestic computing power, transitioning from 'just running' to 'running brilliantly.'

From Meituan open-sourcing the world's first trillion-parameter large model trained with domestic computing power, to the official launch of the Guangdong-Hong Kong-Macao Greater Bay Area's 10,000-card intelligent computing cluster, and the accelerated deployment of domestic computing power ecosystems by companies like iFLYTEK and Baidu, a comprehensive industrial breakthrough centered on 'training domestic models with domestic chips' is well underway. This is not just a technological breakthrough; it is crucial for China's AI industry to achieve true self-reliance and sustainable development.

01 Milestone Breakthrough: Domestic Computing Power Completes Training of Trillion-Parameter Model from Scratch

On July 6th, Meituan made waves across the industry by officially open-sourcing the LongCat-2.0 large model with 1.6 trillion parameters, featuring a Mixture of Experts (MoE) architecture. What many don't know is that this model was trained entirely on over 50,000 domestically produced computing chips, taking just over a month to become the industry's first trillion-parameter ultra-large-scale model trained with domestic computing power, breaking the overseas chip monopoly in large model pre-training.

Almost simultaneously, on July 9th, the Guangdong-Hong Kong-Macao Greater Bay Area's first 10,000-card intelligent computing cluster for 'training domestic models with domestic chips' officially debuted in Shaoguan, Guangdong. Operated by China Telecom with a total investment of approximately 5.5 billion yuan, the cluster deploys 30 Huawei Ascend 384 super nodes, housing 11,520 Ascend 910C chips. Officials confirmed that it is fully capable of handling the full-process training tasks of trillion-parameter large models, serving as a significant benchmark for the large-scale deployment of domestic computing power.

Additionally, the 10,000-card cluster built with Moore Threads' S5000 GPUs, set for mass production in 2025, has already entered commercial use and is assisting clients in training trillion-parameter models. Zhan Molei, a senior principal analyst at Omdia, stated that the successful training of large models on 10,000-card clusters signifies a 'from scratch' breakthrough for domestic AI chips in the training phase—full-process pre-training of trillion-parameter MoE models has been proven feasible at the engineering level.

The significance of this breakthrough extends beyond the technical realm. Previously, domestic AI chips were primarily used for model fine-tuning, post-training, or pre-training of small to medium-parameter models, while pre-training of large-scale parameter models was dominated by international vendors like NVIDIA. This was a 'tough nut to crack' for achieving self-reliance in domestic AI chips. Now, multiple practical cases have demonstrated that domestic computing power is capable of supporting ultra-large-scale model training.

02 Two Approaches to Deployment: Heavy Investment vs. Cautious Exploration

As domestic computing power gradually rises, large model companies in China have made different choices, mainly divided into two camps: the 'heavy investment' camp represented by Meituan and iFLYTEK, which Decisive betting (boldly bets) on domestic computing power, and the 'cautious exploration' camp, such as Baidu and Zhipu, which adopt a more reserved stance. Regardless of the approach, both are steadily advancing the 'chip swap' initiative.

Let's start with the 'heavy investment' camp—Meituan and iFLYTEK. In fact, as early as 2023, they began deploying domestic AI training computing power. Meituan collaborated with domestic computing power vendors on 'model-chip synergy' R&D, progressing from small-scale testing to ultra-large-scale training. The LongCat-2.0 model was trained entirely using AI ASIC chips, with Huawei's Collective Communication Library (HCCL) enhancing training stability. Although the per-card memory capacity of the chips used was smaller than that of NVIDIA's H800, optimization strategies successfully overcame the memory bottleneck.

iFLYTEK is a staunch supporter of domestic computing power. In October 2023, it launched the 'Feixing No.1,' China's first 10,000-card domestic intelligent computing platform, with training performance exceeding 90% of NVIDIA's A800 cluster of the same scale. By September 2025, iFLYTEK plans to invest 2.4 billion yuan in leasing domestic computing power to further expand its capacity. Currently, multiple models in its Spark series are trained on domestic AI chip clusters and have successfully adapted to various domestic chips, including Cambricon's MLU590 and Zhongke Haiguang's 'Shensuan No.3' BW1000.

iFLYTEK stated that its core motivation for choosing a self-reliant and controllable domestic route is to avoid long-term dependence on overseas sources for critical training, inference, and iteration capabilities, thereby mitigating the impacts of external computing power supply fluctuations, cost volatility, and ecological constraints. This 'bet-the-farm' approach has made it a vital force in improve (improving) the domestic computing power ecosystem.

In contrast, companies like Baidu and Zhipu have adopted a more cautious 'trial' approach. Baidu completed training for the important version 5.1 of its Wenxin model using the computing cluster of its majority-owned subsidiary, Kunlunxin, and trained three multimodal models, including the video generation model 'Baidu Steam Engine,' on Kunlunxin clusters. However, these did not involve the core versions of its flagship large models. Zhipu, in collaboration with Huawei Ascend, completed the full-process pre-training of the image generation model GLM-Image, validating the feasibility of domestic computing power for pre-training.

These differing deployment strategies reflect companies' considerations of the technical maturity of domestic computing power and the current industry landscape—domestic computing power is transitioning from 'pilot applications' to 'large-scale deployment.' Whether through heavy investment or cautious exploration, both approaches are accumulating practical experience for improve (improving) the domestic computing power ecosystem.

03 Bottlenecks and Costs: Domestic Computing Power Still Faces Multiple Hurdles

Despite breakthroughs, domestic computing power still lags significantly behind international leaders. The 'chip swap' process encounters numerous challenges and costs, and a complete 'breakaway from NVIDIA' remains a long road ahead.

Computing power is the first hurdle. Take Huawei's latest Ascend 950DT chip, which delivers 1,034 TFLOPS of computing power at FP8 precision, while NVIDIA's H100 achieves 1,979 TFLOPS under the same conditions—a clear gap in per-card computing power. To compensate, domestic chip vendors are promoting super node solutions, combining thousands of chips into a supercomputing node via high-speed interconnectivity, exemplified by Huawei's 384 super node and the upcoming 950 super node.

Memory capacity and bandwidth limitations are another unavoidable issue. Most mainstream domestic AI chips launched before 2026 feature only 64GB or 96GB of memory, whereas NVIDIA's H200 boasts 141GB of memory and a staggering 4.8TB/s memory bandwidth. During LongCat-2.0 training, Meituan faced significant challenges due to insufficient memory and had to implement specialized memory optimizations. Fortunately, improvements are underway, with next-generation products like Ascend 950DT, Alibaba's T-Head Zhenwu M890, and MetaX Xiyi C600 equipped with 144GB of memory and enhanced bandwidth.

Beyond hardware gaps, shortcomings in software ecosystems and engineering optimization are even more pronounced. iFLYTEK noted that NVIDIA's advantage lies not just in its chips but also in its mature software ecosystem, training frameworks, and toolchains. Training large models with domestic computing power tests the entire system's ability to operate stably and efficiently over the long term, involving multiple components such as chips, servers, networks, communication, operators, training frameworks, cluster scheduling, and fault recovery. Any inefficiency or instability in these areas can undermine overall training effectiveness.

Especially in 10,000-card clusters, frequent cross-card and cross-node communication in scenarios like MoE, long contexts, and intelligent agent reinforcement learning can waste significant computing power if communication efficiency is low. Additionally, domestic chips currently suffer from higher fault rates compared to NVIDIA's, posing challenges for large-scale training.

These technical bottlenecks ultimately translate into costs for enterprises. Liu Qingfeng, Chairman of iFLYTEK, revealed that training the Spark X1 model on domestic computing power platforms required an additional two months for adaptation. Moreover, the higher usage costs and longer training times of domestic chips may delay model releases. Zhan Molei also noted that current domestic solutions succeed primarily through active adaptation to MoE architectures and substantial customized engineering efforts, rather than directly replacing NVIDIA's CUDA ecosystem—a full substitution remains unrealistic for now.

04 Collaborative Breakthrough: 'Model-Chip Synergy' Becomes the Industry's Development Focus

Facing numerous challenges, 'collaboration between domestic chips and models' has become a consensus in the industry. In recent years, interactions between domestic model companies and chip vendors have intensified, shifting from passive adaptation to proactive collaboration and simultaneous progress, even achieving 'Day 0 native adaptation' where models are compatible with domestic chips upon release.

Wei Liang, Vice President of the China Academy of Information and Communications Technology, stated that this reflects the increasingly close link between innovative breakthroughs in frontier models and foundational hardware and software. However, current collaboration mostly focuses on the inference phase—ensuring trained models run smoothly on domestic chips—while deep collaboration in the pre-training phase remains limited. iFLYTEK also mentioned that while domestic computing power for inference is relatively mature, completing large model pre-training, especially for cutting-edge techniques, remains engineering-intensive.

True 'model-chip synergy' involves full-chain deep integration. iFLYTEK cited examples such as designing more suitable MoE architectures and attention mechanisms based on chip architectural characteristics, optimizing operators for key processes to enhance computational and memory access efficiency, and refining communication mechanisms to align with cluster communication bottlenecks and reduce performance losses. Such deep collaboration has become a key criterion for enterprises when selecting domestic computing power platforms.

Several notable examples of deep collaboration have emerged in the industry. iFLYTEK and Huawei Ascend have formed a strong partnership, jointly tackling key technologies like efficient model structures and hybrid attention mechanisms on the Ascend 950 platform, with plans to release a flagship large model rivaling the industry's most advanced offerings by October 2026. Suiyuan Technology collaborated with clients to complete pre-training for multiple Qianwen series models and small to medium-parameter Tencent Hunyuan models based on its fourth-generation L600 product, with ongoing efforts for large-parameter model pre-training.

More importantly, this collaboration is mutually beneficial. iFLYTEK noted that the large model training process exposes real issues in chips and systems, driving upgrades in AI chips and software ecosystems. For instance, new trends like long contexts and intelligent agents demand higher memory and bandwidth, pushing chip iterations in storage and interconnectivity. Performance issues with operators and software toolchains identified during training also prompt chip vendors to improve supporting tools, while faults in 10,000-card training scenarios accelerate advancements in cluster management and stability.

05 Industry Outlook: Self-Reliance Becomes the Core Development Direction

From technological breakthroughs to ecosystem improve (improvement), the collaborative development of domestic AI chips and large models is gradually outlining a clear path for China's AI industry to achieve self-reliance. Despite ongoing challenges, the industry's momentum is unstoppable.

At the policy level, the involvement of local governments and telecommunications operators has significantly boosted domestic computing power development. The Shaoguan 10,000-card intelligent computing cluster not only addresses Guangdong's computing power shortage for ultra-large-parameter model training but also serves as a core infrastructure supporting general AI innovation and industrial digital transformation across the province, offering replicable and scalable experiences for other regions. As computing power infrastructure continues to improve nationwide, the application scenarios for domestic computing power will expand.

Technologically, domestic chip performance is rapidly improving, and software ecosystems are maturing. New-generation domestic AI chips are narrowing the gap with international leaders in memory, bandwidth, and computing power, while vendors like Huawei, Cambricon, and Kunlunxin continue to strengthen their technical capabilities. Meanwhile, deep collaboration between model vendors and chip vendors is accelerating solutions to engineering optimization and ecological adaptation challenges, enabling domestic computing power to transition from 'capable of training' to 'training well' and 'training fast.'

From an industrial perspective, forming a closed loop of 'training domestic models with domestic chips' will fundamentally alter China's AI industry's reliance on overseas computing power, reducing external risks to industrial development. As more companies join the 'chip swap' movement, the market size for domestic computing power will grow, gradually forming a complete industrial ecosystem encompassing 'chip R&D—model training—scenario applications' and propelling China's AI industry to new heights.

Of course, we must recognize that the rise of domestic computing power will not happen overnight. Achieving full self-reliance requires sustained efforts. However, the successful training of trillion-parameter large models and the deployment of 10,000-card intelligent computing clusters have injected strong confidence into the industry. Driven by 'model-chip synergy,' domestic AI chips are poised to establish a firm foothold in the global computing power competition, laying a solid foundation for the high-quality development of China's AI industry.

This 'chip replacement' project is not just a technological breakthrough, but also an awakening of the entire industry. As the 'engine roar' of domestic computing power grows louder, the path to independent control of our country's AI industry will certainly become wider and wider.