In the coming weeks on East Eight Time, the news that DeepSeek V4 will officially adapt to Huawei and Cambricon chips is seen as a key move in the landscape of AI chips in China. Against the backdrop of the U.S. continually escalating export controls on high-end GPUs, Chinese domestic AI computing power and chip supply have long been "choked," putting practical pressure on the continuity and security of large model training and inference. At this time, DeepSeek has not only chosen to shift towards domestic chip adaptation with the new generation of models but, according to The Information, multiple Chinese tech companies have pre-purchased hundreds of thousands of Huawei AI chips, which has been assessed by outsiders as an "important milestone in China's semiconductor self-sufficiency." The symbolic meaning of this action far exceeds a single commercial collaboration; it resembles a systemic breakthrough experiment that has been forced to accelerate under regulatory constraints.

From V3 turbulence to V4 betting on domestic chips

During the V3 and R1 phases, DeepSeek has already proven its technological presence on a global scale—the model performance and cost structure triggered a wave of sentiment fluctuations among overseas tech stocks and computing power concept stocks, which momentarily led to a repricing of the competitive landscape of large models on Wall Street and the Asian market. At that time, like most leading models, DeepSeek still heavily relied on the overseas high-end GPU ecosystem, with training and inference infrastructure primarily anchored in the "U.S. supply chain."

By the time V4 arrived, the path began to undergo a substantial shift: the narrative that once focused on scheduling and optimizing overseas GPUs transformed into a reconfiguration of the software stack surrounding domestic AI chip adaptation. This generation of products is clearly positioned to be usable on Huawei and Cambricon chips, meaning the computing power foundation has changed from "default overseas" to "actively introducing domestic options," no longer merely a supplement but written into one of the core goals of version planning.

DeepSeek's choice of Huawei and Cambricon is not an emotional "politically correct" decision but a strategic balance under multiple practical constraints: on one hand, U.S. controls on high-end GPUs have cut off predictable incremental supply, making any medium to long-term planning based on overseas chips full of uncertainty; on the other hand, Huawei and Cambricon have already formed a relatively mature AI chip product line and customer base domestically, capable of undertaking large-scale model adaptation's engineering and ecological conditions. For DeepSeek, this is a bet on redistributing survival risks and growth space onto the domestic supply chain in a constrained environment.

Rewriting underlying code: After power has been choked

To enable a leading large model to run stably on different chip architectures is far from as simple as "changing a few parameters." Adapting to domestic AI chips means that the DeepSeek team needs to rewrite or deeply modify the underlying operators and system stack: from matrix operations, communication scheduling to memory management, all must be custom-optimized for Huawei and Cambricon's instruction sets, parallel architectures, compilers, and driver features. This is a systematic project that continuously consumes top engineering resources.

The research report mentions that Huawei and Cambricon are conducting joint technical adaptation, which effectively connects the entire chain from chips, drivers, middleware to frameworks and models. By binding to DeepSeek's V4 project, this collaboration no longer stays at the "laboratory demo" level but directly faces pressure testing in high-load, long-cycle, real business scenarios. Once V4 can run stably and at scale on domestic chips, it will signify that the domestic chip ecosystem has begun to shift from "single-point breakthroughs" to "systematic collaboration".

This series of technological transformations is hard to understand without considering the U.S. controls on high-end GPUs. It is precisely because of the ongoing "power choking" that Chinese manufacturers have been forced to compress the adaptation and substitution rhythms originally planned for the medium to long term into a shorter time window. In a certain sense, DeepSeek V4's collaboration with Huawei and Cambricon is an innovation experiment that has been compelled: under the premise of unstable external supply, thoroughly opening up the software architecture for deep binding and pressure validation of local chips.

The power war behind hundreds of thousands of orders

According to The Information, Chinese tech companies have pre-purchased hundreds of thousands of Huawei AI chips, providing a very realistic market scenario for DeepSeek V4's domestic adaptation: these chips will not just be for "stock display," but will be the computing power base that cloud vendors and large model companies compete for in the coming years. Whoever can run usable and practical large models on domestic chips first will gain the advantage in this local computing power war.

With limited domestic chip production capacity, the priority game between cloud vendors and model companies will become increasingly intense. On one end, there are cloud vendors eager to encapsulate chips from Huawei and others into their cloud services and create integrated computing power products; on the other end are AI companies like DeepSeek that need massive computing power to iterate models and occupy cognitive heights. Who secures more chips and who receives more early deep adaptation support will determine their service performance, costs, and ecological appeal.

Market sentiment has already labeled this pre-purchase of hundreds of thousands of chips as an "important milestone in China's semiconductor self-sufficiency" and "significant progress for Huawei in China's AI chip ecosystem." These phrases affirm the reality of the orders while also being a concentrated restoration of confidence in the industry: under the premise of uncertain external supply, domestic capital and enterprises are willing to bet real money on local chips, which means the question of "whether the domestic supply chain can support the long run of AI large models" is transitioning from a hypothetical issue to a practical one.

Gaps in U.S. controls: Chinese AI

From a broader perspective, DeepSeek V4's adaptation to domestic chips is a systematic response and pressure test to U.S. technology controls. The U.S. seeks to "choke" the pace of China's AI development at the source of computing power by restricting high-end GPU exports, while leading models like V4 actively turning to domestic chips attempts to reconstruct a parallel path within the existing gaps: enhancing the usability and efficiency of local computing power without directly touching restricted hardware.

The problem is that publicly available channels lack detailed performance comparison data: outsiders see neither fine parameter differences between different domestic chips nor the matching results against overseas GPUs. This lack of transparency in information causes market sentiment to oscillate between "comparable" and "just alternatives." Some people believe that as long as software and hardware are well-coordinated, domestic solutions could potentially catch up or even locally surpass in specific scenarios; others doubt that domestic chips still have clear gaps from international leaders in high-end training and extreme inference scenarios.

Under the long-term regulatory and technological game, China's medium to long-term path for AI may not have a single answer but could coexist with "more models turning to domestic chips" and "multiple parallel lines." On one hand, many general and industry-specific models will complete training and deployment on domestic chips to seek a balance of cost, compliance, and supply chain security; on the other hand, for a few cutting-edge research pursuits seeking extreme performance, there may still be a retained dependency on some overseas computing power under compliance. DeepSeek V4's choice reflects an acceleration of the former path—using the practical migration of leading models to open a larger imaginative space for the domestic chip ecosystem.

Repeated release timelines and market expectation tug-of-war

The research briefing mentions that the release date of DeepSeek V4 was rumored to be delayed from the original plan due to technical adjustments, and this information is currently still in a "to be verified" state; that is, the specific timeline and internal decision-making process have not been detailed to the outside. However, even a rumored delay is sufficient to amplify questions about the maturity of technology and stability of the ecosystem in the market: Is adapting to domestic chips more challenging than expected? Are the underlying operators, frameworks, and toolchains already sufficiently robust?

From the outside, the change in release timelines from clear to wavering, then to the relatively broad time window of "within the next few weeks" is itself a tug-of-war of expectation management. On one hand, DeepSeek needs to prove its reliability in technological route and delivery rhythm; on the other hand, any delays in adaptations around the domestic supply chain will be magnified as signals that "the domestic ecosystem is not yet ready," triggering chain reactions at the industry level.

Many industry observers interpreting this change in rhythm have given the keyword: "Better to be slightly slower and ensure the stability of the domestic supply chain." From this perspective, a delay does not equal failure; rather, it is a choice by the team to prioritize stability and maintainability in the face of complex engineering transformations and high-intensity testing. For a leading model that carries high expectations, if V4 is to take on the role of a benchmark case for domestic chips, then spending more time on compatibility, reliability, and operational tools might hold more value than rushing to release at a certain point.

If V4 runs stably: Chinese AI counts

If DeepSeek V4 can run long-term and stably on Huawei and Cambricon chips, its significance for the autonomy of Chinese AI computing power will reflect in both symbolic and practical dimensions. On the symbolic level, this is a sample submitted by China under the pressure of U.S. technological control—leading models no longer bound to overseas GPUs but completing a closed loop from training to inference on local chips. On the practical level, it will force a tighter feedback loop between domestic chips, system software, cloud services, and industry applications, accelerating the transformation from "usable" to "well usable."

Looking ahead, the rhythm of more models and more industry applications migrating to domestic chips will depend on both technological feasibility and commercial calculations. In sectors like finance, government, enterprises, industry, and education, if there is a desire to build a more controllable AI infrastructure under compliance and security requirements, local chip solutions will naturally be prioritized. However, the migration process will encounter several bottlenecks: including the development thresholds posed by an incomplete toolchain, insufficient performance margins in certain scenarios, and consistency issues between different chips and cloud platforms, all of which will be exposed during large-scale deployment.

To assess how far this path can go, three types of subsequent observation indicators can be focused on: firstly, the completeness of the ecological toolchain—whether the compiler, framework, and tuning tools are friendly enough to allow more model teams to "plug and play"; secondly, the stability of chip supply—beyond hundreds of thousands of orders, whether manufacturers like Huawei and Cambricon can provide predictable commitments in terms of capacity, cost, and delivery cycles; thirdly, the evolution of the regulatory environment—whether U.S. controls continue to tighten, and whether Chinese policies on data boundaries and computing power construction allow for more space for domestic solutions.

DeepSeek's collaboration with Huawei and Cambricon is just one strike in the long run for autonomy of Chinese AI computing power, but if V4 can run stably on domestic chips, this strike may be enough to change the rhythm of the entire game thereafter.

Join our community to discuss together and become stronger!
Official Telegram community: https://t.me/aicoincn
AiCoin Chinese Twitter: https://x.com/AiCoinzh

OKX Benefits Group: https://aicoin.com/link/chat?cid=l61eM4owQ
Binance Benefits Group: https://aicoin.com/link/chat?cid=ynr7d1P6Z

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。