The launch of DeepSeek V4, one of China’s most advanced open-source large language models, has sent ripples through the domestic semiconductor industry. As US export controls continue to restrict access to top-tier Nvidia chips, Chinese AI developers are increasingly reliant on homegrown hardware. The release of V4 has catalyzed a rush among domestic chipmakers to ensure their processors are fully compatible and optimized for the new model, highlighting the rapid maturation of China’s AI hardware ecosystem and the growing viability of a fully domestic AI stack.
According to a recent report by the South China Morning Post, at least eight domestic chip architectures have confirmed compatibility with DeepSeek V4 through FlagOS, a cross-chip software system platform developed by the Beijing Academy of Artificial Intelligence. This coordinated effort underscores a strategic push to build a robust, self-sufficient AI infrastructure capable of supporting frontier models without reliance on foreign hardware.
Huawei Leads the Charge
Unsurprisingly, Huawei is at the forefront of this wave of adaptation. The company announced that DeepSeek V4 is fully adapted to its Ascend 950PR chip platform. Huawei claims that the Ascend 950PR delivers up to 2.87 times the single-card inference performance of Nvidia’s H20, a chip specifically designed for the Chinese market to comply with US export rules.
Furthermore, Huawei stated that its full line-up of AI processors, including the A2, A3, and 950 series, has completed compatibility testing. The optimization has reportedly improved multimodal generation efficiency by 60% and reduced deployment costs to approximately one-tenth of comparable GPT-based services. This performance leap has driven significant demand, with tech giants like Alibaba, ByteDance, and Tencent collectively placing orders for hundreds of thousands of Ascend 950 processors following the V4 launch.
(Related: Chinese Tech Giants Scramble for Huawei Ascend Chips as DeepSeek V4 Triggers Supply Crunch)
The Rising Stars: Cambricon, Moore Threads, and Metax
Beyond Huawei, several other domestic chipmakers achieved “day-one” compatibility with DeepSeek V4, signaling significant improvements in their software ecosystems.
Cambricon, a leading AI chip designer listed on the Shanghai Stock Exchange, achieved full-stack adaptation on the day of release and simultaneously open-sourced its deployment code. The company attributed this rapid integration to its NeuWare software ecosystem, which has been steadily improving to rival Nvidia’s CUDA platform. Cambricon’s decision to open-source its deployment code is strategically significant, as it invites the developer community to contribute to and build on its platform, accelerating the growth of its ecosystem.
Moore Threads, known for its MUSA architecture, announced that its flagship MTT S5000 GPU completed day-one adaptation. Crucially, the chip supports native FP8 precision and extended context lengths, essential features for running massive models like V4 efficiently. Moore Threads has been positioning itself as a full-stack GPU company, developing both hardware and the software tools needed to program it.
Metax (Muxi), a Shanghai-based GPU designer, also achieved day-one compatibility for its Xiyun series chips via the FlagOS platform. Metax utilizes its KernelSwift system for core operator optimization, working closely with the Shanghai AI Laboratory to enhance performance. The company has benefited from strong backing by the Shanghai municipal government, which sees domestic chip development as a strategic priority.
(Related: DeepSeek V4 Triggers Broad Reassessment of China’s AI Chip Stocks, from Cambricon to SMIC)
A Broad Ecosystem Effort
The South China Morning Post provides an extended list of compatible hardware, illustrating the breadth of China’s semiconductor efforts. Here are the additional names on the list:
T-Head (Alibaba): Formed from C-SKY Microsystems and the DAMO Academy chip team, T-Head has shipped 470,000 GPU chips as of February 2026, generating an annualized revenue of approximately 10 billion yuan ($1.5 billion). Its integration with DeepSeek V4 allows Alibaba Cloud customers to run the model on Alibaba’s own silicon, reducing dependence on third-party hardware.
Kunlunxin (Baidu-backed): Its mass-produced AI chips, already widely deployed in Baidu’s search infrastructure and Apollo Go autonomous driving systems, were included in the FlagOS compatibility release. The integration demonstrates the versatility of Kunlunxin’s architecture across both inference and training workloads.
Hygon: The only domestic firm with a permanent x86 license from AMD, Hygon produces high-end CPUs and deep computing units (DCUs) like the Shensuan No. 3 (BW1000), which are now V4-compatible. Hygon’s x86 compatibility gives it a unique advantage in enterprise deployments where software portability is a priority.
Enflame: Its fourth-generation L600 chip, featuring an architecture closer to a purpose-built AI accelerator (similar to Google’s TPU), was adapted to support V4 Pro and Flash models at native FP8 precision. Enflame has been particularly focused on training workloads, where its architecture offers efficiency advantages.
Iluvatar CoreX: The Shanghai-based general-purpose GPU company also announced day-one compatibility with the full V4 suite, rounding out a comprehensive domestic hardware ecosystem.
The Strategic Imperative
The rapid adaptation of DeepSeek V4 across multiple domestic architectures is a critical milestone for China’s AI industry. It demonstrates that Chinese chipmakers are not only designing capable hardware but are also overcoming the software ecosystem challenges that have historically hindered the adoption of non-Nvidia chips.
The software ecosystem, including the libraries, compilers, and tools that allow developers to write code that runs efficiently on a given chip, has long been Nvidia’s most powerful moat. CUDA, Nvidia’s programming platform, has been the industry standard for over a decade, and the vast majority of AI software has been written to run on it. Chinese chips that lack CUDA compatibility have faced significant adoption barriers, regardless of their raw performance.
The coordinated effort around FlagOS and the individual software stacks of companies like Cambricon and Moore Threads represents a serious attempt to overcome this barrier. By ensuring that a model as widely used as DeepSeek V4 runs efficiently on domestic hardware, these companies are creating a compelling reason for Chinese AI developers to consider alternatives to Nvidia.
(Related: DeepSeek V4 Goes Fully Open-Source; Huawei Confirms Ascend Chip Compatibility in Separate Statement)
As geopolitical tensions persist and access to foreign technology remains constrained, the synergy between domestic models like DeepSeek V4 and homegrown silicon will be the defining factor in China’s ability to maintain its momentum in the global AI race.