Huawei Technologies is making significant progress in closing the technology gap with U.S. chipmaker Nvidia through a deepening partnership with artificial intelligence (AI) model start-up DeepSeek. According to a recent report by research firm SemiAnalysis, the collaboration is advancing both hardware and software development for Huawei’s Ascend 950PR chips, signaling a potential challenge to Nvidia’s dominant position in the AI chip ecosystem.
DeepSeek has developed software tools specifically optimized for the Ascend chip platform, introducing six software modules reminiscent of those previously created for Nvidia chips. Among these is an Ascend-compatible version of TileLang, a specialized programming language that simplifies the creation of high-performance kernels—critical components for efficiently operating chip hardware. DeepSeek described the Ascend version of TileLang as a step toward fostering a more “independent and controllable” domestic software ecosystem for graphics processing units.
SemiAnalysis noted that the increasing integration between DeepSeek and Huawei is likely to result in DeepSeek relying predominantly on Huawei’s mature technology stack by 2027 or 2028. This trend aligns with Beijing’s broader push for technological self-reliance amid U.S. export controls restricting China’s access to advanced chips and chipmaking equipment, including Nvidia products. Nvidia’s CEO Jensen Huang has publicly expressed concern over these developments, warning that DeepSeek’s optimization of AI models for Huawei hardware could give China a competitive edge in AI technology.
Despite the advances, Huawei’s Ascend 950PR, unveiled in March 2026, still trails Nvidia’s H100 chip—which began shipping in 2022—in raw performance metrics. Nevertheless, Huawei engineers are achieving near-peak efficiency with their best Ascend kernels, reaching approximately 99.7 percent of their theoretical performance ceiling, according to the report.
The partnership’s progress is evident in how DeepSeek tests model-specific workloads on Huawei’s Ascend ecosystem and provides early access to its foundational AI models. Huawei has announced “Day-0” inference support—running AI models at launch rather than after additional optimization—for major updates such as DeepSeek V4 and V4.1.
Challenges remain, particularly around hardware compatibility beyond DeepSeek’s specialized supernode configurations. Most software optimizations have targeted large-scale supernodes, leaving developers confronting issues when attempting to run Ascend chips on more common x86 host processors instead of Huawei’s native architectures.
Still, the ongoing collaboration presents avenues for broader industry benefits. SemiAnalysis suggested that other AI operators working with DeepSeek could adopt existing kernels and deployment practices, while research labs might develop new models using shared compiler capabilities. If these users contribute back improvements, the compounded gains could accelerate innovation in China’s domestic AI chip ecosystem.