Nvidia CEO Jensen Huang warns optimized DeepSeek models on Huawei Ascend chips pose strategic threat to US AI leadership.
The AMW Read
Updates the DeepSeek case study regarding hardware optimization and signals a high-significance structural shift toward a bifurcated US-CN tech stack (geopolitics/silicon).
Nvidia CEO Jensen Huang warns optimized DeepSeek models on Huawei Ascend chips pose strategic threat to US AI leadership.
Nvidia CEO Jensen Huang stated on the Dwarkesh Podcast that if Chinese AI lab DeepSeek optimizes its upcoming V4 foundation model for Huawei Technologies' Ascend 950PR processors, it would represent 'a horrible outcome' for the United States. Huang framed this as a scenario where a divergent Chinese tech stack, combining DeepSeek's software with Huawei's hardware, could become the global standard as AI proliferates, enabling China to surpass the US. This commentary arrives as DeepSeek prepares to launch its V4 model, with conflicting reports on its training hardware: Reuters suggested it used Nvidia's restricted Blackwell chips, while The Information reported it will run inference on Huawei's latest Ascend processors.
This public warning from the world's most valuable chipmaker's CEO underscores the intensifying geopolitical dimension of AI infrastructure competition. The market implication is a potential bifurcation into US-led (Nvidia GPUs, CUDA ecosystem) and China-led (Huawei Ascend, competitive AI models) technology stacks. For enterprise clients and global developers, this risks creating compatibility and supply chain complexities. Huang's argument that China's 'abundant' energy and large researcher pool could compensate for potentially inferior chips highlights that raw compute access is not the sole determinant in the AI race, shifting focus to systemic advantages like talent and operational scale.
A grounded expert take recognizes this as a pivotal moment for the AI hardware and model ecosystem. Huang's statement is both a genuine risk assessment and a strategic narrative aimed at US policymakers, advocating for calibrated engagement over total decoupling. The specific concern is DeepSeek V4—a leading open-weight model—becoming a performance benchmark for Huawei's Ascend platform, validating an alternative to CUDA and creating a viable exportable stack. For the AI market, the immediate watchpoint is the V4 launch's technical details, which will clarify the depth of the Huawei-DeepSeek optimization and its real-world performance versus Western counterparts, directly impacting global cloud AI service and hardware investment strategies.
