China’s Tech Giants Avoid Huawei Chips Amid Overheating Problems and NVIDIA’s Dominance with CUDA Lock-Ins

This article does not provide investment advice. The author holds no stakes in any of the stocks referenced herein.

Huawei’s Challenges in the AI Chip Market

Huawei aimed to reduce China’s reliance on NVIDIA with its Ascend 910C GPUs, yet faces significant obstacles in achieving this goal. The stronghold of NVIDIA’s ecosystem, bolstered by the CUDA software platform, along with various shortcomings on Huawei’s part, have contributed to this inertia.

Recent reports from The Information reveal that key Chinese tech players such as TikTok’s parent company ByteDance, Alibaba, and Tencent have not placed substantial orders for Huawei’s AI chips.

Factors Impeding Adoption of Huawei’s 910C GPUs

Several critical factors have come together, creating considerable inertia around the adoption of Huawei’s 910C GPUs. The company’s focus has shifted primarily toward catering to larger State-Owned Entities (SOEs) and local governmental bodies, as enthusiasm in the broader tech community remains subdued.

1. Investment in NVIDIA’s Ecosystem

Many of China’s technological giants have heavily invested in the NVIDIA CUDA framework. Transitioning away from this established ecosystem would require considerable time and resources. Companies are reportedly looking for Huawei to adjust its offerings to align with their existing platforms, rather than the reverse.

2. Competitive Landscape

Moreover, China’s largest tech firms, which happen to be Huawei’s competitors, are hesitant to fully commit to its products. This apprehension is fueled by their competitive interests rather than a desire for collaborative innovations.

3. Reliability Concerns

Reliability is another pressing issue, as Huawei’s Ascend 910C chips have been reported to experience overheating problems. Such concerns affect their reputation within the tech industry and lead to hesitance in adoption.

4. Existing NVIDIA Inventory

Additionally, many leading tech companies have stockpiled a significant number of NVIDIA GPUs over the years. This existing inventory reduces the incentive for these firms to switch to Huawei’s products, especially considering the financial implications of such a shift.

5. Regulatory Challenges

The situation has been further complicated by US export restrictions. In May, the US Department of Commerce issued guidance declaring Huawei’s chips as “toxic.”Any entities using these chips without prior authorization risk violation of US export controls, a situation that particularly impacts Chinese companies with international operations.

Technical Comparisons and Developments

In a prior analysis, we noted that Huawei’s Ascend 910C combines two earlier 910B chips to deliver around 800 TFLOP/s of computing power at FP16, along with a memory bandwidth reaching 3.2 TB/s. Thus, it is positioned closely alongside NVIDIA’s H100 GPU in performance.

To further contend with NVIDIA’s supercomputing offerings, Huawei has introduced the CloudMatrix 384, integrating up to 384 Ascend chips. While this product matches NVIDIA’s performance capabilities, it currently lacks robust support for memory-efficient calculation formats such as FP8, though Huawei has created a tool intended for artificial compatibility, which remains less than ideal.

NVIDIA’s Market Position

Despite challenges from competitors, NVIDIA continues to thrive. According to a recent report from UBS, the company confirmed visibility into “tens of gigawatts”of AI infrastructure projects. UBS conducted a theoretical evaluation that estimated NVIDIA could realize between $400 billion and $500 billion annually for its AI data center segment, assuming a 20GW infrastructure pipeline and a realization timeline of 2 to 3 years.

The current market landscape illustrates the complexities and dynamics of the AI chip industry, highlighting the formidable challenges Huawei faces in establishing its products within a heavily saturated market.

Source & Images