In recent developments, DeepSeek’s AI model has captured attention by leveraging Huawei’s latest Ascend 910C chips for inference tasks, highlighting the rapid growth of China’s AI sector in recent months.
DeepSeek, a prominent player in China’s AI scene, has shaken the global AI landscape. Announcing their R1 LLM model, they triggered a $2 trillion dip in the US stock market, causing quite a stir among investors. Despite debates surrounding the computing prowess powering DeepSeek’s R1 model, it’s clear Huawei has played a significant role. According to @dorialexander on social media, DeepSeek R1 leverages the Ascend 910C chips for inference, adding a new layer to the unfolding drama.
For those unacquainted, Huawei’s Ascend 910C AI chip is considered a formidable opponent to NVIDIA’s Hopper H100 AI accelerators. Although details about Huawei’s chips remain somewhat elusive, there’s talk of mass production kicking off in early 2025, with coveted attention from major Chinese tech powerhouses like ByteDance and Tencent. These chips are well-regarded as top-tier alternatives to NVIDIA’s offerings in China and have already grabbed a healthy portion of the market, suggesting a surge in their popularity soon.
While specifics about the computing resources used for DeepSeek’s R1 model remain unclear, there seems to be some confusion in media reports. The much-discussed “$5 million” is not the overall training cost but rather the cost of operating the final model. Moreover, claims suggest DeepSeek has access to more than 50,000 of NVIDIA’s H100s, indicating substantial resources akin to those used by other leading AI models.
Choosing Huawei’s chips for inference is intriguing. They are plentiful for local companies, and pricing is competitive compared to NVIDIA’s more limited variants or even unauthorized accelerators. Meanwhile, Huawei is reportedly advancing its next generation of Ascend AI chips, designed to rival NVIDIA’s Blackwell AI products, which promises to intensify global competition in the AI chip market.