DeepSeek R1 Utilizing Huawei’s Ascend 910C AI Chips Highlights China’s Advancing AI Capabilities

Over the past months, there’s been a significant shift in China’s AI landscape, driven by DeepSeek. Their AI model now runs inference tasks using Huawei’s latest Ascend 910C chips, a testament to the tech industry’s rapid evolution.

DeepSeek has stirred up a lot of discussion recently, especially after their R1 LLM model caused quite a stir in the global markets, contributing to a $2 trillion drop in the US stock market as it rippled through investor sentiment. There’s been some controversy surrounding the actual computing power DeepSeek leveraged to develop their R1 model. Yet, Huawei seems to have played a crucial role, with reports indicating that DeepSeek’s R1 is utilizing the Ascend 910C chips for inference operations.

To put it in perspective, on social media, Alexander Doria pointed out this noteworthy development: “While DeepSeek initially trained using Nvidia’s H800 chips, they’re now running inference on Huawei’s newly minted 910C chips,” he shared, signaling a strategic pivot towards domestic technology.

For those not in the know, Huawei’s Ascend 910C AI chip is positioned as a formidable competitor to NVIDIA’s Hopper H100 AI accelerators. While there are still some uncertainties about the specifics of Huawei’s chip, the plan for mass production is already on the horizon for early 2025, drawing significant interest from major Chinese AI players like ByteDance and Tencent. As these chips provide a viable alternative to NVIDIA’s offerings in China, they are rapidly capturing market share and seem likely to gain even more ground in the future.

However, when it comes to the resources behind DeepSeek’s R1 AI, details remain a bit murky, with some media reports perpetuating misconceptions. Notably, the $5 million figure often cited isn’t for the entire training process but solely covers running the final model. Additionally, the suggestion that DeepSeek has access to over 50,000 NVIDIA H100 units implies that their resource base is in the same league as other leading AI efforts.

It’s interesting that Huawei’s chips are being used for inference, especially considering their plentiful availability to local companies and their competitive pricing compared to NVIDIA’s stripped-down models or illicit market options. Meanwhile, Huawei is rumored to be in the process of developing the next wave of Ascend AI chips, which are anticipated to give a run for the money to NVIDIA’s forthcoming Blackwell AI products, setting the stage for intensified global rivalry in the field.