🧠 Food for thought

Implications, context, and why it matters.

DeepSeek V4 needs unusually large computing resources

• Huawei said its Ascend supernode, built on Ascend 950 AI chips, offers “full support” for DeepSeek V4. That lines up with the model’s size and the hardware it needs 1.
• DeepSeek V4 reportedly has about 1 trillion parameters and is expected to handle text, images, and video 2.
• A 1-trillion-parameter model usually needs a huge memory pool. One deployment guide puts that at about 1 terabyte of video random access memory for bfloat16 weights, with lower needs at 8-bit floating point and quantized precisions 3.
• DeepSeek reportedly spent months rewriting core code and testing it with Huawei plus Cambricon, a Chinese AI chip company, so the model could run on Huawei’s Ascend architecture 1.

Chinese AI companies are leaning harder on local chip suppliers

• DeepSeek’s reported use of Huawei hardware fits a wider pattern among Chinese AI companies dealing with U.S. chip limits 1.
• Alibaba Group, ByteDance, and Tencent have reportedly placed bulk orders for Huawei’s Ascend 950PR chips, adding weight to a homegrown option for AI computing in China 1.
• The shift does not mean Western technology is gone. It does suggest more future models and deployments in China may rely on domestic chips.
• DeepSeek V4 is expected to run on Huawei Ascend chips, though DeepSeek has not confirmed that report 1.