DeepSeek爆火最大受益者是英伟达：特供H20芯片在中国订单大增
发布时间：2026-07-08 15:26:28
英伟达投资者的DeepSeek恐惧症，这下彻底治好了。据报道，六名知情人士表示，由于对DeepSeek低成本人工智能模型的需求激增，中国企业正在增加英伟达H20人工智能芯片的订单。其中的两位知情人士表示，自DeepSeek上个月进入全球公众视野以来，腾讯、阿里巴巴和字节跳动对H20的订单“大幅增加”。H20是英伟达为中国市场定制的人工智能芯片，为遵守美国出口限制而推出。H20基于英伟达Hopper架构，拥有96GB的HBM3内存，提供4.0TB/s的内存带宽。算力方面，H20的FP8性能为296 TFLOPS，FP16性能为148 TFLOPS。与H100相比，其GPU核心数量减少41%，性能降低28%。近日，英伟达CEO黄仁勋也对此前出现的DeepSeek恐慌进行了回应。他表示，DeepSeep-R1推理模型“令人印象深刻”，但未来的人工智能模型仍将依赖英伟达的计算能力，此前市场对英伟达股票的抛售是基于误解。“我认为市场当时对R1的反应是‘哦，我的天哪，人工智能结束了！它从天而降，我们不再需要进行任何计算了！’”黄仁勋表示，“而事实恰恰相反。”“从投资者的角度来看，他们有一种思维模式，他们认为世界就是先进行预训练，然后进行推理，而推理就是向AI提问并立即得到答案。”黄仁勋说，“我不知道这种误解是谁造成的，但显然，这种观念是错误的。”黄仁勋表示，尽管预训练仍然重要，但‘后训练’（post-training）才是“智能最重要的部分”，也是“学习解决问题的关键环节”。教会人工智能模型更好地“推理”是科技行业的下一个扩展前沿，仍然依赖于巨大的计算能力。DeepSeek横空出世后，英伟达股价曾在1月27日暴跌17%。目前，已经修复到1月份水平。
The biggest beneficiary of DeepSeek's popularity is NVIDIA: orders for specially supplied H20 chips have surged in China. NVIDIA investors' DeepSeek phobia has been completely cured. Chinese companies are increasing orders for Nvidia's H20 artificial intelligence chips as demand for DeepSeek's low-cost artificial intelligence models surges, six people familiar with the matter reportedly said. Two of the people familiar with the matter said that orders for the H20 from Tencent, Alibaba and ByteDance have increased significantly since DeepSeek entered the global public eye last month. H20 is an artificial intelligence chip customized by Nvidia for the Chinese market and launched to comply with US export restrictions. The H20 is based on the NVIDIA Hopper architecture and has 96GB of HBM3 memory, providing 4.0TB/s memory bandwidth. In terms of computing power, H20 s FP8 performance is 296 TFLOPS and FP16 performance is 148 TFLOPS. Compared with the H100, the number of GPU cores is reduced by 41% and the performance is reduced by 28%. Recently, Nvidia CEO Huang Renxun also responded to the previous DeepSeek panic. He said that the DeepSeep-R1 inference model is impressive , but future artificial intelligence models will still rely on Nvidia's computing power, and the previous market sell-off of Nvidia stock was based on misunderstandings. I think the market s response to R1 was Oh my gosh, AI is over! It falls from the sky and we don t have to do any calculations anymore! Huang said. And it s exactly the opposite. From an investor s perspective, they have a mindset where they think the world is pre-training and then inference, and inference It s about asking AI questions and getting answers immediately. Huang Renxun said, I don t know who caused this misunderstanding, but obviously, this concept is wrong. Huang Renxun said that although pre-training is still important, post-training is the most important part of intelligence and the key link in learning to solve problems. Teaching AI models to reason better is the next expanding frontier in the tech industry, which still relies on massive computing power. After DeepSeek came out, Nvidia's stock price plummeted 17% on January 27. Currently, it has been restored to January levels
