【Monthly Market Strategy with Julius Bär】With Korean semiconductor stocks down nearly 50% from their June peak, has the AI correction finally run its course? Where could the next leg of the AI rally come from, and how are China’s cheaper AI models — including DeepSeek’s rumoured new release — reshaping the competitive landscape? 韩国半导体股自6月高位以来已回落近50%，AI板块回调是否已接近尾声？下一轮行情将由哪些领域推动？与此同时，低成本AI模型在中国持续发展，又将为全球AI的竞争格局带来什么变化？ 除AI主题外，市场亦聚焦价值股回归、中国经济增长预期、香港交易所推出五年期中国国债期货，以及黄金和白银强劲反弹等重要议题。 本期《市场透视》播客于2026年8月11日录制 ***2026.08.11录制。英文版并非中文版的英文翻译，有更多的细节讨论*** 节目播客链接：

DeepSeek just open-sourced DeepGEMM for Huawei Ascend. Same API. BF16, FP8 + FP4 kernels. The wild number: 99.8% of the Ascend 950’s theoretical compute limit. This isn’t another model release. They’re going after the AI infrastructure stack now. GitHub:

DeepSeek: "...open-sourcing the infrastructure components for Huawei's Ascend computing platform, including the TileLang high-level language compiler, computing libraries, and distributed communication libraries. All components correspond one-to-one with the previously open-source components for the NVIDIA platform. TileLang has already been validated on NVIDIA platforms and currently powers the majority of the operators used in training DeepSeek’s state-of-the-art V4 series models." Qwen says: Currently, DeepSeek uses a hybrid training strategy: NVIDIA for the massive 33T token pre-training phase, and Huawei Ascend for post-training (RLHF/SFT) and inference. However, with DeepSeek recently ordering 160,000 next-generation Ascend 950DT/960 chips, the company is building the infrastructure to move pre-training to Ascend by 2027, once the new chips support the necessary low-precision (FP4) training formats.

DeepSeek says it partnered with Huawei to build programming tools for Ascend AI chips and is open-sourcing infrastructure including TileLang. It also said it advanced a 128-chip “super node” system based on the Ascend 950.

DEEPSEEK REVENUE DOUBLES AS IT FINALIZES $75 BILLION VALUATION DeepSeek’s annualized revenue has reached ~$1B, up from less than $500M just a few months ago, as the company works to raise $7.5B at a $75B valuation by the end of October. The company is also preparing for a Shanghai IPO. DeepSeek raised API prices by 2.3x-4.5x last month but reportedly saw no decline in its customer base. Its API business generated an 82.9% gross margin in the first seven months of the year. More than 70% of DeepSeek’s compute is still being allocated to model training, with less than 30% used for inference. Source: The Information