Memory Solutions for Generative AI Part 5: DeepSeek Moment in Chips
DeepSeek’s Large Language Models (LLM) have gained attention for offering performance similar to ChatGPT but at a significantly lower cost. The immediate market reactions are mixed. Although the missing part might be the cost of data training, such as labelling, and grouping, which is supported by the government, DeepSeek’s achievements in training efficiency and the low cost are clear.
Can DeepSeek Empower China’s Chip Manufacturing?
China’s memory could be already cost competitive: Assuming the Chinese government supports fixed costs, which constitute a large portion of the total cost, China can already achieve a competitive cost compared to peers. For example, if the price per Gb of DRAM in Q1 2024 is $0.34, when the average selling price is less affected the highly priced HBM, and the total cost for South Korean DRAM is about 67% of the price, or $0.23, then China's cost could be as low as $0.20 without the fixed cost. (However, the fixed costs are much higher than in South Korea.)
Efficient software opens market for low-end hardware: China's approach is to compensate for lower performance compared to competitors with volume. Huawei's latest GPU, the Ascend 920, supports HBM2 and HBM2e which are two-year-old standards for peers, instead of the latest HBM3 or HBM3e. However, as DeepSeek has shown, it is expected to be competitive in certain areas of inference.
2025-2026 will be the most critical period for the competitive landscape. The issue is the sustainability of low cost and high efficiency. Among equipment, workforce, and capital, equipment is the most critical bottleneck. However, we are surprised to see China's latest Cryo Etcher from Naura being put into mass production for YMTC. Considering China’s comprehensive efforts in the supply chain for new products like HBM, it seems quite sustainable in the near term.
Access full 4-pager report here.
If you are not a client, reach out to us at info@counterpointresearch.com
Author
MS Hwang
MS Hwang is a research director at Counterpoint, specializing in memory semiconductor research. MS Hwang brings over 30 years of experience from Samsung Electronics and sell-side brokerage research roles including ABN AMRO, Goldman Sachs, Credit Suisse and Samsung Securities.
Related Research
Jan 30, 2025
Jan 16, 2025
Jan 3, 2025
Dec 26, 2024