**[WEEKLY BIZ] DeepSeek’s ‘Cost’ Innovation: “Expanding the AI Industry to Foster Innovation”**
The shockwave of the AI chatbot released by the Chinese AI startup DeepSeek continues to resonate. Unlike its counterparts (U.S. AI companies), who typically spend significant sums on development, DeepSeek achieved impressive results with just about one-tenth of the cost, using low-performance semiconductors to strategically focus on key areas. This strategy has surprisingly led DeepSeek to top its class in certain subjects, catching the attention of global big tech companies.
WEEKLY BIZ analyzed how the executives of nine global IT and semiconductor companies, who released their quarterly earnings after the 27th of last month, view DeepSeek. Instead of reacting with ‘tension’ and ‘shock’, these companies generally expressed a positive outlook, recognizing the potential for a wider range of choices and expansion of AI usage with affordable models.
**American or Chinese Cats, It Doesn’t Matter**
American big tech companies are particularly interested in DeepSeek’s value for money. DeepSeek disclosed that its AI development costs (model V3 standard) amounted to $5,576,000 (approximately 8.1 billion KRW). While there is debate about whether pre-research costs are included, the simple comparison shows that DeepSeek’s costs are about one-eighteenth of OpenAI’s GPT-4 and one-tenth of Meta’s Llama 3 development costs. Satya Nadella, CEO of Microsoft, praised DeepSeek’s achievement as comparable to OpenAI’s launch of the o1 model, emphasizing the importance of reducing inference costs to broaden AI application. Sundar Pichai, CEO of Alphabet (Google), similarly highlighted the importance of both developing and efficiently providing AI models, expressing confidence in their own models like Gemini 2.0 Flash compared to DeepSeek’s V3 and R1 models.
Many companies using large language models (LLM) developed by others share the sentiment that it doesn’t matter if the cat is American or Chinese as long as it catches mice. ServiceNow’s CPO Amit Zavery stated that they plan to explore DeepSeek while collaborating with all LLM providers. Even for Microsoft, which has consistently invested in OpenAI, DeepSeek’s models could be a viable option for their own generative AI service, Copilot. Nadella mentioned the possibility of using DeepSeek’s R1 model with Copilot+PC.
**Positive Prospects for Semiconductor Manufacturers**
The semiconductor industry does not view the ‘DeepSeek shock’ solely as a risk factor. DeepSeek employs a ‘Mixture of Experts’ model, activating only the AI necessary for solving specific tasks, resembling a student who focuses on likely test questions. With ‘low-cost, high-efficiency’ becoming a buzzword in the AI industry, there’s anticipation for the development of lightweight, specialized sLLMs (small large-language models) for finance, healthcare, legal, and other sectors. Arvind Krishna, CEO of IBM, noted the necessity of considering simplified language models and appropriate training times for enterprise LLMs.
While there are concerns that the construction of large and heavy AI infrastructure might slow down, reducing AI semiconductor demand, semiconductor manufacturers see this phase as short-lived, creating new opportunities as the AI industry evolves. Kim Jae-jun, VP of Samsung Electronics’ Memory Division, mentioned monitoring industry trends, given that they supply HBM for GPUs used by DeepSeek. ASML CEO Christoph Fokker shared a similar outlook, focusing on positive trends rather than crises.
**Will Open Source Become the Trend?**
DeepSeek is also igniting debates on AI model development methods. Leading players like OpenAI and Google have championed closed models, keeping their AI workings hidden. In contrast, latecomers like DeepSeek and Meta are opting for open-source development, allowing public use of their information. This approach, through a form of ‘collective intelligence’, can fast-track AI model development while reducing development costs. However, it is noted to have drawbacks like lower profitability and potential misuse by criminal entities.
While the open-source approach has its pros and cons, Meta, which employs a similar strategy, seems to support DeepSeek. Meta CEO Mark Zuckerberg remarked during an earnings call that the open-source method is becoming a global standard.
**Concerns Over Chinese AI and Technology Misappropriation Persist**
Reactions to the development of innovative AI models by China, a competitor to the U.S. in AI dominance, vary among companies. Apple and Meta view DeepSeek’s emergence as beneficial for industry growth. Apple CEO Tim Cook welcomed innovation that enhances efficiency, while Zuckerberg highlighted the essence of the tech industry as learning from peers, irrespective of geographical competition.
Nonetheless, concerns remain. Palantir CTO Shyam Sankar pointed out that the real lesson from DeepSeek’s emergence is the ongoing AI arms race between China and the U.S., indicating that some misappropriation of technology likely occurred during DeepSeek’s distillation process. He also mentioned the increased GPU sales in Singapore as an indication of attempts to bypass U.S. sanctions.
**Open Source AI Model**
This approach involves developers publicly sharing information about the algorithms and architecture of their AI models, in contrast to closed models that keep such details confidential. This openness allows other companies and developers to offer improvements, facilitating rapid AI model development with reduced costs.