Explainer: Why DeepSeek AI is a breakthrough for some, a threat for others
Published: 31 Jan. 2025, 19:25
Updated: 31 Jan. 2025, 19:44
-
- LEE JAE-LIM
- [email protected]
-
- PARK EUN-JEE
- [email protected]
Audio report: written by reporters, read by AI
![The Deepseek app is seen in this illustration taken on Jan. 29. [REUTERS/YONHAP]](https://koreajoongangdaily.joins.com/data/photo/2025/01/31/4c8a2fd1-3aa3-4c94-b098-39893eea2c08.jpg)
The Deepseek app is seen in this illustration taken on Jan. 29. [REUTERS/YONHAP]
The release of new AI models developed by Chinese startup DeepSeek has instigated market gyration across the globe and intense interest from every corner.
The open-source R1 model is making headlines due to its startlingly low training costs and capabilities comparable to OpenAI’s reasoning model.
Still, its significance goes much deeper than that, as the emergence has been upsetting a set of established beliefs and industry norms, sharply dividing winners and losers.

Why is DeepSeek’s development so disruptive?
It boils down to the fact that the performance of DeepSeek's reasoning model, R1, is just as good as OpenAI’s o1 while using relatively fewer low-performing processors — if the Chinese startup's claim about its chip deployment holds true. The breakthrough is the key factor that pushed down training costs to a mere $5.576 million compared to the more than $100 million expenditure by OpenAI.
DeepSeek claims that it trained V3, an underlying large language model for R1, on 2.788 million H800 GPU hours, according to its technical paper. Nvidia’s H800 is a modified version of its high-end chips designed to circumvent U.S. export restrictions against China that possesses far lower energy efficiency and a weaker performance than the H100 chip.
The revelation upended the widely-held notion that the higher the chip performance or computational power behind a model, the better it can comprehend and generate sophisticated responses.
Some experts say that it is evidence that reasoning or inference-oriented tasks can be also done through algorithmic breakthroughs rather than resource-intensive training.
How did DeepSeek make it?
The achievement stems from a combination of its techniques at the architectural level that result in significant efficiency gains. Among them is DeepSeekMoE, a language model architecture designed to route queries to the most relevant “experts” and activate only the ones that are necessary. Another important system is called multi-head latent attention (MLA), which enables DeepSeek models to reduce memory usage.
Such an approach could threaten the monopoly-like status of Nvidia and offer more opportunities to smaller chip designers specializing in inference.
Still, the belief that a training price tag of $5.576 million can automatically guarantee a performance comparable to the R1 reasoning model or V3 language model is a misconception given that it excludes development expenditures for the architecture.
“Assuming the rental price of the H800 GPU is $2 per GPU hour, our total training costs amount to only $5.576 million,” the paper said. “Note that the aforementioned costs include only the official training of DeepSeek-V3, excluding the costs associated with prior research and ablation experiments on architectures, algorithms or data.”
How will this affect the Korean industry?
In the wake of DeepSeek's announcement, shares of domestic AI value chains were impacted, including SK hynix, a major memory chip supplier to Nvidia, as well as tech companies such as Naver and Kakao.
Following the debut of the Chinese firm's low-cost, open-source AI model R1 on Jan. 20, the Korean stock market initially remained insulated due to a temporary holiday on Jan. 27, followed by the Lunar New Year break through Jan. 30.
As trading resumed on Friday, SK hynix shares plummeted on the Kospi bourse, closing at 199,200 won ($137.40), a 9.86 percent dip compared to the previous trading session on Jan. 24. Shares of Samsung Electronics, another Korean chipmaker, slipped 2.42 percent to 52,400 won.
Declining share prices reflected investor concerns over whether the traditional scale-up approach to AI development was still necessary. DeepSeek’s ability to develop advanced models quickly and cheaply without relying on high-performance chips challenged existing industry norms.
Tech companies that benefited from the news of DeepSeek’s entry were Naver and Kakao.
Shares of Naver jumped 6.13 percent to close at 216,500 won while Kakao shares spiked 7.27 percent to 38,350 won.
Naver, Korea’s largest portal site, and Kakao, the operator of the nation’s dominant mobile messenger KakaoTalk, are at the forefront of the domestic industry, opting to advance their respective proprietary AI — although progress has been relatively sluggish compared to U.S. Big Tech or Chinese companies due to a lack of hardware at scale as well as investment, infrastructure and computing power.
“In the future, AI may no longer be exclusive to Big Tech,” said Sangsangin Investment & Securities analyst Choi Seung-ho. “The reason DeepSeek has shaken up the market is that it suggests high-performance models may not necessarily require overwhelming computing resources. If the AI technologies presented by DeepSeek prove to be valid, it could provide an opportunity for domestic AI developers to narrow the gap with Big Tech.”
Shares of other AI startups such as Flitto, Maum AI and Vaiv Company surged by double digits — 29.91, 13.15 and 16.54 percent, respectfully — on the Kosdaq bourse on Friday.
![Nvidia CEO Jensen Huang delivers a keynote address at the Consumer Electronics Show (CES) in Las Vegas, Nevada on Jan. 6. [AFP/YONHAP]](https://koreajoongangdaily.joins.com/data/photo/2025/01/31/f1a8b232-5f26-40a8-ac9d-844547abdc5d.jpg)
Nvidia CEO Jensen Huang delivers a keynote address at the Consumer Electronics Show (CES) in Las Vegas, Nevada on Jan. 6. [AFP/YONHAP]
What about chips? Does cheaper AI training signal retracting demand for advanced chips?
The answer remains divided due to opposite claims about whether DeepSeek’s AI capabilities demonstrated by V3 and R1 can be touted as technological innovations that ultimately shake up the global industry and upend the traditional scale-up approach.
One side claims that — albeit recognizing reduced expenses — what DeepSeek released is the natural cycle of technology diffusion, and not a factor that hinders the expansion of AI infrastructure investment.
“Its impact on demand for high bandwidth memory chips remains limited,” said Samsung Securities analyst Lee Jong-wook.
“The memory bandwidth shortage in language models has not been resolved. In fact, the push for model optimization highlights this very limit,” adding that the optimization techniques are “not particularly groundbreaking.”
“Not only Nvidia, but even chips custom-designed for AI inference are seeing a surge in HBM adoption,” Lee commented.
Another side argues that the introduction of DeepSeek models reflects the shifting AI trend of optimizing pretrained models and improving them through reinforcement learning — which, although maybe not as revolutionary as OpenAI’s GPT series, will still have a significant impact on AI hardware dependency, including semiconductors that power the AI models.
“Right now, companies are forced to use high-powered GPUs due to the traditional scale-up approach,” said an industry insider who spoke on condition of anonymity.
“If DeepSeek’s open-source approach allows other companies to replicate its efficiency, then the market will naturally move away from relying on high-performance AI chips and other related hardware. While this trend is positive for the overall AI industry, it poses risks for Big Tech companies that rely on expensive AI hardware, such as Nvidia.”
BY PARK EUN-JEE, LEE JAE-LIM [[email protected]]
with the Korea JoongAng Daily
To write comments, please log in to one of the accounts.
Standards Board Policy (0/250자)