DeepSeek

DeepSeek — 중국 스타트업이 AI 판을 흔든 사건

2025년 1월에 무슨 일이 있었냐면, 중국의 작은 AI 회사가 모델 하나 공개했는데 엔비디아 주가가 하루에 17% 폭락했음. 이게 'DeepSeek 쇼크'임. 도대체 뭔데?

DeepSeek이 뭐야?

중국 항저우에 있는 AI 스타트업. 직원 200명밖에 안 됨. 근데 ChatGPT 만든 OpenAI가 수조 원 쏟아붓는 걸 600만 달러(약 85억 원)로 비슷한 성능을 냈다고 해서 난리남.

왜 이렇게 충격이었냐?

AI 만들려면 비싼 엔비디아 GPU 엄청 써야 하는 줄 알았는데, DeepSeek이 미국이 수출 규제해서 못 쓰는 고성능 칩 없이도 비슷한 성능을 낸 거임. 그것도 오픈소스로 공개해서 누구나 무료로 쓸 수 있게 했음.

어떤 기술이 핵심임?

• MoE 구조: 모든 기능을 한꺼번에 쓰지 않고, 필요한 부분만 켜서 씀. 전기 절약처럼 계산 절약
• 강화학습: 정답 없이 스스로 더 나은 답 찾는 훈련
• 오픈소스: MIT 라이선스로 공개 → 전 세계 개발자들이 무료로 씀

논란 있음

• 중국 회사라서 개인정보 중국 서버로 갈 수 있다는 우려
• 천안문·대만 같은 민감한 주제는 대답을 안 함 (검열)
• 미국·한국 정부기관에서 사용 금지
• 600만 달러 비용이 전체 비용 아니라는 의혹

결론

DeepSeek은 "돈 많이 써야 최고 AI 만든다"는 공식을 깨버렸음. AI 효율 혁신의 시작점이 됐고, 미중 AI 패권 경쟁을 더 치열하게 만들었음. 오픈소스라 지금도 전 세계에서 쓰고 있음.

DeepSeek — 중국에서 만든 깜짝 AI

DeepSeek이 뭔가요?

DeepSeek(딥시크)은 중국에서 만든 AI예요. AI는 컴퓨터가 스스로 생각하는 기술인데, DeepSeek은 2025년에 아주 똑똑한 AI를 세상에 내놓아서 깜짝 놀라게 했어요.

왜 특별한가요?

보통 아주 똑똑한 AI를 만들려면 엄청난 돈이 필요해요. 그런데 DeepSeek은 다른 회사들보다 훨씬 적은 돈으로 비슷하게 똑똑한 AI를 만들었어요. 마치 비싼 재료 없이도 맛있는 요리를 만든 것처럼요!

어떻게 가능했나요?

DeepSeek 연구원들이 새로운 방법을 발명했어요. 필요한 부분만 골라서 쓰는 방법인데, 마치 형광등 전체를 켜는 대신 필요한 부분만 켜는 것처럼 에너지를 아꼈어요. 그래서 더 적은 자원으로 더 좋은 결과를 냈어요.

무료로 공개했어요

DeepSeek은 자신들이 만든 AI를 누구나 쓸 수 있게 무료로 공개했어요. 그래서 전 세계의 많은 사람들이 DeepSeek을 사용하거나 연구할 수 있게 됐어요.

더 알아보기

DeepSeek 이후로 AI를 더 효율적으로 만드는 방법에 대한 연구가 더 활발해졌어요. 적은 자원으로도 좋은 AI를 만들 수 있다는 걸 보여줬으니까요!

DeepSeek (딥시크)

In January 2025, the U.S. NASDAQ market experienced a seismic shock when NVIDIA's stock plummeted by 17% in a single day, wiping out approximately $590 billion in market capitalization. The catalyst was a groundbreaking AI model unveiled by China's relatively unknown startup, DeepSeek (딥시크). This model reportedly achieved performance rivaling that of OpenAI and Google's heavily funded AI endeavors at a fraction of the cost—just $60 million (about $85 million) in training expenses. This event became known as the "DeepSeek Shock."

Company Overview

Founded in Hangzhou, China, in 2023, DeepSeek is an AI research startup backed by High-Flyer, a quantitative hedge fund established by Liang Wenpeng (량원풍). Initially focused on internal AI research for investment purposes, DeepSeek evolved into an independent AI research entity with a team of around 200 employees, significantly impacting the global AI landscape despite its modest size.

R1 Shock: The Mechanics Behind It

DeepSeek-R1, unveiled in January 2025, demonstrated inference capabilities comparable to OpenAI's o1 model but with training costs reduced to a fraction—around one-fortieth of the original investment. This breakthrough stemmed from three key strategies:

1. Leveraging Export Restrictions: By developing efficient algorithms capable of learning effectively even with less powerful H800 chips, circumventing the U.S. ban on exporting advanced H100 chips to China spurred innovation under constraint.

2. Mixture of Experts (MoE) Architecture: This approach selectively activates only necessary expert modules rather than utilizing all parameters simultaneously, dramatically enhancing computational efficiency.

3. Reinforcement Learning for Inference Training: DeepSeek refined its inference capabilities through self-generated reward signals without relying on extensive labeled data, effectively disclosing techniques previously kept secret by OpenAI through open-access publications.

4. Open-Source Release: By releasing major models like R1, V2, and V3 under an MIT license, DeepSeek enabled global developers to freely utilize, modify, and distribute these models, fueling rapid global adoption.

Model Lineage

DeepSeek-V2 (May 2024): Enhanced efficiency with MoE architecture, significantly lowering cost relative to performance gains within China.

DeepSeek-V3 (December 2024): Featuring 67.1 billion parameters, this large model achieved performance comparable to GPT-4 despite requiring substantially less GPU training time.

DeepSeek-R1 (January 2025): This model sent shockwaves globally, matching OpenAI's o1 in inference capabilities.

DeepSeek-V4 (April 2026): Previewed in April 2026, aiming for a valuation exceeding $20 billion, with claims of operational costs being one-sixth those of competitors.

Geopolitical Implications

DeepSeek transcends being merely an AI company; it stands at the forefront of the U.S.-China technological rivalry.

The U.S. imposed stringent restrictions on exporting advanced AI chips to China to curb its AI advancement. However, DeepSeek ingeniously developed highly efficient algorithms using less sophisticated hardware, challenging the efficacy of these export controls while demonstrating China's capability to overcome hardware disadvantages through software innovation.

U.S. congressional bodies and the Department of Defense have banned DeepSeek applications on government devices, and South Korea has similarly restricted certain public sector usage due to AI security concerns.

Controversies

Data Security Concerns: Allegations that DeepSeek's applications might transmit user data to Chinese servers raise significant privacy issues, compounded by China's stringent data access laws.

Training Cost Transparency: Questions arise regarding whether the reported $60 million training cost represents the entirety of development expenses or merely the final fine-tuning phase, suggesting a potentially higher overall investment.

Chip Smuggling Allegations: There are claims that DeepSeek circumvented U.S. export restrictions by sourcing H100 chips through unofficial channels.

Content Moderation Issues: DeepSeek has been scrutinized for potentially censoring or refusing to address sensitive topics such as the Tiananmen Square protests, Tibet, and Taiwan.

Impact on the Global AI Industry

Following the DeepSeek Shock, the AI industry's focus shifted towards efficiency rather than mere scale. There is now a greater emphasis on developing algorithms that achieve superior performance with fewer resources, prompting companies like Meta and Google to prioritize efficiency improvements in their models.

Future Outlook

Continuing its rapid release schedule into 2026, DeepSeek aims to maintain its technological lead through frequent new model launches and robust academic contributions. With fundraising targets exceeding $20 billion, the company plans to balance open-source accessibility with commercial solutions to monetize its innovations effectively.

Related Topics

OpenAI | GPT-4 | Google AI | Google Gemini | Microsoft Azure AI | NVIDIA | AI Chip Export Restrictions | MoE Architecture | Reinforcement Learning | Open Source AI | AI Dominance Race | Liang Wenpeng | High-Flyer Fund

English version not yet available.

English version not yet available.