DeepSeek: What is it? and what impact is it currently having?

DeepSeek, the Chinese firm specializing in artificial intelligence, is revolutionizing the sector with its open-source, low-cost language models, competing directly with large American companies.

It has long been believed that developing advanced language models required enormous technical and financial resources. In fact, that was one of the reasons why the US announced the Stargate Project $500 billion, promoted by President Donald Trump.

However, DeepSeek has debunked that idea. On January 20, 2025, the company released its R1 language model at a fraction of the cost of its competitors. It also published it under an open source license, allowing it to be used for free.

Within days, the DeepSeek AI Assistant mobile app, which offers an R1-based chatbot interface, rose to number one in the Apple App Store download rankings, surpassing even OpenAI's ChatGPT app. Its meteoric rise triggered a stock market sell-off on January 27, 2025, raising doubts about the future of major players in the sector such as Nvidia, Microsoft, Meta, Oracle, and Broadcom.

What is DeepSeek?

DeepSeek is a AI development company Headquartered in Hangzhou, China, it was founded by Liang Wenfeng, a graduate of Zhejiang University and co-founder of the investment fund High-Flyer, which currently owns DeepSeek. Although the total funding and valuation are unknown, the company operates as an independent research lab focused on open-source language models.

Their first model was launched in November 2023, but it was with the DeepSeek-R1 model in January 2025 that they achieved international fame. They offer their services through a web interface, mobile app, and API.

DeepSeek vs OpenAI

DeepSeek presents itself as an alternative to OpenAI, known for its popular ChatGPT. While OpenAI uses proprietary technologies, DeepSeek focuses on transparency and open source, which represents a direct threat to the business model of many American companies.

	OpenAI	DeepSeek
Founding year	2015	2013
Headquarters	San Francisco, California	Hangzhou, China
Development focus	Broad AI capabilities	Efficient, open source models
Key models	GPT-4o, o1	DeepSeek-V3, DeepSeek-R1
Specialized models	Dall-E (image generation), Whisper (speech recognition)	DeepSeek Coder (coding), Janus Pro (vision model)
API pricing (per million tokens)	o1: $15 (input), $60 (output)	DeepSeek-R1: $0.55 (input), $2.19 (output)
Open source policy	Limited	Mostly open source
Training approach	Supervised and instruction-based fine-tuning	Reinforcement learning
Development cost	Hundreds of millions of dollars for o1 (estimated)	Less than $6 million for DeepSeek-R1, according to the company

DeepSeek Technical Innovations

The R1 model was trained in less time, with fewer resources, and at a lower cost. Some of the keys to its success are:

• Large-scale reinforcement learning focused on reasoning tasks.
• Reward engineering with rule-based systems that are more efficient than traditional neural models.
• Knowledge distillation, achieving compact models with 1.5 billion parameters.
• Emergent behavior network, discovering complex patterns without explicit programming.

Language models developed by DeepSeek

• DeepSeek Coder (November 2023): oriented to programming tasks.
• DeepSeek LLM (December 2023): General-purpose model.
• DeepSeek-V2 (May 2024): improved performance and lower training costs.
• DeepSeek-Coder-V2 (July 2024): 236B parameter model for advanced programming.
• DeepSeek-V3 (December 2024): "mixture-of-experts" architecture, 671B parameters.
• DeepSeek-R1 (January 2025): focused on reasoning, it competes with OpenAI's o1 model.
• DeepSeek-R1-0528 (May 2025): Supports JSON prompts, functions, and output. Improves accuracy and reduces hallucinations.
• DeepSeek-R1-0528-Qwen3-8B: A reduced version based on Alibaba’s Qwen3 model.
• Janus-Pro-7B: multimodal model launched in January 2025.

Why is DeepSeek a concern in the US?

The success of DeepSeek has raised concerns in the United States for several reasons:

• Cost disruption: DeepSeek claims to have developed the R1 model for less than $6 million.
• Technical achievement without advanced chips: Despite export restrictions, they have achieved top-level results.
• Threat to the business model: they offer for free what others sell by subscription.
• Geopolitical tension: It represents a direct challenge to US technological dominance.

And in the European Union?

According to his own approach, the European Union has established a pioneering legal model with its Artificial Intelligence Act (AI Act), defining four levels of risk—from low to unacceptable—and mandating security, transparency, and accountability requirements depending on the type of system.

Key regulatory framework

• The AI Act is the first global regulation to categorize AI systems by risk: unacceptable risk (prohibited), high risk (subject to strong controls), limited risk, and minimal risk.
• It includes specific requirements for general-purpose models (LLMs), such as DeepSeek, which must comply with transparency, documentation, and systematic risk assessment obligations.
• It establishes a CE marking-type system for high-risk systems and allows the use of legal sandboxes to encourage innovation within compliance.

Governance and implementation

• The European AI Office, created in February 2024, coordinates the implementation of the AI Act together with the European Artificial Intelligence Board and national offices in each Member State.
• The European approach combines innovation and security, adopting measures such as mandatory technical audits for high-risk systems.

DeepSeek in the European regulatory spotlight

• Several European bodies—France, Italy, Belgium, Germany, Ireland, and Spain—have launched investigations into privacy and GDPR violations related to DeepSeek.
  • The Italian data protection authority blocked its use, considering it non-compliant with regulations.
  • The Berlin commissioner has requested that Apple and Google remove DeepSeek from their stores for illegally transferring user data to China.
• Independent evaluations such as COMPL-AI conclude that DeepSeek models have critical gaps in cybersecurity and bias mitigation, although they perform well in preventing toxic content.

Prohibitions and regulations on DeepSeek

As we've seen, numerous governments and institutions have banned the use of DeepSeek due to privacy, security, and ethical concerns. Some examples include:

• Government agencies from Australia, India, Taiwan, South Korea, and Texas.
• NASA, Congress, and the U.S. Pentagon
• In the EU, DeepSeek could violate regulations such as the GDPR and the AI Act by transferring data to China without adequate safeguards.

Cyberattacks and data breaches

On January 27, 2025, DeepSeek fell victim to a massive cyberattack, possibly a DDoS attack, forcing it to limit new registrations. Shortly after, Wiz Research discovered an exposed internal database containing sensitive information such as API keys, chat histories, and operational data.

Conclusion

What future awaits DeepSeek?

DeepSeek has demonstrated that it is possible to develop advanced language models with fewer resources and open source code. Its disruptive approach, advances in reasoning, and low cost are reshaping the global landscape of artificial intelligence.

For businesses and users interested in AI, DeepSeek is positioned as a powerful, accessible, and highly competitive option compared to the tech giants.