There have been very few AI conversations lately that have not included DeepSeek. What’s all the hype about?
DeepSeek R1 model, which is on par or better than industry-leading models in the U.S. at a fraction of their cost, threatens to disrupt the multibillion-dollar approach to AI development.
Less than a month after its launch, the groundbreaking Chinese AI model caused a bloodbath in AI tech stocks. For example, Nvidia stock (NVDA) lost 17%, wiping almost $600 billion in market value – a record single-day drop for a U.S. stock.
However, the cost-efficiency of Chinese AI models isn’t the only concern.
Digging deeper, some experts fear that the advancements in Chinese AI models could threaten U.S. national security. Are these safety concerns real, or just an attempt to shut down a competitor?
Key Takeaways
- DeepSeek R1 is a low-cost, high-performance AI model that competes with U.S. industry leaders like OpenAI.
- The model’s efficiency comes from reinforcement learning and Mixture of Experts (MoE) architecture, reducing computing power and memory usage.
- Privacy and security concerns have led to bans in multiple countries, including Italy, the Netherlands, Taiwan, and U.S. government agencies.
- Experts are divided: some see DeepSeek as a threat to U.S. dominance, while others view it as a win for open-source AI development.
What Is DeepSeek?
DeepSeek is a Chinese artificial intelligence company that develops low-cost, open-source, advanced large language models (LLMs) and reasoning systems. It specializes in high-performance AI models like DeepSeek R1.
DeepSeek successfully competes with the leading AI models, challenging a unipolar AI development market dominated by U.S. tech giants.
DeepSeek Cost: Is It Truly Much Cheaper than U.S. Models?
How Much Does DeepSeek Cost?
Released in January 2025, DeepSeek is 100x time cheaper for input and 30x cheaper for output than OpenAI’s o1 model. However, both models show comparable performance results.
Parameter | DeepSeek R1 | OpenAI o1 |
---|---|---|
Context length | 64K | 200K |
Input price (1M tokens) | $0.14 | $15.00 |
Output price (1M tokens) | $2.19 | $60.00 |
So, why is the Chinese model much cheaper?
According to DeepSeek’s research paper, the company spent under $6 million on computing power per training the model, below the costs for popular AIs like ChatGPT or Google’s Gemini.
In their research paper, DeepSeek’s engineers said they used about 2,000 Nvidia H800 chips, which are less advanced than the latest AI chips, to train its model. In contrast, the big U.S. firms use around 16,000 chips or more.
DeepSeek’s R1 model uses an innovative training approach that emphasizes reinforcement learning (RL) over traditional supervised fine-tuning. This approach is also more cost-efficient.
Like OpenAI’s o1, R1 is a reasoning model. This means it produces responses incrementally, simulating a process resembling human reasoning. It requires less memory than its rivals, decreasing the cost of performing tasks.
DeepSeek R1’s cost-effectiveness and efficiency are also largely driven by its Mixture of Experts (MoE) architecture. It has 671 billion parameters organized into multiple expert networks; however, only 37 billion are used at a time, ensuring the model only uses what’s needed for the task.
DeepSeek Doesn’t ‘Do for $6M What Cost U.S. AI Companies Billions’
According to Dario Amodei, CEO of Anthropic, the statement about DeepSeek “doing for $6M what cost U.S. AI companies billions” is largely exaggerated.
To make it more accurate, he suggested a fairer statement would be:
“DeepSeek produced a model close to the performance of U.S. models 7-10 months older, for a good deal less cost (but not anywhere near the ratios people have suggested).”
Amodei went further, suggesting that DeepSeek’s cost aligns with the normal historical trend of the cost curve decreases. He said:
“DeepSeek-V3 is not a unique breakthrough or something that fundamentally changes the economics of LLM’s – it’s an expected point on an ongoing cost reduction curve. What’s different this time is that the company that was first to demonstrate the expected cost reductions was Chinese.”
Still, Amodei admitted that DeepSeek achieved strong performance results due to “some genuine and impressive innovations, mostly focused on engineering efficiency.”
He particularly highlighted “innovative improvements in the management of an aspect called the ‘Key-Value cache,’ and the method called Mixture of Experts.
Deepseek Privacy Concerns, Bans & Millions of Downloads
DeepSeek rapidly surged to the No. 1 most downloaded app on Apple’s App Store and Google Play Store in the U.S. last week.
However, while millions of people entrust their personal search inquiries to a new AI chatbot, many U.S. entities and international governments banned DeepSeek AI over privacy issues:
- Italy’s Data Protection Authority (DPA) launched an investigation into DeepSeek’s data collection practices and compliance with the GDPR.
- The Netherlands’ privacy watchdog AP is also investigating DeepSeek’s data collection practices and urging Dutch users to exercise caution when using the company’s software.
- Taiwan’s Ministry of Digital Affairs said DeepSeek “endangers national information security” and banned government agencies from using it.
- No wonder that, in the face of US-China tech rivalry and mistrust, the U.S. Congress, Pentagon, U.S. Navy, and NASA reportedly banned employees from using DeepSeek AI.
So, what’s wrong with DeepSeek’s privacy policy?
According to the latest report by AI security researchers from Robust Intelligence, now a part of Cisco, the results of investigating DeepSeek’s vulnerabilities were alarming. The report stated:
“DeepSeek R1 exhibited a 100% attack success rate, meaning it failed to block a single harmful prompt. This contrasts starkly with other leading models, which demonstrated at least partial resistance.
“Our findings suggest that DeepSeek’s claimed cost-efficient training methods, including reinforcement learning, chain-of-thought self-evaluation, and distillation, may have compromised its safety mechanisms. Compared to other frontier models, DeepSeek R1 lacks robust guardrails, making it highly susceptible to algorithmic jailbreaking and potential misuse.”
John Meah, InfoSec Author & Cybersecurity Consultant, highlighted a broader concern of global misuse. He told Techopedia:
“China’s strict surveillance, powered by advanced tools like DeepSeek, raises serious concerns about privacy and security. This goes beyond the usual worries tied to AI chatbots — it brings fears of misuse and global instability. While innovative technologies improve our daily lives in many ways, we must ask: at what cost?
“Nations need to tackle how these advancements impact data privacy while finding a balance between progress and responsibility. Protecting individual freedoms is essential in today’s connected world – it cannot be overlooked or compromised.”
A Win for Open-Source AI Models or a Battle in the U.S.-China Dominance War?
Besides the obvious cybersecurity concern that DeepSeek can provide a vast amount of data that bad actors in China could harvest and use for malicious purposes (among them, DeepSeek can be obliged to share data with China’s intelligence agencies), there is a deeper one – the intensified competition between nations.
Antropic’s Amodei sees a bipolar world where both the U.S. and China have equally powerful AI modes as a threat. He said:
“It seems likely that China could direct more talent, capital, and focus to military applications of the technology. Combined with its large industrial base and military-strategic advantages, this could help China take a commanding lead on the global stage, not just for AI but for everything.”
Therefore, Amodei advocates for more export controls, banning Chinese developers from getting powerful AI chips, to keep the U.S. in the leading position.
“I don’t see DeepSeek themselves as adversaries, and the point isn’t to target them in particular…But they’re beholden to an authoritarian government that has committed human rights violations, has behaved aggressively on the world stage, and will be far more unfettered in these actions if they’re able to match the U.S. in AI.
“Export controls are one of our most powerful tools for preventing this.”
However, it’s a very controversial statement regarding the U.S. military history. Is the U.S. AI lead a virtue in itself?
Niels Rogge, Machine Learning Engineer at ML6 & Hugging Face, completely disagrees. He said in a LinkedIn post:
“We need to view DeepSeek AI’s release as a win for open-source over proprietary models. Not as a war between U.S. and China. Don’t believe his [Amodei’s] claims about “AI safety” – it’s arrogance in disguise.”
According to Rogge, the monopoly on owning the most powerful AI models is the “exact opposite of what Hugging Face stands for – building AI technology in a collaborative way.”
He believes that open-sourcing work through code, paper, and model releases allows people to build on each other’s work to improve, innovate, and accelerate.
Rogge raises the opposite concern about AI decentralization and the need to avoid the situation when the most powerful AI models in the world are in the hands of just a few companies like OpenAI and Anthropic. He said:
“By making everyone have access to the same technology, you avoid having everyone access the same model, to which the provider can add their own biases on how to view the world.”
The Bottom Line
Despite the DeepSeek controversy, which shakes the existing AI development market, it could be perceived as a step towards AI democratization – the more players in the game, the more innovation, faster progress, and more affordable tools we have.
Dr.Kiran Nair, Associate Professor and Head of Internationalization and Partnerships at Abu Dhabi University, said:
“For businesses, investors, and policymakers, DeepSeek’s rise is not just about AI – it’s a strategic shift that will redefine the future of technology and global competition. The question is no longer who has the best AI but who can build it smarter, cheaper, and faster.”
FAQs
Who owns DeepSeek AI?
Is it safe to use DeepSeek AI?
Does DeepSeek collect your data?
Is DeepSeek more sustainable?
What does DeepSeek mean for AI?
Can you invest in DeepSeek AI stock?
References
- DeepSeek (DeepSeek)
- Chart: The Extreme Cost of Training AI Models (Statista)
- DeepSeek-V3 Technical Report (Arxiv)
- Dario Amodei — On DeepSeek and Export Controls (Dario Amodei)
- Italian Data Protection Authority Investigates DeepSeek AI’s Data (NatLawReview)
- Dutch privacy watchdog warns about uploading info to DeepSeek – DutchNews.nl (DutchNews)
- Government agencies banned from using DeepSeek, digital ministry says (Taipei Times)
- Evaluating Security Risk in DeepSeek (Cisco Blogs)
- Dr.Kiran Nair on LinkedIn: DeepSeek: China’s Game-Changing AI and the Future of Global AI Competition (LinkedIn)