Artificial intelligence (AI) has become the new scenario of a global technological war.
On one side, giants like OpenAI, Google or Meta, backed by billions of dollars and access to cutting-edge resources. On the other, a Chinese AI company, DeepSeek, whose R1 model has emerged as a precision missile: cheaper, more efficient and, in some respects, just as lethal.
While the West invested in pharaonic projects like Stargate – the $500 billion initiative led by Trump and OpenAI – DeepSeek showed that innovation does not always depend on the budget.
With just $5.5 million and a revolutionary approach, this company has shaken Wall Street, plummeted stocks like Nvidia, and questioned the future of ChatGPT.
How did it do it? These are the six reasons why DeepSeek-R1 is not just a rival, but a turning point in the AI race.
Unprecedented cost efficiency: Do more with less
Each training of a GPT model costs tens of millions of dollars. However, DeepSeek-R1, according to its own creators, was trained with $5.5 million, 3-5% of the cost of OpenAI.
The key is in the method. While OpenAI uses Supervised Fine-Tuning (SFT)—a process that requires huge amounts of human-labeled data—DeepSeek opted for Reinforcement Learning (RL).
This approach, based on rewards and automatic feedback, reduced dependence on human supervision and optimized the use of resources.
Furthermore, DeepSeek trained its model with only 50,000 GPUs (10,000 from Nvidia, acquired before the US veto), compared to the 500,000 GPUs used by OpenAI or Google.
By the way, if you don’t know what a GPU is, you might want to read our article on AI chips.
As Wedbush’s Dan Ives noted: “DeepSeek has taken the market by storm by achieving the impossible: rivaling ChatGPT at a fraction of the cost.”
Open-Source: Transparency as a weapon
What is now ChatGPT promised to be “open” in its beginnings, not for nothing is the company called “OpenAI”, but today it is a black box. Users cannot access or run your code locally.
The reality is that DeepSeek-R1 is an open source model, downloadable and modifiable by anyone. This transparency solves two critical problems:
- Privacy: When running R1 on local devices, data never leaves the user’s computer.
- Censorship: Although the China-hosted version applies government filters, developers can remove these restrictions by modifying the code.
For Jay Woods, of Freedom Capital Markets, “DeepSeek’s open-source is a masterstroke. It attracts developers and companies that are wary of sharing data with Big Tech.”
Superior performance in technical tasks
Preliminary testing indicates that DeepSeek-R1 outperforms OpenAI’s o1 model in logical reasoning, coding, and mathematics. According to some users, R1 solves complex equations more accurately and offers more optimized code solutions.
An example cited in a recent VentureBeat article: When asking for an algorithm for sorting data, DeepSeek not only wrote it, but explained its efficiency using “Big O.”
Although both models are tied in general tasks, DeepSeek’s technical advantage is a magnet for sectors such as engineering, research or finance.
Economic accessibility: The price that changes the rules
Accessing OpenAI’s API to use its o1 reasoning model costs $7.50 per million tokens, something like 750,000 words. And ChatGPT Plus subscription costs $20 per month.
DeepSeek’s API, which is fully compatible with OpenAI’s, allows access to R1 from $0.14 per million tokens (53 times cheaper) and is free to use for individuals.
For companies, this difference is abysmal. As Matt Binder explains in a Mashable article: “If two models are equivalent, you will choose the one that costs 95% less.”
DeepSeek not only threatens OpenAI, but the entire AI API economy, where Google (Gemini) and Meta (Llama) also play.
Innovation under restrictions: Ingenuity in the face of sanctions
The US banned the export of advanced chips (such as the Nvidia A100) to China in 2023. However, DeepSeek had already accumulated 10,000 of these GPUs, using them as a base for R1.
Faced with the shortage, DeepSeek engineers devised two strategies:
- Inference-time computing: They only activate the parts of the model necessary for each query, saving energy and power.
- Extreme resource optimization: According to its white paper, R1 achieves 92% energy efficiency compared to ChatGPT’s 78%.
“It’s like they built a Ferrari with scrap parts,” commented Ben Reitzes of Melius.
Geopolitical impact: The new face of the technological Cold War
The rise of DeepSeek is not just technical, but geopolitical. With China pushing its technological self-sufficiency agenda, R1 demonstrates that the country can compete in AI despite sanctions.
For Angelo Zino of CFRA Research, “this questions whether the West has underestimated China. As the US spends trillions on projects like Stargate, DeepSeek shows that innovation doesn’t always require huge checks.
Furthermore, the Chinese model has exposed two Western vulnerabilities:
- Hardware dependence: Nvidia shares fell 16.9% yesterday after the launch of R1, reflecting fear of lower demand for GPUs.
- Overvaluation of technology stocks: The Nasdaq fell 3% in one day, with investors reassessing the AI hype.
There was an earthquake, aftershocks are expected
By way of conclusion, DeepSeek-R1 is not perfect. Its bias towards China, which we discussed in a recent article, and the opacity of its real costs are shadows on its success.
But it has achieved the unthinkable: putting ChatGPT (and other competitors) on the ropes with a fraction of its resources.
As Marc Andreessen wrote in X: “DeepSeek R1 is one of the most amazing advances I have ever seen. And being open-source, it is a gift to the world«.
The question now is whether the West will replicate its methods or double down in the race for AGI (Artificial General Intelligence). The only thing clear is that, on the AI battlefield, DeepSeek has redefined the rules of the game. And this is just the beginning.
This post is also available in: