Early 2025 has seen the world stride towards the tech version of an Iron Curtain division, with two technological ecosystems battling it out. As was the case during the Cold War, a global AI race can create a competitive landscape where different parts are dominated by varying AI standards. Brics+ countries may, for example, support the Chinese-made technologies, while the collective West may be using the US-made counterparts.
The Great AI Divide has been recently precipitated by the Americans responding to the launch of the DeepSeek R1 model created by a small Chinese-based firm of around 200 employees whose budget is less than that of a single team at Meta. Crucially, what they did was break the industry-wide rules set by the US. Hence the West’s chaotic reaction simultaneously punctuated by outrage and admiration. The Chinese model boasts reasoning faculties and a productivity that is well on par with OpenAI o1, the industry’s benchmark. At the core of the production process were several key differences that eventually led to the market crash.
First, DeepSeek R1 is much more efficient as it was pre-trained using fewer calculation steps and fewer funds too. This is an important achievement. A widely held concept used to view the training of large language models as a luxury. OpenAI, Anthropic and other tech giants spent hundreds of millions of dollars doing it. The process required huge data centres equipped with expensive GPUs, and that field was monopolised by Nvidia. To put it into perspective, it is like a factory requiring a separate power station to be able to operate. The shocking turn of events was that DeepSeek said, ‘You know what? It will only take us a few million bucks and 2,000 GPUs as opposed to 100,000.’
Secondly, DeepSeek R1 is an open-source code, making the results fully accessible to the public. Everyone can download the model to their PC using Ollama or LM Studio and run it offline. The code is freely available. That is no-nonsense, honest engineering. The creators are not hiding anything from you. The model is MIT-licenced, which allows for the unlimited use of intelligible weights and results by researchers and developers alike. That gives the model an edge over its biggest rival created by OpenAI, which closed its code and banned access to its weights while only providing the API. Plus, this software is free and currently has no limits.
Thirdly, DeepSeek is a user-friendly app, easy to download to your smartphone and to figure out. Worldwide, it has taken the lead from ChatGPT by total downloads. This fact briefly overwhelmed the app, but they have already made the fixes.
Finally, DeepSeek has revolutionised the large model training principle by ditching the concept of a jack-of-all-trades juggernaut as a daft idea. Why would you want to activate 600 billion all at once while each task requires just a narrow subject-matter expert? They ended up with an expert system with only the necessary parameters being active, roughly 30 billion of those, which economically is a substantial improvement.
The AI bubble is about to go bust. Everyone will start readjusting to a new reality, which is going to usher in new rules. DeepSeek has called into question the enormous investments made by US businesses to help create new AI solutions. The Chinese researchers have now turned the tables on their Western counterparts, spurring a catch-me-if-you-can race. Mark Zuckerberg has already put together emergency response task forces where engineers are supposed to figure out how a smaller Chinese firm could roll out a disruptive AI technology. They will be cutting the costs and drawing on DeepSeek’s expertise. But the opinions differ on this one. For instance, Yann LeCun, chief AI scientist at Meta, believes a model’s quality can only be ensured by a tech company’s economic sustainability. In other words, for all of its technological competitiveness, DeepSeek’s capacity to reshape the current market will hinge upon its monetisation prospects. But this is debatable since the leading companies have been in the red for quite a while. OpenAI’s losing streak spans 10 years.
The Americans are now claiming that China has stolen their technologies. However, it is hard to tell so far. According to OpenAI’s official statement, Chinese startups are trying to use the US businesses’ R&D in a bid to upgrade their AI models. One training method is called distillation where a model is being trained based on the existing ones. This is a way for smaller models to be trained on larger ones. The White House alleges that the Chinese firm may have been using the American developments. The new administration refers to it as a threat to national security.
By the war, Donald Trump’s response has so far been remarkably even-keeled. According to him, if the Chinese can train their models more efficiently, so can the American scientists, ‘the world’s best’. The fun part here is that DeepSeek R1 has already been asked about its training algorithms. The AI assistant claims it is based on OpenAI’s GPT-4. If this is the case, it means that (1) the training methodology is going to evolve and (2) there will be certain ‘red lines’ set to limit the use of foreign AI technology. This is the policy initiated under the Joe Biden administration. Back then, the tech giants mounted a severe backlash, with Nvidia forecasting a future blow to the market. But now the odds are that they are going to reconsider their stance.
We are undoubtedly watching a global AI race getting underway. It may have been set off by Donald Trump’s Stargate AI project that involves $500bn in investment money, which is comparable to the funds allocated to the Moon exploration programme and three times the aggregate budget of the US venture-capital industry. But DeepSeek has already fought back. Around the same time, OpenAI debuted Operator, its first-ever agent capable of making restaurant reservations, booking tickets and ordering food. The Chinese-based ByteDance responded with an agent of their making, UI-TARS. Alibaba launched the Qwen2.5-1M model supporting a context of 1 million tokens.
This is eight times ChatGPT’s capabilities. Meta, in turn, stormed into it with an ambitious project of building a data centre the size of Manhattan, with a capacity on par with that of a nuclear power station. In the end, China announced the launch of its own Stargate, with the government vowing to invest $137bn in AI development. True, the US investments will surpass it threefold, but the thing is, the US-based OpenAI is the sole company using private funds, whereas the Chinese government will subsidise five vendors, including DeepSeek.
The latest update is that OpenAI has released their free o3-mini model, a move apparently spurred by the newly competitive market. Mind you, these launches have occurred within a span of several days. The fast-paced AI race has definitely kicked off.