For over two years, OpenAI, the San Francisco-based artificial intelligence (AI) powerhouse, has led the field with its generative pre-trained language models. The company’s chatbot could craft poems, generate long-form content, debug code, and assist with online searches (albeit with a cutoff date). Its ability to produce seamless, coherent text left users worldwide astounded.
Meanwhile, across the Pacific in Beijing, China began taking steps to challenge America’s dominance in AI. In March 2023, Baidu secured government approval to launch its AI chatbot, Ernie Bot. Marketed as China’s response to ChatGPT, Ernie gained over 30 million user sign-ups within a day of its release.
However, the initial enthusiasm surrounding Ernie gradually faded as the chatbot struggled with politically sensitive topics. When questioned about President Xi Jinping, the Tiananmen Square crackdown, or the treatment of Uyghur Muslims, Ernie sidestepped the queries with a standard response: “Let’s talk about something else.”
As excitement over Ernie diminished due to censorship constraints, experts pointed out the challenges of developing large language models (LLMs) in China’s tightly regulated environment. Former Google CEO Eric Schmidt, speaking at Harvard Kennedy School in October 2023, commented, “China was late to the party. They didn’t enter the LLM AI space early enough.”
Schmidt further noted that China’s AI efforts were hindered by limited access to diverse training data and an unfamiliarity with open-source collaboration. As Chinese tech giants lagged behind, their U.S. counterparts surged ahead, with OpenAI and Microsoft leading advancements in AI-powered reasoning models. OpenAI’s ‘O’ series, for instance, introduced inference-time scaling, optimizing how models handle vast amounts of data in real time.
While China’s major tech firms struggled, an unexpected player emerged in the AI space. High-Flyer, a hedge fund based in Hangzhou, Zhejiang, known for its AI-driven trading strategies, established its own AI lab, DeepSeek, in April 2023. Within a year, DeepSeek released its second-generation AI model, DeepSeek-v2, which performed exceptionally well in benchmark tests while being significantly more cost-effective than other Chinese LLMs.
DeepSeek’s next model, DeepSeek-v3, launched in December, shocked the AI community. The model utilized a Mixture-of-Experts (MoE) architecture, pre-trained on 14.8 trillion tokens, boasting 671 billion total parameters—though only 37 billion were activated per token. MoE models use specialized sub-models or “experts” that activate only when relevant, making them more efficient and computationally effective.
Former U.S. President Donald Trump called DeepSeek’s advancements a “wakeup call” for American industry, emphasizing the need for continued AI innovation in the U.S.
DeepSeek-v3 was trained for 2.78 million GPU hours using Nvidia’s H800 GPUs. For comparison, Meta’s Llama 3.1, trained with Nvidia’s H100 chips, required 30.8 million GPU hours. This efficiency allowed DeepSeek to develop its most advanced reasoning models—DeepSeek-R1-Zero and DeepSeek-R1—which have shaken up the AI industry due to their unprecedented affordability.
Compared to OpenAI’s O1 model, DeepSeek-R1 reduces costs by an astonishing 93% per API call. This dramatic reduction in expenses is a game-changer for businesses and developers seeking to integrate AI without excessive costs.
Additionally, R1 is designed for local computing, eliminating the need for costly cloud services and restrictive API rate limits. The model’s efficiency allows it to run on high-end personal computers, making AI-driven applications more accessible. By leveraging cloud-based task batching, DeepSeek further optimizes performance and cost efficiency.
Although DeepSeek’s R1 does not yet surpass OpenAI’s O3, it is comparable to O1 on several key benchmarks. LiveBench scores show that O1 outperforms R1 slightly, with a global average of 75.67 compared to DeepSeek’s 71.38. OpenAI’s model remains stronger in complex reasoning tasks, maintaining an edge in problem-solving and language processing.
In areas like coding, mathematics, and data analysis, however, DeepSeek’s R1 is proving to be a formidable contender. It has demonstrated superior performance in analyzing large datasets, a crucial capability for AI-driven decision-making in finance and research.
Even OpenAI CEO Sam Altman acknowledged DeepSeek’s progress, calling R1 “impressive.”
Despite its technological strides, DeepSeek-R1 still adheres to China’s stringent censorship policies. Like its predecessor, Ernie Bot, R1 avoids politically sensitive topics. Questions about President Xi Jinping, the Tiananmen Square protests, or human rights concerns elicit the familiar response: “Let’s talk about something else.”
Nevertheless, DeepSeek’s AI assistant has gained international popularity. Its app has surpassed competitors like ChatGPT, Gemini, and Claude to become the most downloaded AI assistant worldwide. The rapid adoption of R1 underscores its appeal as a cost-effective and high-performing alternative.
OpenAI’s O4 remains the gold standard in AI, but DeepSeek’s innovations suggest that smaller, more efficient models may soon redefine the landscape. DeepSeek’s approach involves distillation—a process where smaller models inherit the reasoning capabilities of larger ones. This method reduces computational costs and improves accessibility.
DeepSeek’s research indicates that its distilled models outperform similarly sized models trained using traditional reinforcement learning (RL). Specifically, its 32-billion-parameter distilled model, DeepSeek-R1-Distill-Qwen-32B, surpasses Alibaba Cloud’s Qwen-32B-Preview across multiple benchmarks.
This shift towards efficient AI models could decentralize computing, reducing reliance on large data centers and reshaping the AI infrastructure market.
sWhile distillation presents an exciting avenue for AI development, it has its limitations. Distilled models remain dependent on their larger “teacher” models, inheriting both their strengths and weaknesses. They may struggle with complex reasoning tasks beyond the capabilities of their predecessors.
Ultimately, as AI advances, demand for powerful GPUs and high-performance computing will continue to rise. While DeepSeek’s R1 showcases remarkable cost efficiency, the U.S. remains better positioned in terms of computational resources and cutting-edge chip technology.
DeepSeek’s R1 and OpenAI’s O1 represent the first successful reasoning models in AI, with R1 pioneering reinforcement learning for reasoning tasks. Moving forward, the industry will require greater computational power to train, experiment, and push the boundaries of AI. The global AI race is far from over, and China’s DeepSeek is proving to be a formidable competitor.
