DeepSeek R1: A game changer for AI?

To help make sense of this pivotal moment, OMMAX AI experts Dr. Felix Gerslbeck, Dr. Hardy Kremer, and Anatoli Kantarovich address some of the most pressing questions about DeepSeek and its implications for the industry.

1. How does DeepSeek R1 perform?

In short: exceptionally well. On key AI benchmarks, including math problems, coding challenges, and general knowledge tests, R1 matches or outperforms OpenAI’s flagship model, o1. Although not a groundbreaking leap in performance, R1 has quickly positioned itself as a leading contender in the AI space.

2. What does it mean that DeepSeek is open source?

There are two main reasons that DeepSeek’s rise is so exciting. Firstly, DeepSeek's AI models are open source and available under the MIT license, meaning the code can be freely used, modified, and commercialized. This open-access approach stands in stark contrast to proprietary models from OpenAI and Google, which require licensing fees and where the methodology remains a trade secret. Secondly and more importantly, this also means that researchers and developers can closely examine DeepSeek’s techniques, accelerating the path towards greater transparency and innovation.

3. Has DeepSeek reduced AI training costs?

A key differentiator is DeepSeek’s efficiency. Compared to its competitors, DeepSeek managed to achieve comparable performance with significantly fewer resource investments. The company reportedly has a stockpile of 10,000 H100 Nvidia chips, substantially fewer than those of American models, such as the 100,000 recently installed by Elon Musk’s xAI. Additionally, training costs were reported as approximately $5.5 million, a fraction of what is typically required for cutting-edge AI models. While total operational costs (including salaries, infrastructure, and research) are certainly higher, this streamlined approach represents a major cost-saving breakthrough.

4. Why is this moment significant for AI users? Is this a watershed moment in AI?

Before DeepSeek, developing high-performance AI models was thought to be exclusive to tech giants with vast financial resources. DeepSeek’s success challenges this assumption, proving that advanced AI can be developed at a fraction of the previously assumed cost. While personal AI model training remains out of reach for most individuals, this shift opens opportunities for universities, corporations, and investors to enter the AI space. The result? Increased competition and a wave of innovation driven by more accessible AI development.

5. Why are Silicon Valley and Nvidia concerned?

The rise of DeepSeek and democratization of AI development introduces significant competition for established AI players like OpenAI, Google, xAI, and Anthropic. These companies rely on high subscription and licensing fees to recover their substantial AI investments (e.g., ChatGPT Pro costs $200 per month). DeepSeek’s emergence is likely to drive pricing pressure, forcing a reassessment of business models across the industry.

How about Nvidia? The concerns are different. While AI models still require chips (which will still come from Nvidia for the foreseeable future), Nvidia’s previous sky-high valuation was based on the assumption of continuously increasing hardware demand. DeepSeek’s efficiency-first approach challenges this trajectory, prompting a market correction in Nvidia’s stock.

6. How did DeepSeek achieve this breakthrough?

DeepSeek’s success is attributed to several strategic design choices:

• Training process optimization: DeepSeek’s engineers put significant effort on refining the mathematical computations within the LLM training process, reducing the number of necessary calculations down to the essentials. This highlights the untapped potential for massive efficiency gains through foundational AI research.
• Mixture-of-Experts (MoE) architecture: Unlike general-purpose (“distilled”) models such as ChatGPT, which handle all types of queries with a single massive model, DeepSeek uses specialized sub-models tailored for tasks like mathematics, coding, and text generation. A top-layer model efficiently assigns queries to the most relevant expert, drastically cutting computational costs. If ChatGPT is the equivalent of the CEO calling a meeting with the entire management of the company to decide on all marketing, sales, product, and operations topics together, DeepSeek is the equivalent of the CEO delegating marketing topics to the head of marketing, product topics to the head of product, and so on.
• Reinforcement learning for reasoning: DeepSeek R1 enhances step-by-step reasoning capabilities through reinforcement learning. Instead of relying on vast manually labeled datasets, DeepSeek claims their R1 model learns by trial and error, receiving rewards for correct responses. With enough repetition, the model learns by itself which reasoning steps most help in leading it to correct outcomes. While reinforcement learning has been widely used in fields such as chess AI, DeepSeek has successfully applied it to language model training, significantly improving reasoning efficiency.

Closing thoughts

DeepSeek’s launch has created a strong paradigm shift in AI development by challenging existing business models and accelerating the democratization of AI technology. In the coming months, competition is expected to sharply increase, and time will reveal how industry leaders adapt to this new reality.

Do you want to know more about OMMAX’s experience in artificial intelligence and how we can support you in maximizing ROI in tech, data & AI investments? Get in touch with our experts through the form below!

Sign up for our Decoding Tech, Data & AI series to stay up-to-date with the latest news from the world of data & AI and learn from the most successful data & AI use cases.