DeepSeek disrupts market with budget-friendly R1 model
Rivalling larger AI models
Chinese artificial intelligence (AI) company DeepSeek launched its R1 model last week, causing significant disruption in the AI industry. The announcement sent shockwaves through the stock market, causing the Nasdaq to drop by 3 per cent on Monday, a decline attributed to a nearly 17 per cent loss in shares of chip maker Nvidia.
A research paper released alongside the DeepSeek-R1 launch is at the heart of the controversy. The paper reveals that the company spent just $6 million on computing power for each training run of the model. This figure is notably lower than the estimated costs for training major AI systems like OpenAI’s ChatGPT and Google’s Gemini, which rely heavily on high-end computing infrastructure.
Despite the relatively modest budget, the DeepSeek-R1 has garnered positive reviews for its impressive performance, which rivals much larger and more expensive AI models. Central to the development of these AI systems are powerful computer chips, such as those manufactured by Nvidia, which have traditionally been seen as crucial for delivering top-tier performance.
However, DeepSeek’s breakthrough suggests that high-quality AI results can be achieved with a smaller budget, fewer chips, and even less advanced technology.
According to chatbot provider DocsBot, which recently published a detailed pricing breakdown, DeepSeek-R1’s cost-effectiveness extends to its commercial pricing. Uploading 1 million tokens (the smallest unit of AI processing, approximately four characters long) costs just 55 cents.
In contrast, downloading 1 million tokens from the system costs $2.19. By comparison, the most similar AI model on the market—OpenAI’s ChatGPT-o1 Mini—charges considerably higher fees. While ChatGPT-o1 Mini is also a text-based model like DeepSeek-R1, its commercial costs range from $3 to $5 for 1 million tokens input and $12 to $15 for 1 million tokens output.
Regarding performance, the two models—DeepSeek-R1 and ChatGPT-o1 Mini—are comparable across several industry benchmarks assessing subject knowledge, understanding, reasoning, and accuracy. DeepSeek-R1 even slightly outperforms its US counterpart in certain areas despite having similar token capacities. DeepSeek-R1 can handle up to 32,000 tokens in a single request, while ChatGPT-o1 Mini can process up to 65,500 tokens.
Elon Musk’s xAI has also entered the AI race with its Grok model. This model boasts a larger output window of 128,000 tokens and supports text and image inputs. However, Grok’s benchmark performance is slightly below that of DeepSeek-R1, and it is also much more expensive for commercial use.
Google and Amazon models, such as Gemini and Nova, are cheaper than OpenAI and xAI but still cost more than DeepSeek-R1. However, these models are primarily designed for multimedia inputs, making their output only partially comparable to DeepSeek-R1.
At the more budget-friendly end of the spectrum is Nvidia’s Llama 3.1 Nemotron 70B Instruct, an open-source, text-only model based on Meta’s Llama AI. While it has received strong user feedback and good performance results, it still does not match DeepSeek-R1’s comprehensive capabilities.
DeepSeek’s entry into the market has sparked discussions about the future of AI development, particularly in terms of cost-effectiveness and accessibility. With its ability to deliver impressive results at a fraction of the cost of its competitors, DeepSeek-R1 is expected to have a lasting impact on the AI landscape.
Image: DeepSeek’s breakthrough suggests that high-quality AI results can be achieved with a smaller budget, fewer chips, and even less advanced technology. Credit: Tara Winstead
Last Updated on 2 days by Arnold Pinto
