On the same day of President Trump's inauguration, Chinese artificial intelligence company DeepSeek released DeepSeek-R1 — an AI reasoning model that has emerged as a fresh competitor in the space with stellar results, specifically rivalling the renowned OpenAI-o1 model developed by industry leader OpenAI. The news sent US tech stocks plunging, especially for those in the AI and semiconductor space. In particular, Nvidia (NASDAQ:NVDA) lost almost $600 billion in market capitalisation in a single day, making this the largest single-day-loss in stock market history. Let's dive into the details.
The Value Proposition of DeepSeek's R1
While it may be debatable as to whether DeepSeek's R1 is truly the most precise model in the space, how the model got there is the story that's knocking most people's socks off.
Firstly, R1 was trained for about $5.5 million, which is about 2.5% of the training cost of OpenAI's GPT-4. The training cost of OpenAI's o1 is not disclosed, but it is rumoured to be in the hundreds of millions. The significant edge in training costs translates to DeepSeek's Application Programming Interface (NASDAQ:TILE) (API) pricing is about 25 times cheaper than OpenAI's.
Secondly, DeepSeek's R1 achieved its results with 2.78 million GPU hours, which is significantly lesser than the 30.8 million GPU hours required for Meta (NASDAQ:META)'s Llama model, which operates on a similar scale.
Finally, DeepSeek R1 was allegedly trained using Nvidia's H800 AI GPUs, which have significantly lesser capabilities (as compared to Nvidia's other offerings) due to trade restrictions imposed on sales to China.
The above was made possible due to enhanced memory efficiency through DeepSeek's usage of the Multi-Head Latent Attention (MLA) system. In addition, DeepSeek mainly uses FP8 (8-bit numbers) instead of FP32. Also, through a feature known as Multi-Token Prediction (MTP), DeepSeek's R1 is able to predict multiple tokens without compromising much on accuracy, which greatly enhances its inference speed.
All in all, under relatively tight financial and technological constraints, DeepSeek has managed to rival the performance of Silicon Valley's best AI models. This breakthrough highlights potential inefficiencies that exist within today's AI models which we all know and love.
Implications on Nvidia
Since DeepSeek's R1 has managed to produce such excellent results while training with Nvidia's sub-standard chips, it then begs the question — is there really a need for Nvidia's more powerful and more expensive offerings? In recent years, the world's largest tech companies have been channelling enormous amounts of resources into AI model training, and many apply a brute-force approach by investing in the most powerful chips — which are, of course, produced by Nvidia. With the release of DeepSeek's new R1 model, which is completely open-source, we could see companies repositioning their AI ventures by focusing more on resource efficiency as opposed to hoarding the best chips in terms of computing power.
Nvidia's Response
Nvidia has praised DeepSeek's new model for being a perfect example of test time scaling, and is optimistic about the company's trajectory in the AI space. That being said, they also re-emphasised that DeepSeek's models are largely reliant on inference-time computing, which still requires significant numbers of inference-focused GPUs and high-performance networking. As such, even in an evolving landscape, Nvidia believes it can hold its own by offering the appropriate chips to companies to enhance efficiency. After all, they're here to provide the most competitive chips on the market, which is vastly different from providing the most competitive AI reasoning model in the space.
Is there a need to panic?
Ultimately, it's important to understand the reason behind recent market movements, and the key takeaways from DeepSeek's release of R1. It is not necessarily a certainty that the US tech sector has completely lost its competitive advantage, and that the best move is to rotate into other sectors, or even consider investments in different regions. However, this breakthrough definitely puts the largest tech firms in the world on their toes. It will be interesting to see how the biggest players in the space respond to this over the next few months — will they remain complacent, or fight back to maintain their lead? We will see in due time.