Should DeepSeek R2 Worry Nvidia Investors?
In late January, Chinese AI research lab DeepSeek released its open-source and highly efficient AI model, DeepSeek-R1. This shook up the AI world and triggered a large but brief sell-off in Nvidia stock (NASDAQ:NVDA). At one point on January 27th, Nvidia stock was down by as much as 17% following the news. The R1 model was able to match the performance of more established models including OpenAI’s O1 and Meta’s Llama AI, while being cheaper to run and more resource-efficient. This was seen as a threat to Nvidia stock, as lower computational requirements could translate into slower growth in GPU demand. However, Nvidia stock soon bounced back as the market reassessed its long-term prospects, although it faced renewed pressure amid an escalating U.S.-China tech trade war. Now, just as the headwinds appear to be fading, some reports suggest that DeepSeek’s next model, dubbed R2, could be close to launching. Should Nvidia investors be worried?
If you seek upside with lower volatility than individual stocks, the Trefis High Quality portfolio presents an alternative, having outperformed the S&P 500 and generated returns exceeding 91% since its inception.
DeepSeek’s Innovations
DeepSeek’s models emphasize software-driven resource optimization over hardware dependency. These optimizations result in drastically lower costs compared to traditional large language models. R1 was able to deliver competitive performance with limited computing power, driven by these advanced techniques. DeepSeek optimized communication between chips, adopted memory-saving methods, and leveraged reinforcement learning to reduce resource demands. This cost efficiency is reflected in the API pricing for DeepSeek-R1, which costs just $0.55 per million input tokens and $2.19 per million output tokens, dramatically undercutting OpenAI’s API rates of $15 and $60, respectively. See How DeepSeek’s AI Model Impacts Nvidia Stock.
What Could R2 Have In Store?
The new R2 model is reportedly developed using a so-called hybrid Mixture-of-Experts (MoE) architecture, which divides an AI model into separate sub-networks that activate selectively based on the input. This approach could considerably reduce compute costs further pre-training, and enable faster inference performance. The model is reported to have a total of 1.2 trillion parameters, making it up to 97.3% cheaper to train than OpenAI’s GPT-4o [1].
DeepSeek reportedly is training R2 using China’s own Huawei Ascend 910B chips instead of Nvidia GPUs. Systems built on Ascend chips have achieved up to 91% efficiency compared to a similar-sized Nvidia A100-based cluster. DeepSeek is also said to be working on building a local hardware supply chain, further reducing its reliance on U.S.-made chips. This contrasts with the R1 version, which apparently used tens of thousands of Nvidia’s H100 and H200 GPUs to train its models despite U.S. export controls. While Chinese firms have faced constraints in accessing Nvidia chips, they have reportedly been sourcing them via countries such as Singapore and Malaysia to bypass these restrictions. If DeepSeek successfully moves toward Chinese chips, other big tech players in China could follow suit.
What Does This Means For Nvidia?
Over the past two years, companies have funneled massive resources into building AI models, driving Nvidia’s revenue up by over 125% in FY’24 to $61 billion, with net margins nearing 50%. Revenue surged further to over $130 billion in FY’25. If the industry begins to take inspiration from DeepSeek’s approach to open-source model development, we could see a cooling in demand for AI computing power. Moreover, if chips from the likes of Huawei and others are better utilized by DeepSeek’s innovative code, it could enable companies to diversify their demand away from Nvidia’s chips, which have become increasingly pricey. That said, leaner and more efficient models could also have a counter-effect, helping Nvidia to an extent. Leaner models that help to reduce the compute requirements might end up democratizing access to AI, bringing in new users and use cases, which could ultimately expand total demand. We value Nvidia stock at about $101 per share, roughly 10% below the current market price. See our analysis of Nvidia valuation: Expensive or Cheap for more details.
Learn more about Trefis RV strategy that has outperformed its all-cap stocks benchmark (combination of all 3, the S&P 500, S&P mid-cap, and Russell 2000), to produce strong returns for investors. Separately, if you want upside with a smoother ride than an individual stock, consider the High Quality portfolio, which has outperformed the S&P, and clocked >91% returns since inception.
Invest with Trefis Market-Beating Portfolios
See all Trefis Price Estimates