Nvidia stock: China’s DeepSeek AI model surprises the tech world. Should you consider selling?

Could Nvidia‘s (NASDAQ: NVDA) impressive two-year growth be nearing its end? Thus far, the demand for Nvidia’s cutting-edge graphics processing units (GPUs) has been relentless. With the AI competition escalating, both major tech firms and startups have been hastily acquiring or leasing as many of Nvidia’s top-tier GPUs as possible to enhance their models.

However, last week, Chinese AI startup DeepSeek unveiled its R1 model, which has taken the tech community by surprise. This “reasoning” model reportedly matches or surpasses OpenAI’s recently launched o1 reasoning model, all while being offered at a much lower price.

If AI companies can now produce advanced large language models (LLMs) without relying heavily on costly computational resources, the impact on Nvidia and other AI frontrunners from last year could be significant.

Yet, as usual, the reality is more nuanced.

What is DeepSeek?

DeepSeek is an AI laboratory that originated from a quantitative hedge fund named High-Flyer. Its CEO, Liang Wenfeng, founded High-Flyer in 2015 and initiated the DeepSeek project in 2023 following the groundbreaking launch of ChatGPT.

Since then, DeepSeek has been developing AI models, reportedly having procured 10,000 Nvidia A100 GPUs before they faced restrictions; these are two generations behind the current Blackwell chip. Additionally, it is believed that DeepSeek operates with a cluster of Nvidia H800s, a modified version of the Nvidia H100, geared towards the Chinese market. It’s worth mentioning that the H100 represents Nvidia’s latest GPU generation before the recent Blackwell release.

R1 shocks the world

On January 20, DeepSeek launched the R1, its first “reasoning” model based on its V3 LLM architecture. Reasoning models are a relatively new concept that employ a technique called reinforcement learning. This approach encourages an LLM to follow a line of thought, backtrack when necessary, and explore alternative avenues before arriving at a conclusion. As a result, reasoning models are capable of answering intricate questions with greater accuracy than traditional Q&A models.

Remarkably, R1 has been shown to match or even surpass OpenAI’s o1 across various benchmarks, reportedly trained at a mere fraction of the cost.

So, just how affordable is this? The R1 research document claims the training equated to only $5.6 million in rented GPU hours, a tiny fraction compared to the hundreds of millions invested by OpenAI and other U.S.-based leaders. Moreover, DeepSeek charges around one-thirtieth of the running cost of OpenAI’s o1, with Wenfeng stating that DeepSeek charges only a modest profit margin on costs. Industry experts have estimated that Meta Platforms‘ (NASDAQ: META) Llama 3.1 405B model incurred about $60 million in rented GPU hours, in contrast to approximately $6 million for V3, even as V3 surpassed Llama’s latest iteration in multiple benchmarks.

How DeepSeek Achieved This

As per an insightful blog by Kevin Xu, DeepSeek’s achievements can be attributed to three distinct advantages.

First, Wenfeng designed DeepSeek as an idealistic AI research hub eschewing a fixed business model. Presently, DeepSeek charges a nominal fee for other developers building products on its platform, while offering its open-source model for free. He also brought on board mainly young graduates and Ph.D. candidates from China’s premier institutions. This setup fostered a culture of free experimentation, enabling a trial-and-error approach, unlike the more rigid structures found in larger Chinese tech firms.

Secondly, DeepSeek operates its own data center, allowing it to fine-tune the hardware setup to meet its specific needs.

Lastly, DeepSeek optimized its learning algorithms in various ways, which cumulatively enhanced the effectiveness of its hardware.

For example, DeepSeek developed a proprietary parallel processing algorithm called the HAI-LLM framework, which improved how computing workloads were managed across its limited available chips. Additionally, it utilized the F8 data input framework, a less precise but more memory-efficient system compared to F32. Although F8 sacrifices some precision, R1’s other processes compensated for this shortcoming by performing a larger volume of efficient calculations. DeepSeek also optimized its network’s load-balancing kernel to maximize the output of each H800 cluster, ensuring hardware was never idle awaiting data.

These are just a few of the innovations that enabled DeepSeek to achieve impressive results with fewer resources. By integrating all these adjustments, they significantly boosted their performance.

This suggests that Nvidia may face challenges as AI firms become less reliant on high-end hardware due to these software innovations, potentially impacting Nvidia’s sales and profit margins.

Counterpoints to the doom narrative

While R1 presents a considerable challenge for Nvidia, there are several points to argue that Nvidia is not “doomed.”

First, some experts doubt the accuracy of the cost estimates provided by the Chinese startup. Machine learning researcher Nathan Lampbert suggests that the $5.6 million estimate for GPU hours likely omits several significant expenses. These include extensive pre-training time before the actual training of the large model, capital investments associated with purchasing GPUs and constructing data centers (if DeepSeek didn’t rent from external cloud providers), and considerable energy costs. Additionally, the R1 project involved 139 technical contributors whose salaries might not be fully accounted for in the claim. While DeepSeek’s model is open-source, it’s probable that many of these contributors are employees with salaries that contribute substantially to operational costs.

Lampbert speculates that DeepSeek’s annual operational costs could range between $500 million and $1 billion. Although this is still less than their U.S. competitors, it is far more than the $6 million stated in the R1 research paper.

Some skeptics even question DeepSeek’s claims regarding its chip access. In a recent interview, Scale AI’s CEO Alexandr Wang stated that he believes DeepSeek may possess access to a 50,000-strong H100 cluster that hasn’t been disclosed, as these chips are banned in China due to export limitations imposed in 2022.

Nevertheless, since DeepSeek has openly shared the methods used in developing the R1 model, other researchers should theoretically be able to replicate its efficiency with limited resources. Currently, it appears the R1’s efficiency breakthrough has some validity.

Even if true, it may not spell the end for Nvidia

While DeepSeek has undoubtedly made significant strides, ex-OpenAI executive Miles Brundage also cautioned against overestimating the implications of R1’s launch. Brundage highlights that OpenAI has already introduced its o3 model and will be rolling out its o5 model soon. Although DeepSeek has made inventive advancements to reach R1, its limited computational capacity may hinder its progress moving forward.
The speed at which it can grow and develop from its initial reasoning model.

Brundage also emphasizes that limited computational resources will play a crucial role in how these models can operate concurrently in the real world:

Even if you’re working with the smallest feasible version while keeping its intelligence intact — the already-optimized version — you’d still want to deploy it in various real-world applications at the same time. You wouldn’t want to have to pick between enhancing cyber capabilities, assisting with school assignments, or tackling cancer. You would aim to do all of these things simultaneously. This necessitates running numerous copies at once, providing hundreds or thousands of attempts to address tough problems before determining the optimal solution. To draw an analogy with humans, imagine Einstein or John von Neumann as the most intelligent individuals you could fit in a human brain. You would still desire more of them. You would want additional copies. Essentially, this is what inference compute or test-time compute means — replicating the intelligent entity. Having an hour of Einstein’s insights is far better than a minute’s worth, and I find no reason why that wouldn’t apply to AI.

The Jevons Paradox

Lastly, it’s essential for investors to consider the Jevons paradox. Introduced by British economist William Stanley Jevons in 1865 concerning coal consumption, this phenomenon arises when a technological process becomes more efficient. According to Jevons’ paradox, when a resource is utilized more effectively, rather than a reduction in its usage, consumption tends to grow dramatically. The heightened demand typically more than compensates for the gain in efficiency, resulting in an overall uptick in the need for that resource.

In the context of AI, if the expenses associated with training advanced models decrease, we can expect AI to become increasingly integrated into our everyday lives. According to the paradox, this will likely heighten the demand for computing power — primarily for inference rather than for training purposes. This could unexpectedly be a favorable outcome for Nvidia. Conversely, it’s believed that AI inference might become more competitive in comparison to training for Nvidia, potentially leading to a downside. However, this downside would stem from heightened competition rather than a reduction in computing demand.

The key takeaway is that the appetite for AI computing is anticipated to continue its significant growth in the coming years. For instance, on January 24, Meta Platforms CEO Mark Zuckerberg declared that Meta would establish an AI data center almost the size of Manhattan and plans to ramp up its capital expenditures to between $60 billion and $65 billion this year, up from a previous forecast of $38 billion to $40 billion in 2024.

This announcement came just four days following the release of DeepSeek, indicating that Zuckerberg was likely aware of its implications. Yet, he still believes that a sizable increase of more than 50% in AI infrastructure investment is justified.

Undoubtedly, the introduction of DeepSeek will influence the AI landscape. However, instead of a definitive endgame for Nvidia and the other “Magnificent Seven” firms, the situation will be more complex.

As the AI landscape evolves, investors must evaluate which companies truly possess an AI “moat,” since AI business models are transforming rapidly and in unexpected ways, as evidenced by DeepSeek R1.

Randi Zuckerberg, former director of market development and spokesperson for Facebook, and sister to Meta Platforms CEO Mark Zuckerberg, serves on The Motley Fool’s board of directors. Billy Duberstein and/or his clients hold stakes in Meta Platforms. The Motley Fool also owns shares in and endorses Meta Platforms and Nvidia. The Motley Fool has a formal disclosure policy.

The Motley Fool is a content partner with YSL News, offering financial news, analysis, and insights designed to empower individuals in managing their financial decisions. Its content is created independently of YSL News.

Don’t let this second chance at a potential opportunity pass you by

Offer from The Motley Fool: Have you ever felt like you missed out on investing in highly successful stocks? Then you need to pay attention to this.

Occasionally, our team of expert analysts releases a “Double Down” stock recommendation for companies they believe are on the verge of significant growth. If you’re concerned you might have already missed your chance, now is the perfect time to invest before it’s too late. The statistics are compelling:

• Nvidia: if you had invested $1,000 when we doubled down in 2009, it would now be worth $369,816!
• Apple: if you had invested $1,000 when we doubled down in 2008, it would now be worth $42,191!
• Netflix: if you had invested $1,000 when we doubled down in 2004, it would now be worth $527,206!

Currently, we’re issuing “Double Down” alerts for three amazing companies, and opportunities like this may not present themselves again soon.