“`html
Could Nvidia’s Leading Role in AI Be at Risk?
Nvidia‘s (NASDAQ: NVDA) impressive growth over the past two years may be in jeopardy. This surge was largely driven by soaring demand for Nvidia’s high-performance graphics processing units (GPUs). As the AI competition intensified, both tech giants and startups clamored for Nvidia’s cutting-edge GPUs to develop advanced models.
However, last week, the Chinese AI start-up DeepSeek launched its R1 reasoning model, which has notably matched or even surpassed OpenAI’s o1 reasoning model—released just last December—at a fraction of the cost.
Where to invest $1,000 right now? Our analyst team just revealed what they believe are the 10 best stocks to buy right now. See the 10 stocks »
The ability to produce advanced large language models (LLMs) with minimal computing power suggests AI companies might not need to rely on as many high-cost resources in the future. This could significantly impact Nvidia and other leading AI firms.
Nevertheless, the situation is more intricate than it appears.
Understanding DeepSeek
DeepSeek is an AI research lab created from the quantitative hedge fund High-Flyer, founded by CEO Liang Wenfeng in 2015. The DeepSeek project was initiated in 2023 in response to the groundbreaking launch of ChatGPT.
Since then, DeepSeek has been developing AI models and reportedly secured 10,000 Nvidia A100 GPUs before restrictions were imposed. Additionally, they have a cluster of Nvidia H800s, a capped version of the H100 tailored for the Chinese market. Notably, the H100 is the latest in Nvidia’s line of chips prior to the Blackwell launch.
R1’s Astounding Impact
On January 20, DeepSeek unveiled the R1, its initial “reasoning” model derived from the V3 LLM. Reasoning models are innovative and utilize reinforcement learning, which enables an LLM to navigate a thought process, backtrack when necessary, and explore alternative paths to reach a conclusive answer. This method allows for more precise answers to complex questions compared to traditional models.
Remarkably, R1 has matched or exceeded OpenAI’s o1 in multiple benchmarks while being trained for a fraction of the cost. The training reportedly used just $5.6 million worth of rented GPU hours, a stark contrast to the hundreds of millions spent by OpenAI and other U.S. leaders. Furthermore, DeepSeek charges about one-thirtieth of the operating cost for OpenAI’s o1, claiming to operate with a “small profit” margin. Comparatively, estimates suggest that Meta Platforms‘ (NASDAQ: META) Llama 3.1 405B model required around $60 million in rented GPU hours, whereas V3’s cost is approximately $6 million, despite V3 outperforming Llama in various tests.
Key Factors Behind DeepSeek’s Success
According to insights shared by Kevin Xu, DeepSeek owes its success to three main advantages.
First, Wenfeng designed DeepSeek as an experimental AI lab without a strict business model. While DeepSeek does charge a small fee for commercial use of its platform, the core model remains open-source and freely available. This approach attracted many recent graduates and Ph.D. students from top Chinese universities, fostering an environment of creativity and open experimentation, unlike more traditional tech companies.
Second, DeepSeek operates its own data center, which enabled them to fine-tune the hardware specifically for their needs.
Lastly, they optimized their learning algorithms extensively, enabling better performance from their hardware setups. For example, DeepSeek created a parallel processing algorithm called the HAI-LLM framework, designed to distribute workloads effectively across their limited resource base. They also adopted an F8 data input framework, which trades off some precision for enhanced efficiency, allowing R1 to achieve remarkable results with extensive calculations. Furthermore, DeepSeek optimized its networking load-balancing, ensuring that no hardware resources were underutilized.
These innovations allowed DeepSeek to efficiently utilize its hardware while achieving extraordinary performance gains.
However, this means AI firms could become less reliant on costly hardware, potentially impacting Nvidia’s sales growth and profit margins.
Image source: Getty Images.
Counterarguments to the Pessimism
Despite the challenges posed by R1, there are several perspectives suggesting Nvidia may not be facing imminent doom.
First, skeptics question DeepSeek’s transparency regarding its cost estimates. Machine learning researcher Nathan Lampbert suggests the $5.6 million in GPU hours does not factor in pre-training hours, capital expenditures for hardware, energy costs, and salaries for DeepSeek’s engineering team, which reportedly includes 139 technical contributors. Lampbert estimates DeepSeek’s annual operational costs could fall between $500 million and $1 billion—still less than their American rivals, but significantly higher than claimed.
Some experts suspect DeepSeek might have undisclosed access to a 50,000 H100 GPU cluster, which would be illegal for them to possess under 2022 export limitations. This raises questions about the legitimacy of DeepSeek’s claims.
That said, the methods DeepSeek used for R1 are published publicly, suggesting other researchers could replicate its efficiency breakthroughs with smaller budgets. The consensus currently indicates that R1’s performance improvements are credible.
What Lies Ahead for Nvidia
Although DeepSeek’s advancements are notable, former OpenAI executive Miles Brundage advises caution against overestimating the impact of R1’s introduction. He emphasizes that OpenAI’s features and capabilities still set it apart, indicating that Nvidia’s future may not be as bleak as anticipated.
“““html
AI Competition Heats Up: DeepSeek and the Future of Computing Power
DeepSeek has launched its o3 model and is set to unveil the o5 model shortly. While the company has reached R1 through innovative methods, its limited computing power may hinder its ability to scale and develop further from its initial reasoning model.
Brundage points out that restricted computing resources will impact the performance of these models in real-world applications:
Even if that’s the smallest possible version while maintaining its intelligence — the already-distilled version — you’ll still want to use it in multiple real-world applications simultaneously. You wouldn’t want to choose between using it for improving cyber capabilities, helping with homework, or solving cancer. You’d want to do all of these things. This requires running many copies in parallel, generating hundreds or thousands of attempts at solving difficult problems before selecting the best solution. … To make a human-AI analogy, consider Einstein or John von Neumann as the smartest possible person you could fit in a human brain. You would still want more of them. You’d want more copies. That’s basically what inference compute or test-time compute is — copying the smart thing. It’s better to have an hour of Einstein’s time than a minute, and I don’t see why that wouldn’t be true for AI.
Understanding the Jevons Paradox in AI
Investors should be aware of the Jevons paradox, a term introduced by economist William Stanley Jevons in 1865 about coal consumption. It describes how increasing efficiency in a resource does not lead to reduced use. Instead, this often results in exponentially higher consumption. When a resource becomes more efficient, demand usually grows to exceed previous levels, creating an overall increase in usage.
In the context of AI, if the costs associated with training advanced models decrease, we can expect AI to become integrated into our daily routines more frequently. This rising use could indeed increase demand for computing power, particularly for inference rather than training. Oddly enough, this scenario might actually benefit Nvidia. However, there is a possibility that AI inferencing will become more competitive compared to training, creating challenges for Nvidia. Such challenges would stem from increased competition rather than any slashes in computing demand.
Ultimately, the need for AI computing is projected to expand significantly over the next several years. For instance, on January 24, Meta Platforms CEO Mark Zuckerberg announced plans to build an AI data center nearly the size of Manhattan, boosting the company’s capital expenditure from a range of $38 billion to $40 billion in 2024, to between $60 billion and $65 billion this year.
This move followed closely after DeepSeek’s release, which makes it likely that Zuckerberg was aware of its implications. His decision shows confidence in a substantial increase of over 50% in spending on AI infrastructure.
The introduction of DeepSeek will undoubtedly impact the AI competition landscape. Nevertheless, it is not the endgame for Nvidia and the other firms known as the “Magnificent Seven.” Instead, the competition will play out in more complex and multifaceted ways.
As the AI industry progresses, investors must closely evaluate which companies can establish a sustainable AI “moat,” especially as business models continue to develop rapidly, as evidenced by DeepSeek R1.
A New Opportunity for Savvy Investors
Do you sometimes feel like you lost your chance to invest in successful stocks? If so, you should pay attention here.
Occasionally, our expert analysts pinpoint a “Double Down” stock recommendation for firms they believe are on the verge of significant growth. If you worry about missing your moment to invest, now is an ideal time to act before it’s too late. Here’s what the numbers reveal:
- Nvidia: if you had invested $1,000 with us in 2009, it would have grown to $369,816!*
- Apple: if your investment of $1,000 in 2008 had been with us, it would now be $42,191!*
- Netflix: a $1,000 investment in 2004 would be worth $527,206!*
Currently, we are issuing “Double Down” alerts for three impressive companies, and another chance like this may not arise soon.
Learn more »
*Stock Advisor returns as of January 27, 2025
Randi Zuckerberg, former director of market development for Facebook and sister to Meta Platforms CEO Mark Zuckerberg, sits on The Motley Fool’s board of directors. Billy Duberstein and/or his clients have positions in Meta Platforms. The Motley Fool holds positions in and recommends Meta Platforms and Nvidia. The Motley Fool has a disclosure policy.
The views and opinions expressed herein belong solely to the author and do not necessarily reflect the views of Nasdaq, Inc.
“`
