This article was originally published on Fool.com. All figures quoted in US dollars unless otherwise stated.
Last week, semiconductor stocks like Nvidia (NASDAQ: NVDA), Advanced Micro Devices, and Micron Technology plunged on news that a Chinese start-up called DeepSeek had figured out how to train artificial intelligence (AI) models for a fraction of the cost of its American peers.
Investors were concerned that DeepSeek's innovative approach would trigger a collapse in demand for graphics processors (GPUs) and other data centre components, which are key to developing AI. However, those concerns might be overblown.
Meta Platforms (NASDAQ: META) is a huge buyer of AI chips from Nvidia and AMD. On January 29, CEO Mark Zuckerberg made a series of comments that should be music to the ears of investors who own AI hardware stocks.
DeepSeek background
Successful Chinese hedge fund High-Flyer has been using AI to build trading algorithms for years. It established DeepSeek as a separate entity in 2023 to capitalise on the success of other AI research companies, which were rapidly soaring in value.
Last week's stock market panic was triggered by DeepSeek's V3 large language model (LLM), which matches the performance of the latest GPT-4o models from America's premier AI start-up, OpenAI, across several benchmarks. That isn't a concern at face value, except DeepSeek claims to have spent just $5.6 million training V3, whereas OpenAI has burned over $20 billion since 2015 to reach its current stage.
To make matters more concerning, DeepSeek doesn't have access to the latest data centre GPUs from Nvidia, because the U.S. government banned them from being sold to Chinese firms. That means the start-up had to use older generations like the H100 and the underpowered H800, indicating it's possible to train leading AI models without the best hardware.
To offset the lack of computational performance, DeepSeek innovated on the software side by developing more efficient algorithms and data input methods. Plus, it adopted a technique called distillation, which involves using a successful model to train its own smaller models. This rapidly speeds up the training process and requires far less computing capacity.
Investors are concerned that if other AI firms adopt DeepSeek's approach, they won't need to buy as many GPUs from Nvidia or AMD. That would also squash demand for Micron's industry-leading data centre memory solutions.
Nvidia, AMD, and Micron power the AI revolution
Nvidia's GPUs are the most popular in the world for developing AI models. The company's fiscal year 2025 just ended on January 31, and according to management's guidance, its revenue likely more than doubled to a record $128.6 billion (the official results will be released on February 26). If recent quarters are anything to go by, around 88% of that revenue will have come from its data centre segment thanks to GPU sales.
That incredible growth is the reason Nvidia has added $2.5 trillion to its market capitalisation over the last two years. If chip demand were to slow down, a lot of that value would likely evaporate.
AMD has become a worthy competitor to Nvidia in the data centre space. The company plans to launch its new MI350 GPU later this year, which is expected to rival Nvidia's latest Blackwell chips that have become the gold standard for processing AI workloads.
But AMD is also a leading supplier of AI chips for personal computers, which could become a major growth segment in the future. As LLMs become cheaper and more efficient, it will eventually be possible to run them on smaller chips inside computers and devices, reducing reliance on external data centres.
Finally, Micron is often overlooked as an AI chip company, but it plays a critical role in the industry. Its HBM3E (high-bandwidth memory) for the data centre is best in class when it comes to capacity and energy efficiency, which is why Nvidia uses it inside its latest Blackwell GPUs. Memory stores information in a ready state, which allows the GPU to receive it instantaneously when needed, and since AI workloads are so data intensive, it's an important piece of the hardware puzzle.
Mark Zuckerberg might have put recent concerns to bed
Meta Platforms spent a whopping $39.2 billion on chips and data centre infrastructure during 2024, and it plans to spend as much as $65 billion this year. Those investments are helping the company further advance its Llama LLMs, which are the most popular open-source models in the world, with 600 million downloads. Llama 4 is due to launch this year, and CEO Mark Zuckerberg thinks it could be the most advanced in the industry, outperforming even the best closed-source models.
On January 29, Meta held a conference call with analysts about its fourth quarter of 2024. When Zuckerberg was quizzed about the potential impact of DeepSeek, he said it's probably too early to determine what it means for capital investments into chips and data centres. However, he said even if it results in less capacity requirements for AI training workloads, it doesn't mean companies will need fewer chips.
Instead, he thinks capacity could shift away from training and toward inference, which is the process by which AI models process inputs from users and form responses. Many developers are moving away from training models by using endless amounts of data, and focusing on "reasoning" capabilities instead. This is referred to as test-time scaling, and it involves the model taking extra time to "think" before rendering an output, which results in higher-quality responses.
Reasoning requires more inference compute, so Zuckerberg thinks companies will still need the best data centre infrastructure to maintain an advantage over the competition. Plus, most AI software products haven't achieved mainstream adoption yet, and Zuckerberg acknowledges that serving many users will also require additional data centre capacity over time.
So, while it's hard to put exact numbers on how DeepSeek's innovations will reshape chip demand, Zuckerberg's comments suggest there isn't a reason for Nvidia, AMD, and Micron stock investors to panic. In fact, there is even a bullish case for those stocks over the long term.
This article was originally published on Fool.com. All figures quoted in US dollars unless otherwise stated.
