KUALA LUMPUR (Jan 28): Despite stringent chip restrictions imposed by the US, DeepSeek has managed to deliver performance comparable to leading AI models like OpenAI’s ChatGPT, but at a fraction of the cost.

The development has sent selling waves across global markets, particularly hitting hard on US chipmakers, as well as AI-proxy stocks, including those on Bursa Malaysia.

But what exactly is DeepSeek, how does it work, and what are its broader implications?

In a research note to its clients, Kenanga explains that DeepSeek is an open-source AI large language model (LLM) developed in China that has demonstrated remarkable capabilities in reasoning and language processing.

What sets it apart is its cost efficiency.

According to Kenanga Research, while OpenAI reportedly spent between US$80 million to US$100 million to train its latest ChatGPT model, DeepSeek’s V3 model was trained for less than US$6 million.

The cost advantage is largely attributed to DeepSeek’s innovative approach to leveraging widely available models and computing resources, as well as its ability to overcome hardware limitations imposed by US export controls on high-end chips.

DeepSeek’s R1 model employs reinforced learning, a technique that allows the AI to learn through trial and error.

By using a reward system, the model develops reasoning capabilities while consuming significantly less memory — reportedly 75% less than traditional models.

This efficiency is achieved by focusing on delivering results that are "adequately accurate" rather than striving for perfection, which reduces computational demands, explained Kenanga Research.

The model was reportedly built using just 2,048 NVIDIA H800 chips, a modified version of the H100 chip that has been subject to US export restrictions since 2023.

"This has raised questions about whether DeepSeek’s success could challenge the dominance of US chipmakers like NVIDIA, which have been at the forefront of supplying high-end chips for AI development," said the research outfit.

The rise of DeepSeek has led to a significant sell-off in US chipmaker stocks, with NVIDIA’s share price dropping by 17% in a single day.

The concern is that DeepSeek’s cost-efficient model could reduce the demand for high-end chips, which are essential for training advanced AI models.

"The impact was more broad based as ASML, manufacturer of high-end equipment for chipmaking, slipped 7%, while energy stocks such as Vistra Corp (nuclear power energy) also slid 28%. This signals to us more broad-based apprehension," said Kenanga Research.

In Malaysia, NVIDIA proxies such as YTL Power International Bhd (KL:YTLPOWR) and Nationgate (KL:NATGATE) also bear the brunt of heavy selldown.

YTL Power, an NVIDIA cloud partner, is expected to receive state-of-the-art GB200 chips, but the timing of demand for these chips is now uncertain, said Kenanga Research. 

Similarly, NATGATE, which relies heavily on AI server sales, faces challenges due to tighter export controls and the potential of reduced demand for high-end chips, it added.

The emergence of DeepSeek is expected to prompt big tech firms to reevaluate their AI strategies.

Companies like Meta and Microsoft, which have pledged billions in AI infrastructure, may relook at their capital expenditure (capex) plans. Meta’s projected capex for 2025 is between US$60 billion to US$80 billion, while Microsoft plans to spend US$80 billion on AI-enabled data centres.

DeepSeek’s success could lead to a more cautious approach, as companies assess whether they need to invest in the most advanced chips, or if more cost-effective solutions could suffice, said Kenanga Research.

While the demand for high-end chips may wane, the need for data centres is expected to grow. Kenanga Research pointed out that DeepSeek’s emergence has highlighted the importance of cost efficiency in AI development, which could accelerate the construction of data centres globally.

In Malaysia, contractors like Gamuda Bhd (KL:GAMUDA) are well-positioned to benefit from this trend, according to Kenanga Research, as the country offers competitive advantages in terms of regulatory clarity and power availability.

The construction of data centres is likely to continue, if not accelerate, as companies seek to build the infrastructure needed to support AI development.

The US government’s AI diffusion interim final rule (IFR), which imposes tighter controls on chip exports, has added another layer of uncertainty to the AI landscape. However, the success of DeepSeek could prompt a reevaluation of these regulations.

"We also await any policy response by Trump on Biden’s AI diffusion interim final rule, a process that should avail clarity by 120 days (i.e. mid May), and may help remove uncertainty on local AI/DC (data-centre) stocks.

"While we are not expecting Trump to rescind the IFR in whole, a less restrictive chips exports ruling following the emergence of DeepSeek would be a positive outcome for Malaysian AI/DC stocks," said Kenanga Research.