Will DeepSeek slow down data centre boom?

Chinese artificial intelligence (AI) startup DeepSeek made headlines and sent ripples through the global technology market, thanks to its low-cost AI app.

Within a week of its release on Apple’s app store, DeepSeek’s AI application has become the most downloaded free app.

As the app gained momentum, herd behaviour kicked-in and the exceptional premium of US equity has narrowed as nervous investors amplified their selling of the US Big Tech in favour of cyclicals.

In particular, Nvidia Corp, which is at the centre of the AI revolution for its graphic processing units (GPUs), saw its market capitalisation plunge by over US$600bil as investors question the risk rewards trade-off from AI.

At home, Nvidia’s partner in the AI data centre space, YTL Power International Bhd’s shares fell into a tighter grip of the market bears with the rise of DeepSeek, adding to the pressure from the quota limits through semiconductor law signed by former President Joe Biden earlier.

With fears arising about its ability to gain Nvidia’s GPUs for its data centres in question, YTL Power shares fell from a high of RM4.10 per share on Monday to a low of RM2.93 on Wednesday, before rebounding to close at RM3.11 before the Chinese New Year break, thus erasing nearly RM8bil of its market value.

YTL Corp Bhd, the controlling shareholder in YTL Power, saw similar losses, dipping from a high of RM2.41 to a low of RM1.81, recovering to RM1.90 over the same period, wiping out RM5.5bil in market cap.Is the data centre thematic over?

The DeepSeek launch raises the question about the need for data centre capacity in growing markets like Malaysia.

To note, in December 2023, Nvidia partnered with YTL Power to develop an AI data centre with a pledged investment of RM20bil.

In 2024 alone, Malaysia secured a total of US$16.9bil in investments from global tech giants like Amazon Web Services, Microsoft Corp, Google and Oracle Corp, driven by the growing demand of information technology infrastructure like data centres driven by growth of cloud computing and AI.

A tech analyst tied to a local brokerage told Starbiz 7: “We have heard from a few industry players on this, and it looks like those investments are not at threat, or at least the ones that have already come in.”

He explains that there are two types of data centres – one caters to hyperscalers for cloud computing and enterprise purposes, while the other focuses on AI applications.

“A data centre corresponds to the number of chips.

“Every data centre holds X number of servers, and each server holds Y number of chips, for example.

“What makes a data centre an AI data centre is the presence of these advanced AI chips in the servers.”

In contrast, hyperscaler data centres, such as those being developed by major players like Microsoft, Google and Oracle in Malaysia, require only traditional servers without the need for advanced AI chips.

Therefore, the analyst believes the hyperscaler data centre boom, especially in Johor, is likely to continue.

“It’s also because land in Singapore is really expensive and its data centre market is relatively mature compared to Malaysia. Johor offers cheaper land and better access to power, making it a natural choice for expansion.”

However, the situation with the proposed AI data centres by companies like YTL Power differs, he says.

It’s a view echoed by MIDF Research. In its latest report the research house noted demand for data centre’s could still be there.

“We believe this could present an opportunity to moderate the impact of the United States AI chips restrictions that was announced by President Biden during his last days in office.

“Not all data centres require the most advanced chips, so not all AI-ready data centres require the latest Nvidia GB200.”

However, the analyst suggest that Nvidia might slow down its investments in the AI sector if apps like DeepSeek continues to gain momentum and potentially lead the AI race.

“In the worst-case scenario, we could see Nvidia pull back on their investments, including with YTL Power,” he explains.

“It might not cancel the project but could put on hold or slow down investment and expansion plans.

“Instead of investing RM20bil, for instance, Nvidia now might scale it down to RM10bil.”

Meanwhile, on the plunge of YTL Power’s shares, the analyst says the stock market is sentiment-driven at the end of the day.

“The surge in YTL Power’s share price is purely based on sentiment because it hasn’t disclosed who are the off takers for the proposed AI data centre or whether they are even going to be customers for that AI data centre.

“The stock market at the end of the day is just sentiment-driven to a great extent,” he says.

On the impact of DeepSeek, the analyst says there are significant uncertainties surrounding this development, especially given the lack of “empirical evidence”.

“On a global level, with a lot of things that come out of China, you have to put some level of doubt until empirical evidence is produced to actually show that this is the case. Right now, not much is known about DeepSeek,” the analyst explained.

Assuming the claims are accurate, he said the global implications could be significant, particularly for Nvidia and its contract manufacturer Taiwan Semiconductor Manufacturing Co Ltd (TSMC).

“Nvidia and TSMC’s growth was premised upon the fact that ChatGPT was so successful that every large corporation in the world wants to use AI and will need their own supply of chips.

“So, it’s based on this simple fact – the only way that ChatGPT could operate was by buying lots and lots of the Nvidia GPU chips, driven by raw computing power.”

Turning his focus to Malaysia, the analyst warned the DeepSeek low cost model could dim Malaysia’s role as a prime location for AI data centre investments.

“If Malaysia was going to be the location for more AI data centre investments over the next five years, the trajectory of that now is obviously going to slow,” he concludes.

But with tech giants like Microsoft intending to invest some US$80bil this year into information technology infrastructure worldwide as substantial fresh investments into AI are announced, the data centre thematic may still have some legs to carry on.

Nevertheless DeepSeek’s AI model, DeepSeek-R1, has changed the race somewhat for delivering performance comparable to leading AI systems like OpenAI’s GPT-4, but at a fraction of the cost.

The model was trained using about 2,000 Nvidia H800 GPUs, incurring a cost of around US$5.58mil – significantly less than the estimated over US$100mil spent by OpenAI for GPT-4’s training in 2023.

For instance, DeepSeek-R1’s application programming interface (API) costs just US$0.55 per million input tokens and US$2.19 per million output tokens, compared to OpenAI’s API, which costs US$15 and US$60, respectively.