This article first appeared in The Edge Malaysia Weekly on February 3, 2025 - February 9, 2025

The artificial intelligence race, which in recent years supported the case of a global multibillion-dollar AI data centre infrastructure boom, has reached an inflection point which investors fear could put the entire investment story at risk.

Tighter US restrictions on AI processor exports, aimed at limiting China’s access, have raised concerns about their impact on companies assembling, buying and leasing AI computing power.

Then China sprang a surprise with DeepSeek’s latest chatbot, which rivals its US counterparts despite claims of less funding, less computing power (and using second-tier processors) and reduced energy consumption.

Days after, Alibaba Cloud released its latest large language model (LLM) dubbed Qwen 2.5-Max on Chinese New Year, while MoonShot AI launched the Kimi k1.5, joining a growing list of Chinese AI chatbots, proving that the AI competition is far from over.

Regulatory and technological risks involving the world’s two largest economies have stalled the multi-year AI frenzy.

Nvidia Corp, the nearly US$3 trillion market cap “poster boy” for AI, whose industry-leading chips were so sought after that they were used as collateral for financing, shed as much as 20% in value from its January peak.

In Malaysia, premium valuations ascribed to companies positioned along the AI value chain, which helped drive some stocks to double in value, are being tested.

Over RM35 billion in market capitalisation was erased in these companies so far this year, led by Nvidia-powered data centre developer YTL Power International Bhd (KL:YTLPOWR) (down RM10.74 billion in market cap); national utility firm Tenaga Nasional Bhd (KL:TENAGA) (RM7.79 billion); construction firms Gamuda Bhd (KL:GAMUDA) (RM3.99 billion) and IJM Corp Bhd (KL:IJM) (RM2.35 billion), and Nvidia server assembler NationGate Holdings Bhd (KL:NATGATE) (RM1.68 billion).

The sell-off discounted more than AI-related ventures in some of the companies’ valuation.

Dave Plummer, a former software engineer known for authoring the Task Manager app on Microsoft Windows, warns in his review of DeepSeek that “companies heavily reliant on AI licensing, cloud infrastructure, Nvidia’s chips, or API (Application Programming Interfaces) integrations could face downward pressure as investors factor in lower projected growth or increased competition”.

A Malaysia- and Philippines-based co-location data centre developer cautioned that companies “should tread carefully in the next 48 months”.

However, New York University economist Nouriel Roubini argues that the “positive supply shock” will drive massive demand for AI computing.

A founder and managing director of a regional digital services firm echoes this, saying cost reductions and lower barriers to entry could accelerate mass adoption. “It’s definitely a Sputnik moment [when Russia surprised the US with the first satellite in orbit in the 1950s], leading to the Jevon Paradox,” he says, pointing to how the increased efficiency will result in more — not less — consumption of the required resources such as electricity and computing power.

AI data centres have been Malaysia’s key investment theme, earning it the title of the top Asean data centre destination for 2023/24 by Knight Frank. Until 2021, the country had only a handful of such facilities.

The story has it that data centre owners eye the abundance of land, cheap water and cheap energy here. Since 2023, Johor and Selangor have seen major land sales and data centres being constructed for hyperscalers. The boom also hints at higher water demand, with YTL Power acquiring Johor’s water operator Ranhill Utilities Bhd (KL:RANHILL) in 2024.

Electricity demand is projected to rise by 35% by 2030 from the data centres alone, according to Tenaga, translating into more grid infrastructure work for contractors. Sime Darby Property Bhd (KL:SIMEPROP) has pivoted to leasing entire buildings to data centre operators for stable recurring income, while NationGate is seen as a proxy to Nvidia sales.

In a Jan 31 note, MIDF Research says it believes that fears over US chip restrictions and Malaysia’s role as a global data centre hub “may be overstated”. “Malaysia remains a prime location for data centres due to its abundant power and water resources, cost advantages and established infrastructure. Consolidating all data centres in the US may not be entirely feasible given the massive energy and cooling requirements.”

MIDF also flags how Microsoft CEO Satya Nadella reaffirmed a US$80 billion OpenAI investment, while Meta CEO Mark Zuckerberg committed US$65 billion to data centre capital expenditure. This “confirms that big-tech investment in AI infrastructure is not slowing down and data centre investments in Asean countries such as Malaysia are not expected to experience a slowdown”, the research house adds.

Still, Malaysian AI-linked stocks remain above their long-term forward price-earnings multiples, except YTL Power, IJM and Southern Cable Group Bhd (KL:SCGBHD), along with newer Bursa Malaysia listings such as NationGate, MN Holdings Bhd (KL:MNHLDG) and SNS Network Technology Bhd (KL:SNS).

In a separate note, Kenanga Research says YTL Power’s data centre valuation “is fully discounted” below RM3.96 per share and is supported by its power business in Singapore as well as water and property business in the UK.

The selldown on AI-linked stocks also affected firms with secured contracts, such as data centre construction jobs under Gamuda, Sunway Construction Group Bhd (KL:SUNCON) and SimeProp.

Gamuda, according to a JP Morgan report downgrading the stock, says data centres accounted for just 7% of its outstanding order book.

SunCon’s portion makes up 70% of its 2024 orders, although its target order book replenishment of at least RM4 billion this year takes into account non-data centre infrastructure tenders as well, says MIDF Research.

SimeProp’s build-and-lease agreements involving two data centres for up to RM7.6 billion are set for commissioning in 2026. This makes up a small portion of the analysts’ target prices thus far — 20 sen per share in Kenanga’s target price of RM1.78, for example.

Mah Sing said in a Jan 28 statement that the DeepSeek narrative is a promising development “especially in the context of the growing demand for high-performance computing”.

“In this environment, land near energy hubs will become increasingly valuable,” says Kenanga Research.

The same trend is unfolding in the US, where Chevron on Jan 29 said it plans to build gas power plants adjacent to data centres, together with partners Engine No. 1 and GE Vernova.

Despite the share prices stabilising at current levels, an investment bank research head cautioned that upside catalysts are lacking. “Downside is limited, but any rebound will be short-lived — it is a rangebound consolidation. The situation remains fluid.”

Mah Sing, for example, is targeting to firm up its joint venture with Bridge Data Centres in May.

Over at YTL Power, investors are still waiting for the company to firm up its agreements with Nvidia and off-takers. The agreements are “needed to re-rate” the stock, says Kenanga Research.

NationGate too has yet to officially project the impact from the US chip restrictions and the DeepSeek reveal on its sales.

In December, Kenanga Research reported that NationGate had targeted a turnover target of RM10 billion in FY2025 — it made sales of RM2.24 billion in the first nine months of 2024 — with a 1:10 delivery ratio between AI and non-AI servers.

As such, one may look at Nvidia’s chip sales to assess its impact on NationGate. US investment management firm VanEck points to two scenarios — where the improved efficiency in developing AI models “does not eliminate Nvidia but accelerates a timeline where Nvidia’s near total dominance in AI training normalises”. Even if DeepSeek’s claims are overstated, Nvidia “may continue leading in training. But as AI matures, more players will compete for different parts of the AI hardware stack”, says the firm.

After losing nearly US$600 billion in market value, Nvidia dismissed demand concerns, stating DeepSeek’s influx of users would require “significant numbers” of graphics processing units or GPUs.

OpenAI founder Sam Altman commended the R1 “particularly around what they’re able to deliver for the price” but asserted that OpenAI “will obviously deliver much better models”.

DeepSeek’s pricing — US$0.55 per million input tokens and US$2.19 per million output tokens — is just 3.65% of OpenAI’s comparable O1 model pricing of US$15 (input) and US$60 (output).

However, some are disputing DeepSeek’s claims.

SemiAnalysis, an independent research firm specialising in the semiconductor and AI industries, notes that the US$6 million cost mentioned by DeepSeek refers to pre-training costs and excludes what it estimated to be over US$500 million in capital expenditure.

As for the US chip restrictions, an analyst who covers the tech sector believes there will be minimal impact on Malaysian AI data centre ventures, as hyperscalers will still require third-party facilities for excess capacity at competitive prices.

However, the commoditisation of chips could translate into data centres becoming agnostic, with less advanced options (think buying last year’s phone rather than the latest version, which is more expensive) and from multiple manufacturers to potentially meet the demand.

This again would affect the superior margins and valuations associated with those leasing out Nvidia products.

Nonetheless, open source models like the DeepSeek R1 “democratises access to AI capabilities, as companies and governments could build upon it without fears of restrictions imposed by US firms”, Plummer was quoted as saying.

The progress should be good news as it means countries like Malaysia will increasingly be able to afford AI developments, says the founder and managing director of the regional digital firm.

“Silicon Valley’s argument is this: It does not matter if DeepSeek is able to cut costs by multi-fold, as long as Silicon Valley scales and achieves artificial general intelligence (AGI) first,” he says, referring to the milestone to watch in the AI race — when the model matches human capabilities across a wider range of cognitive tasks.

“What matters is who has the best proprietary data for training [the AI models]. Two schools of thought — the first is that scaling laws have plateaued, the other is that as long as hyperscalers keep investing, they will make a breakthrough. We will soon know who is right,” he adds.

Save by subscribing to us for your print and/or digital copy.

P/S: The Edge is also available on Apple's App Store and Android's Google Play.