MIDF Research said Monday that the fear over the United States’ artificial intelligence (AI) chip export curbs on Malaysia’s data center development may have been overblown.
The research house said in a note that the market needs to establish at the onset that not all data centers are AI data centers.
“While most of the new buildings are AI-ready, they may eventually host normal servers or provide public cloud offerings,” it noted.
According to MIDF, hyperscale data centers are not necessarily AI data centers.
It highlighted that AI data centers can exist both independently and also within an existing hyperscale facility in an AI-specific cluster.
For instance, the YTL Green Data Centre Park is allocating 100MW out of its 500MW for AI infrastructure.
MIDF noted that hyperscale facilities are designed for general-purpose scalability including cloud services, web hosting, storage, and enterprise applications and may include AI capabilities but their primary focus is broader.
AI data centers on the other hand are tailored specifically for more intense workloads such as AI and machine learning training, it explained.
Secondly, based on its 7 percent estimates above on the 40,806 H100 GPU equivalents, MIDF said the amount is more than sufficient to train cutting-edge AI models and to support global-scale AI services.
“And that, is coming only from Google, which we believe should have no issue obtaining the Universal VEU (UVEU),” it said.
MIDF also expects other US tech giants with presence in Malaysia to join the bandwagon, such as Microsoft, Amazon and Oracle, which should increase the total amount of AI chips that can be deployed to Malaysia.
It also remains cognizant of the fact that these tech giants are increasing their total AI computing capabilities as they continue to train and develop more sophisticated AI models.
News reports have it that Meta is utilizing more than 100,000 Nvidia H100 to train Llama 4 while Microsoft was said to have purchased 485,000 of the same GPU.
As the tech giants continue to enhance their AI TPP in the US, MIDF opined that the base for the 7 percent will continue to grow.
“We can also expect colocation data center providers in Malaysia such as NEXTDC, AirTrunk, Vantage, NTT and others, who are headquartered in Tier-1 countries, to apply for the UVEU or the National VEU (NVEU) at the very least, though we do not expect AI to take up a large portion of their offtakes here as many other services require data centers,” said MIDF.
Perhaps, it noted there may not be any need for that until it is required as not all data centers require the most advanced chips.
According to the research house, graphics processing units (GPUs) are some of the most powerful chips used for AI workloads due to their parallel processing capabilities, and there are other lower powered but still highly relevant AI chips that do not come under the ambit of the ECCN 3A090.a.
It noted not all AI-ready data centers require the latest GB200.
It is noted that the year 2024 was one that was driven by the AI or tech frenzy among other factors, as companies invested in AI capabilities, on top of a growing demand for data centers that drove a surge in such infrastructure investments.
The growth in data center demand made Malaysia a hotspot for such developments, benefitting from a four-year data center moratorium in Singapore that was lifted in Jan-22.
Most of the new builds are hyperscale and, in order to ride on the AI proliferation, a majority of the upcoming data centers are AI-ready.
The announcements of hyperscale and AI-ready data center projects from double to three digits in terms of megawatt (MW) drove optimism in the market, but it all came crashing down when the United States unexpectedly introduced export controls on advanced computing integrated circuits (ICs), or loosely called AI chips.
Analysts see US AI chip export curbs has limited impacts on Malaysia’s tech space