If one wants to know what triggers a broad base sell off in DC stocks - this could be the answer today.
It’s a great write up from SemiAnalysis (like always), been following them for years - never knew they will move the Malaysian market too haha.
@terence775 @boncos @wsk20 @nicholascheh @eugenewong @tapdance
$MNHLDG / 0245 (MN HOLDINGS BERHAD) $GAMUDA / 5398 (GAMUDA BERHAD) $YTLPOWR / 6742 (YTL POWER INTERNATIONAL BHD)
https://cutt.ly/3e9Xjfc4
—
We believe that the exponentially growing Johor market, in Malaysia, has the most to lose. Johor, just a few miles north of Singapore, is by far the most aggressive overseas near-term AI build out. In Malaysia we see a critical IT power of nearly 800MW exiting 2024, forecasting growth to 1,798 MW by 2025, and 3,240 MW by 2027. We believe that AI Accelerator deployments represent close to half of the total estimated 2027 capacity, almost all of which is earmarked for very high rack densities of up to 130kW per rack – suitable for the GB200 NVL72.
However, unlike Brazil and India, one of Johor’s largest end-users is US-restricted company ByteDance. TikTok’s owner is, we believe, the largest user of the “GPU Cloud loophole”, working with both established hyperscalers like Oracle and Microsoft, and Neoclouds.
We show below the scale of the Oracle-ByteDance complex operations, with ByteDance leasing capacity to deploy CPUs and some unrestricted GPUs like B20 and H20, while also renting state-of-the-art GPU systems from Oracle for more resource-intensive workloads. In addition, we know of several Neoclouds with H100 nodes deployed in Johor.
In addition, Malaysia boasts the largest Sovereign AI initiative in the world, by a wide margin, carried out by YTL.
This Malaysian energy-focussed conglomerate has planned the project with multiple angles – buildings for traditional datacenter colocation, and other facilities set aside specifically for providing AI Neocloud services, which include plans to “deploy and manage Nvidia’s AI supercomputer in Johor”.
They first made their big entrance into the datacenter market in 2022, where they announced a 500MW total power datacenter powered by entirely green energy. Soon after, the company stated its partners, tenants, and phases for the project to go live. Sea Ltd would be the anchor tenant with 36MW, another phase that would allocate MW to whichever hyperscaler signed, finally announcing the foray into the AI Neocloud market backed by a USD$4.3B partnership with Nvidia.
A significant partnership with Nvidia and deals with hyperscalers spurred rapid development in the site. Images taken of YTL’s site as of Jan 1, 2025 show the buildings developed for Sea, the respective buildings for the signed Hyperscaler, and the AI Neocloud sites under construction. YTL is building for high density Direct-to-Chip liquid cooling, making it one of the largest and highest power density datacenter facilities globally.
YTL is Malaysia’s featured Sovereign AI initiative, and it is the largest Sovereign AI initiative in the world by far. However, the AI Diffusion Framework presents a key challenge to key aspects of YTL’s Neocloud business plan. If YTL cannot obtain an NVEU license they may not be able to grow Nvidia’s AI Cluster to a significant size, let alone establish one of the world’s largest Neoclouds with potentially hundreds of MW of AI GPU capacity. The Managing Director of YTL has come out saying that there is nothing to worry about, and that the Nvidia partnership should give them preferential treatment, but we would like to await further clarity.
The other part of their business, colocation, is also facing a significant hit to its growth potential, with the ByteDance-Oracle complex and the broad Neocloud market likely to be restricted, limiting the pool of potential customers. They will need to focus squarely on seeking out business with US Hyperscalers instead. We show below the pipeline of one of Johor’s largest operations, GDS Holdings, and its rebranded international affiliate DayOne.
In general, for a Datacenter developer, the worst case scenario is when carrying out “speculative builds “ (i.e. without a pre-leasing commitment), and the speculatively built facility ends up only partially leased or even worse, completely empty. A 100-200MW datacenter campus requires capex in the billions of dollars, the impact to the company’s ROI and cashflow could be massive. Real estate projects are by nature highly leveraged, landing sponsors and developers in financial distress very quickly if they are wrong footed.