TL;DR
· The United States has brought remote computing power uses at some overseas data centers under advanced chip license review and added destination license requirements for controlled chips shipped to Malaysia, Singapore, Indonesia, Thailand, and Brazil.
· Goldman Sachs tracks six major data center hubs in Southeast Asia, which are expected to have 4.4 gigawatts (GW) of capacity come online by the end of 2026. Based on an average utilization rate of 70% to 80%, actual usage demand is estimated at about 3.1 to 3.5 GW.
· Some large-scale model training projects in Malaysia have relatively concentrated customers. Once tenants delay expansion or reduce loads, operators may need to re-lease space, and reserved power may sit idle.
· Goldman Sachs expects that the six hubs will have 8.4 GW of capacity online by the end of 2028 under the base case. If only projects explicitly named in the research report are included, the online capacity is about 7.1 GW.
· If announced U.S. data center projects are slow to break ground, some cloud providers' incremental demand may shift to Southeast Asia after 2028.
GPUs are already installed in data centers in Malaysia, but whether they can continue to provide computing power to overseas customers may still depend on U.S. license approval. On November 16, the U.S. Department of Commerce's Bureau of Industry and Security (BIS) published an interim final rule that brings who overseas data centers provide computing power to, and how much computing power they provide, under advanced chip license review. For Southeast Asian operators, existing clusters need to confirm whether they can continue serving their original customers, while new projects depend on whether equipment can arrive on schedule and whether data center space can be leased out in time.
According to a pre-publication document in the Federal Register, the interim final rule titled "Addressing Remote Access to Advanced Semiconductors" is scheduled to be published on November 19 and take effect upon publication, with a 60-day comment period.
The newly added Section 744.24 addresses end uses at overseas data centers. If an exporter, re-exporter, or transferor knows that controlled advanced chips or servers will be used to provide computing power to enterprises within the scope specified by the rule, and the Total Processing Performance (TPP) exceeds 12 million, the relevant transaction must obtain a license. The rule's scope of application is determined by the location of the enterprise's headquarters or the headquarters of its ultimate parent company. The computing power threshold can be aggregated across digitally or physically interconnected facilities, and indirectly provided computing power may also be counted. Relevant enterprises must also verify customers' direct and indirect ownership or control relationships.
Overseas data centers that have already provided relevant computing power above the threshold and wish to continue service must, in principle, obtain licenses within 90 days after the rules take effect. Commercially recommended systems that meet strict definitions may qualify for exceptions. Generative systems, even if they also have recommendation functions, must still be assessed under the relevant rules to determine whether a license is required.
The new rules also add destination licensing requirements, covering specific advanced chips exported, re-exported, or transferred locally to Malaysia, Singapore, Indonesia, Thailand, and Brazil. Eligible U.S.-headquartered companies may qualify for relevant exceptions and must fulfill monthly reporting obligations.
Goldman Sachs expects that by the end of 2026, the operational capacity of six Southeast Asian data center hubs—Johor, Bangkok, Jakarta, Singapore, Kuala Lumpur, and Batam—will reach 4.4 gigawatts. Based on an estimated average utilization rate of 70% to 80%, actual usage demand will be approximately 3.1 to 3.5 gigawatts.
Goldman Sachs believes that some large-scale model training projects in Malaysia serving overseas customers are more vulnerable to the new rules. Such projects have relatively concentrated customers. If tenants delay expansion or reduce computing power demand, operators may need to re-lease, and completed facilities and reserved power may sit idle.
Operators with more diversified customer structures have more potential sources of demand, including major U.S. cloud providers, sovereign AI projects promoted by local governments, approved enterprise customers, and local model inference and recommendation businesses. When new tenants deploy equipment and how much capacity they need will determine how quickly idle data centers can be put back into use.
In Goldman Sachs' base case, operational capacity at the six hubs will increase from 4.4 gigawatts at the end of 2026 to 8.4 gigawatts at the end of 2028, and reach 13 gigawatts in 2030. If only projects explicitly named in the research report are included, operational capacity in 2028 will be approximately 7.1 gigawatts, 1.3 gigawatts lower than the base case. This gap reflects the sensitivity of forecasts to project implementation progress.
U.S. projects under construction provide some support for its data center capacity growth through 2028. How much capacity can be added in 2029 to 2030 depends more on whether announced projects can begin construction. Goldman Sachs believes that if U.S. project construction progress is slow, some cloud providers may shift new demand to Southeast Asia, and the actual scale of implementation will still depend on local regulation and construction conditions.
Welcome to join the official BlockBeats community:
Telegram Subscription Group: https://t.me/theblockbeats
Telegram Discussion Group: https://t.me/BlockBeats_App
Official Twitter Account: https://twitter.com/BlockBeatsAsia