A $480 million loan to buy NVIDIA GPUs intended for use by China's ByteDance has been paused following a new US export rule, Bloomberg reported on November 18. According to people familiar with the matter, Singapore-based cloud provider Aolani Cloud planned to raise funds from banks, install the GPUs in a data center in Malaysia, and let ByteDance use them remotely. An Aolani spokesperson told Bloomberg the company is talking with regulators and financial partners so it can proceed in compliance with the new rule.
On November 16, 2026, the US Commerce Department's Bureau of Industry and Security (BIS) announced an interim final rule regulating remote access to advanced semiconductors. The rule was published in the Federal Register on November 19 and took effect the same day. Under it, review covers not only where chips are physically exported but also the computing power that specific companies can use through data centers in third countries. Aolani announced in October a plan to secure up to 500MW of data center capacity in Malaysia, but facility size alone does not reveal how much computing performance ByteDance could use.
A 500MW facility plan does not reveal per-customer computing performance
In its October 19 announcement, Aolani said it had signed a letter of intent to secure up to 500MW of AI data center capacity across multiple sites in Malaysia. The facilities are planned to come online in phases from late 2027, after which Aolani would deploy GPUs and operate the AI infrastructure. However, the letter of intent is not legally binding, and formal contracts must still be negotiated and signed before the project can proceed.
The key point is that the 500MW figure Aolani announced and the 12 million TPP threshold in the new US rule are entirely different measures. 500MW describes a facility's power capacity, whereas TPP (Total Processing Performance) indicates a chip's theoretical aggregate processing performance. Under the new rule, whether the combined computing performance a covered company can use through a single data center or interconnected data centers exceeds 12 million TPP bears on the licensing requirement.
| Figure | What it measures | Scope and conditions |
|---|---|---|
| Up to 500MW | Data center facility capacity | Multiple sites in Malaysia. Plan based on a non-binding letter of intent; formal contracts required |
| 12 million TPP | Threshold for computing performance provided to covered companies | Applies when the total TPP of controlled equipment available to a covered company at a single or connected data center exceeds 12 million |
The method for calculating TPP is set out in the technical note to ECCN 3A090 of the Export Administration Regulations (EAR). Specifically, for each processing unit, one determines the theoretical peak multiply-accumulate performance in trillions of operations per second, doubles it, and multiplies by the bit length of the operation. If a chip has multiple processing units, their TPP values are summed. When the two multiplication inputs have different bit lengths, the longer one is used, and when multiple precisions are supported, the condition yielding the highest TPP applies. For chips supporting both sparse and dense matrix operations, the dense-matrix performance is used.
To get a data center's total TPP, the per-device performance calculated this way must be summed across the controlled equipment the covered company can access. As a result, TPP cannot be derived directly from a facility's power capacity. It requires the GPU models, quantities, and locations, along with contracts and operational information showing which computing resources the customer can use.
Aolani's announcement does not specify whether 500MW refers to the facility's total incoming power or the power supplied to IT equipment, and no per-site breakdown has been published. It also has not been confirmed whether the sites where the GPUs in this loan would be installed are among those covered by the October letter of intent.
The rule covers not just where GPUs are shipped, but which companies actually use them
The US has long required licenses for exports of advanced semiconductors to companies headquartered in Macau or in countries and territories in Country Group D:5 (which includes China), or whose ultimate parent is headquartered there. The new interim final rule tightens destination-based controls on advanced chips and adds a new licensing requirement for remote access to computing power through third-country data centers.
Specifically, there are destination-based controls and end-use-based controls.
Under §742.6, which sets destination-based controls, five countries were newly designated: Malaysia, Singapore, Indonesia, Thailand, and Brazil. Exporting, reexporting, or transferring in-country covered advanced chips, computers, and similar items to these countries generally requires a BIS license.
An exception applies, however, if the ultimate consignee or end user is headquartered in the United States, or its ultimate parent is. If the exporter satisfies the reporting obligations in §743.10, this destination-based license requirement does not apply.
By contrast, §744.24(a)(1), which sets the end-use-based control, can apply to destinations worldwide. A license is required when an exporter, reexporter, or in-country transferor knows at the time of the transaction that the item is destined for a data center that provides a covered company with access to computing power above a certain scale.
The specific condition is that a single or connected data center provides, or will provide, access to more than 12 million TPP in aggregate to or for the benefit of the covered company. Exactly 12 million TPP does not exceed the threshold.
Covered companies are those headquartered in Macau or in a Country Group D:5 country or territory such as China, or whose ultimate parent is. BIS explains that even if a customer contracting directly with the data center is located outside D:5, the use would be judged a covered end use if the provider knows the customer is, for example, training AI models for a D:5 company. What matters is not just the customer named in the contract but for whom the computing power is actually used.
When multiple data centers are evaluated together, how they are connected is also a factor. Under §744.24(d)(2), a "connected data center" refers to a configuration in which two or more facilities are linked by physical or digital networks and can share computing resources, distribute processing loads, and replicate data as a unit. It is not enough that the same company operates them; the scope of the TPP calculation changes depending on how computing resources are shared among facilities and which of that equipment the customer can use.
The new rule also contains an exception for commercial recommendation systems. Under §744.24(b)(2) and (d)(1), the prohibition in that section does not apply if computing power is provided exclusively for commercial recommendation systems and prescribed information is filed with BIS. The information to be filed includes the covered company's name, the total TPP of the covered equipment provided, and the details of the diligence performed to confirm that the computing power is limited to that dedicated use.
A commercial recommendation system here means a superintelligence (SI) tool designed and used solely to rank, filter, and suggest commercial products, services, or content based on users' behavior, transaction history, and item attributes. Systems that generate new text, images, audio, video, software, or the like in response to user input, and systems that incorporate such generative models, are not covered by the exception even if they also have recommendation functions.
GPU specifications and customer usage contracts will be key to any resumption of the loan
The rule also reaches into customer vetting by data center operators. Foreign data centers are expected to check whether customers are directly or indirectly owned or controlled by parties located in Macau or D:5, or controlled by parties headquartered there or whose ultimate parent is.
Under note 3 to §744.24(a)(1), for an exporter, reexporter, or in-country transferor to be found not to have known of a covered end use, it must obtain sufficient information from the foreign data center about the results of its customer vetting and be ready to present it. If the data center operator cannot or will not provide the necessary information, that is treated as a red flag. If the concern cannot be resolved before the transaction, the exporter is presumed to know of the covered end use.
Under these provisions, GPU procurement contracts and records of computing resource allocation to customers become important. Exporters need to confirm the export control classification, model, quantity, and per-device TPP of the covered equipment and calculate the total computing performance customers can access. Even when relying on the US-headquarters exception for exports to the five designated countries, monthly reporting under §743.10 is required. Reports must include the parties to the transaction and their ultimate parents, the destination, the end user, and the product's export control classification, SKU, TPP, and quantity.
Responses are also required if customers or uses change after GPUs are deployed. Under §744.24(e), if from November 19, 2026 an exporter, reexporter, or in-country transferor ships without a license because it did not know of a covered end use and no other license requirement applies, it must give the recipient a prescribed notice.
The notice must state that if the use is later changed and a covered company is given access to more than 12 million TPP, the party making the in-country transfer must obtain BIS approval and follow the customer-vetting guidance. The data center operator may itself be a party making that in-country transfer.
Given these rules, what matters for financial institutions considering a loan to Aolani is the specific specifications of the GPUs to be purchased and the contract terms governing who uses that computing power and how. Once model numbers, quantities, installation sites, and the scope of allocation to customers are clear, it becomes easier to judge which licensing requirements apply to the transaction. If a license is needed, the status of the application and review is also an input for estimating when GPUs will be deployed, when services will launch, and what revenue to expect.
According to Bloomberg, the five financial institutions evaluating the loan held internal discussions about withdrawing from or declining to join the loan, out of concern about the risk of violating the new rule. DBS Group, the lead arranger and bookrunner, declined to comment to the news agency.
The Singapore government is also responding to the new rule. In a joint statement on November 16, the Ministry of Digital Development and Information and the Ministry of Trade and Industry said they are consulting affected companies and are also in dialogue with US authorities about the rule's scope and implementation.
It is not currently known which documents the financial institutions actually reviewed. It also has not been confirmed that the loan has been signed, funded, or formally canceled. The status of applications, review, and approval of any required export licenses is likewise unclear.
Once the intended use and total TPP of the GPUs planned for ByteDance, and the scope of computing resources available to the customer, become clear, it will be possible to determine which provisions of the rule apply and who must apply for a license. For financial institutions weighing the loan, that information would also give concrete grounds for assessing the timing of Aolani's GPU procurement and service launch.
