Megawatts measure power, not who can use the GPUs. At Malaysia’s Project Mercury, the 60 MW first phase of IT capacity was fully contracted to BytePlus, but the Asian Infrastructure Investment Bank’s 24 April 2026 disclosure does not identify the GPUs or their downstream users.
A new US Bureau of Industry and Security (BIS) rule, published and effective on 19 November 2026, makes that distinction matter. Buying capacity, receiving equipment and running a computing service are separate milestones.
Key facts
- The US rule adds a licence requirement for specified items sent to five countries, including Malaysia, subject to applicable exceptions and exclusions. A separate end-use test applies worldwide.
- Project Mercury’s campus is designed for 120 MW of IT capacity in total. Only the first 60 MW phase is identified in the disclosure as fully contracted to BytePlus.
- The disclosure does not establish which GPUs are installed, who can access them, whether the service is running, or whether the rule applies to Mercury.
Start with the building
Megawatts measure electrical power, not computing performance. They do not show which processors are installed, how many there are or how capacity is split between customers. A campus power figure can also differ from the power available to IT equipment.
The Mercury disclosure names its first-phase customer and describes the planned development. It does not identify downstream users or workloads, and an April document cannot show that the service is running in November.
Readers should keep three columns: contracted, delivered and running. An announced investment goes in none of them automatically. A signed agreement is stronger evidence, but it still needs the right label.
A reading map, not a confirmed chain for either example: equipment → data centre → cloud provider or reseller → final user.
Each arrow needs its own evidence. The company paying for capacity may not run the workload, and a provider’s registered address may not be where its equipment sits.
Read who is responsible for what
A separate agreement dated 15 July 2026, filed with the US Securities and Exchange Commission (SEC), shows the next layer. It sets terms for Dimension AI, a Singapore-incorporated company, to supply compute services to Exascale Labs.
Sections 7.3–7.4 require Exascale’s downstream agreements to cover customer identity, location, intended use, screening and suspension for unlawful use. Exascale stays responsible to Dimension even if it resells capacity. Sections 10.3–10.4 cover escalation and record-keeping.
Section 5.3 gives the customer a proportional refund for the unexpired term if a regulatory suspension lasts more than 30 days. It does not promise payment within 30 days.
Schedule 1 specifies B200-class or better GPUs and sets performance commitments, but leaves the facility location to be supplied on delivery.
The agreement predates the November rule. It shows how risk was allocated, not that delivery, a suspension or a later amendment took place.
Update, 20 November 2026: Exascale Labs’ quarterly report (Form 10-Q, filed with the SEC on 16 November 2026) gives the agreement’s current status. The filing puts the capacity at 3,931 PFLOPS over a three-year term, with a one-time fee of $71,430,000 payable within 30 days after acceptance and no advance payment. As of the date of the report, Exascale had not accepted any capacity, the services had not commenced and no payment obligation had arisen. On 13 November 2026, the two companies confirmed to each other that the services will not commence before 30 November 2026, that the agreement will terminate automatically on that date in accordance with its terms, that no service fee or other amount has been paid or become payable, and that after termination neither party will have any outstanding obligations to, or claims against, the other.
That is the commercial reading. Who must collect information from the next customer? Who can act when it changes? Who bears the cost while access is interrupted?
A refund clause allocates one kind of loss. It does not show a loss occurred or that every downstream customer has the same protection. One contract cannot tell us how common these terms are.
Keep two regulatory questions separate
In its 16 November 2026 announcement, BIS said the rule was intended to prevent existing restrictions from being circumvented through remote computing at data centres in third countries. The announcement itself did not put the rule in force that day.
The destination question. Section 742.6 requires a licence to export, re-export or transfer specified advanced computing items to or within Indonesia, Malaysia, Singapore, Thailand or Brazil, subject to applicable exceptions and exclusions.
An exclusion applies when the ultimate consignee or end user, or its ultimate parent, is headquartered in the US and the exporter, re-exporter or transferor meets the reporting requirements in §743.10. An American investor alone does not establish that exclusion.
The end-use question. Section 744.24 adds a second test, which applies anywhere in the world. It covers exports, re-exports and in-country transfers of specified controlled items where the required knowledge exists that a data centre provides, or will provide, access to more than 12 million aggregate TPP.
The test applies when that access goes to or for customers headquartered in Macau or a Country Group D:5 destination, or whose ultimate parent is headquartered there. The calculation can include linked data centres.
TPP (total processing performance) is a regulatory measure of chip performance, not a power figure. So it cannot be read off a megawatt number.
The rule does not make every large cloud service need a licence: the item, transaction, customer, knowledge conditions and exclusions all matter. Nor does staying below the threshold remove separate destination or other export-control requirements.
Section 744.24 also addresses later changes in end use and gives data centres providing specified services before publication 90 days from the effective date to obtain a licence if they wish to continue those services. The rule excludes access used exclusively for qualifying commercial recommender systems if the required information is submitted. It does not exempt every workload on a platform.
Five questions to take away
- What equipment is involved? Look for models, export-control classifications and the performance calculation. Do not convert MW into GPUs.
- Who is in the customer chain? Separate the contracting customer, reseller, ultimate parent and actual user.
- What can that customer access? One address may not describe linked facilities or the capacity available through them.
- What is the workload? Establish its purpose and whether the contract lets it change.
- What proves the stage? Distinguish permissions, contracts, acceptance, delivery and operation.
For each answer, record either “documented” or “not established by this source”.
Gaps cut both ways. Missing public information does not mean a company withheld it from a regulator, and a compliance clause does not prove every check was carried out.
Key questions
Does a 60 MW deal mean the GPUs are running?
No. MW measures power. The AIIB disclosure identifies contracted IT capacity, not installed GPUs, downstream users or a running service.
When did the new US chip rule take effect?
November 19, 2026, the day it appeared in the Federal Register.
Does every large cloud deal now need a US license?
No. Destination and end use are separate tests. The items and transaction, the knowledge requirement, the customer or its ultimate parent’s headquarters, the computing power the customer can access, and applicable exceptions and exclusions all matter.
