top of page
Image by Akif Waseem

APAC TMT Notes

A collection of articles sharing operator-side perspective on telecommunications, digital infrastructure, governance, and technology across APAC covering key topics in investments, market dynamics, emerging technologies, and organisational development.

Export Controls Can Stop GPUs. Can They Stop Compute?

Writer: Kaye Hau
Kaye Hau
Nov 13
6 min read

Updated: 3 days ago


Processor_chip
Credit: Canva

The AI race is increasingly becoming a race for access to compute.


The canonical logic is simple: control the flow of physical hardware, and you can limit access to the compute power required to train advanced AI models. However, while governments have responded by imposing restrictions on the export and resale of advanced GPUs and AI accelerators, GPU infrastructure is no longer only sold as hardware.


Through GPU-as-a-Service (GPUaaS), customers can rent access to GPU clusters without owning the chips, operate the hosting data centre, or physically moving the equipment.


The hardware sits in one country, owned and financed by another party, operated by a cloud provider, and accessed remotely by a customer somewhere else.


And that opens up a larger regulatory question:


If the GPU remains in an approved country while the compute is accessed remotely by a customer elsewhere, where has this technology flowed to?

From Owning GPUs to Accessing Compute

The business model of separating hardware ownership from access to compute is one of the oldest playbooks in the market, and foundational to the economics of a cloud service provider. A cloud / neocloud provider procures servers, deploy them in leased or partner data centres and sells virtualised or dedicated compute to customers. The customer pays for capacity without making the upfront capex investments while avoiding the operational risks normally associated with hardware ownership.


This is commercially efficient as well since customers can scale more quickly, while GPU neocloud providers gain a way to monetise and maximise the use of the capacity of expensive hardware across multiple users.


But that also changes the nature of the infrastructure asset.


The value is no longer in the GPU alone. It sits in the combination of GPU, data-centre location, power, cooling, networking, software layer, customer contracts, and the right to access the compute.

For investors, this changes the underwriting question. It's no longer enough to just count GPUs, megawatts, or contracted capacity. The more important question is whether the provider has durable and compliant access to compute, and whether that access can still be monetised if export controls, customer restrictions or geopolitical conditions change.


Megaspeed is a useful case study of this business model, apart from the current controversy surrounding the company and its employees.


Megaspeed describes itself as a specialised cloud provider offering NVIDIA GPU-accelerated virtual servers, managed Kubernetes containers, and customised compute capacity. This is straight out of a cloud infrastructure playbook. However, it makes the geography of AI infrastructure more complicated given the added layer of export controls.


The GPU may be deployed and hosted in Malaysia while the cloud service is sold through a Singaporean company. The customer may be headquartered elsewhere and the workload is accessed remotely. And the financing may involve an entirely different set of counterparties.


From an infrastructure perspective, this is a normal multi-party operating model driven by commercials and market dynamics.


From an export-control perspective though, it creates a chain of ownership, access and control that is much harder to manage and control.

The Regulatory Gap Is Becoming a Geopolitical Contest

The debate has moved from whether a chip has physically moved to whether a customer in a restricted jurisdiction can access the compute remotely through a cloud provider or overseas data centre.


US policymakers and researchers have identified this challenge for several years now. Traditional export controls are designed around physical goods, while cloud services transcend physical borders, providing access to computing resources without necessarily transferring ownership of the underlying chip.


The US Bureau of Industry and Security has since issued guidance covering advanced computing integrated circuits, infrastructure-as-a-service (IaaS) providers, end-user certification and potential diversion risks. Its latest interim final rule goes further by targeting overseas data centres that provide significant volumes of compute to companies headquartered in China or other Country Group D:5 jurisdictions, including companies whose ultimate parent is headquartered there. It also requires greater due diligence on direct and indirect ownership, control and remote end users.


The rule also introduces destination-based licensing requirements for certain advanced semiconductors exported, re-exported or transferred to Malaysia, Singapore, Thailand, Indonesia and Brazil. This is not a blanket prohibition on data-centre investment in these markets. It creates an additional licensing and compliance layer around the hardware, the operator and the customer.


The direction is clear from a policy point of view. Regulators are beginning to look beyond where a shipment is headed, and towards how compute is accessed, used and transferred.

This is a familiar pattern. Regulation is usually written around the technology and business models that exist at the point of enactment. Markets then evolve around those rules. Physical infrastructure becomes a service, ownership becomes access, and domestic operations become cross-border platforms.


Historically, the market moves much faster than the regulatory perimeter, and will likely continue to do so in this case. The question for investors and operators then is whether they can identify the next control point and adapt before it is formally regulated.


Two Different Perspectives

The policy response is not coming only from Washington.


While the US is trying to control where these advanced chips and compute can flow, China is trying to control how much foreign hardware enters its market and accelerate the adoption of domestic alternatives.


This is becoming part of a strategic-autonomy policy aimed at reducing reliance on foreign chips. Beijing is trying to ensure that its AI ecosystem is not beholden to overseas suppliers, technologies or export-control decisions. Nvidia has disclosed that sales of H200 products to China were restricted by the Chinese government, limiting the amount it could actually ship.


In international-relations terms, AI infrastructure is becoming a contest over strategic chokepoints. The US has concentrated leverage over parts of the semiconductor and technology stack, while China is trying to reduce its exposure by building domestic alternatives and controlling the supply of foreign hardware.


These chokepoints extend beyond GPUs and include HBM memory, advanced packaging, semiconductor equipment, cloud platforms, data-centre capacity, power, networking and software ecosystems. Whoever controls these points can influence who gets access to compute, where it can be deployed and how it can be used.


The notion of “good enough” for scale is not about matching the absolute peak performance of foreign chips, but having sufficient performance, availability and cost efficiency to build a self-sustaining ecosystem.

This is where the latest rules matter for Malaysia and Singapore. Their role is no longer assessed only by the physical location of the data centre. The ownership structure, ultimate parent, customer base and remote end users are becoming equally important.


This brings us to the investor question. If access to compute is becoming more closely regulated, what is the investor actually underwriting?


What This Means for Investors

The value of a GPUaaS platform is dependent on whether it has durable, compliant and monetisable access to compute.


That depends on where the hardware is deployed, who owns and operates it, who controls the cloud platform, who the end customers are, and whether the provider can effectively perform KYC, monitor or withdraw access when the risk profile shifts.


For investors, customer concentration now carries a regulatory dimension. A Malaysian data centre serving US or allied customers may have a different risk profile from one whose capacity is contracted to a Chinese-headquartered group, or accessed through a regional intermediary.


The same analysis applies to the capacity that has not yet been contracted. A data centre may be built with only 50% of its capacity backed by customer offtake, leaving the the remaining 50% to be filled after construction. While the identity and quality of the contracted customers matter, the scrutiny on the uncontracted capacity manifests as well due to the uncertainty of the customer profile amid a shrinking potential customer pool.


Investors need to ask who is expected to take up the balance, whether these customers are from the same risk category, and whether the provider can onboard them under the same regulatory and customer-due-diligence requirements.


However, the existence of these cases doesn't mean that the broader demand for compute is illegitimate. There's real demand from hyperscalers, enterprises, AI developers and digital service providers that need access to GPU capacity in the region.


As the rules tighten, the market may even begin to correct itself. Transparent operators with diversified customer bases, clear ownership structures and credible compliance systems are well positioned to retain access to capital, hardware and customers. Whereas gray or opaque arrangements face higher financing costs, delayed deployment or pressure to restructure their customer base.

The infrastructure, operator, customer and financing parties may all sit in different jurisdictions, creating a different form of infrastructure risk for investors. The asset sits in the right to access and monetise compute across a network of locations, counterparties and regulatory regimes.


The right might be valuable, but it's conditional. Its value can be impaired by changes in export controls, customer restrictions, data-centre access, ownership structures or the ability to transfer the resulting model outputs.


For investors, the key question is how resilient that access remains when the regulatory perimeter shifts.


Access Is the New AI Currency

The strategic question has moved beyond the physical location of the hardware. It now concerns who controls the compute, who can access it, what workload is being run, where the resulting model can be deployed and whether that access can be verified and withdrawn.


The AI infrastructure race is creating a new regulatory layer between semiconductors, cloud services, data centres, trade controls and national security.


While export controls can restrict the movement of GPUs, their effectiveness will depend on how well the broader chain of access is governed given that compute and model outputs can cross borders digitally.


The GPU may be the scarce physical asset, but access to compute is becoming the new AI currency.

Getting Started

Reach out for a confidential initial conversation on how we can help.

Every engagement begins with a complimentary discovery session to understand your objectives, challenges and priorities. If we're not the right fit, we'll tell you honestly and, where possible, point you in the right direction.

bottom of page