• South-east Asia telcos are going all in on GPU-as-a-service as they scramble to support the region’s AI transformation
• Telekom Malaysia is the latest to launch such a service

Telekom Malaysia is the latest prominent South-east Asia telco to set itself up to play a major role in the region’s fast-developing AI services landscape by launching a GPU-as-a-service (GPUaaS) offering to its B2B customer base. 

Telekom Malaysia (TM) recently announced that it’s ramping up its digital infrastructure to enable an offering for its government and enterprise customers.

By launching graphics processing unit (GPU)-as-a-service, telcos such as Telekom Malaysia are able to offer business users access to AI computing resources without the attendant up-front capital costs and set-up time usually involved in building the capability in-house. The technology enables telcos to offer capacity for large language model (LLM) training and generative AI (GenAI) inference for their major customers’ AI applications, a capability that’s quickly becoming a global ‘must have’ for corporate transformation.

TM’s offering is based on Nvidia’s GPU technology at the telco’s Uptime Tier-III certified datacentres, providing, it claims, low-latency, real-time, high-performance AI computing while ensuring data sovereignty and security.

According to Amar Huzaimi Md Deris, TM’s group CEO, the TM GPUaaS investment is designed to support Malaysia’s AI aspirations. “By offering cost-efficient access to high-performance computing for AI model training and inferencing, we enable businesses and the government sector to adopt AI-driven solutions with greater ease, ensuring both operational efficiency and scalability,” he claimed, pointing out that the TM infrastructure is capable of sustainably hosting GPU technology from other chip vendors in the future. 

The CEO added: “Our GPUaaS design is modular and scalable, catering to various user needs. Leveraging this agility, we are pleased to announce that TM has secured an international customer for GPUaaS requirements to be delivered by year-end. This initiative underscores TM’s commitment to empower industries and unlock new opportunities through AI infrastructure, as we align our Digital Powerhouse 2030 aspiration to fast forward the future of a Digital Malaysia,” he added.

The past year has seen several of the largest South-east Asia telcos throw their hats into the GPUaaS ring, all offering similar rationales and strategies, albeit with key differences in approach.

In March, Indonesia’s Indosat Ooredoo Hutchison (IOH) announced a hook-up with Nvidia for GPUs to fuel its transition from ‘telco to techco’ and to help spur Indonesia’s AI transformation. IOH is Indonesia’s second largest operator with nearly 100 million mobile customers and a healthy 10% revenue growth, but it senses the opportunity to use its momentum to play a strategic role in Indonesia’s AI development with its GPUaaS offering along with the development of a sovereign AI cloud architecture for Indonesia.

IOH CEO Vikram Sinha noted in a recent earnings announcement that “with the adoption of the latest technology, supported by collaboration with various parties based on the spirit of mutual cooperation, we will continue our infinite journey in realising Indosat’s grand goal of connecting and empowering all Indonesian society by accelerating the nation’s digital transformation.”

Regional progress

At its recent AI Summit 2024, South Korea’s SK Telecom laid out its AI plans. It intends to go regional, first building out a national network of AI datacentres from which it will launch a GPUaaS offering for enterprise and government users before expanding its reach into other markets in South-east Asia and beyond, using high-capacity submarine network links in collaboration with partners.

Its first AI datacentre testbed in Pangyo,  south of Seoul, is to be driven by SK group  technologies harnessed with GPU virtualisation solutions, AI energy optimisation tools and SK Hynix’s HBM (high-bandwidth memory) chips – as well as products from key partners, such as Nvidia. The testbed will also use three types of liquid-cooling solutions, namely direct liquid cooling, immersion cooling, and precision liquid cooling. 

Meanwhile, Singtel, which outlined an ambitious GPUaaS strategy in early 2024, and partnered with regional telcos AIS, Maxis and Telkomsel, has beefed up its regional GPUaaS credentials by agreeing a partnership with GMI Cloud, a GPUaaS specialist that is already offering services from datacentres in the Asia Pacific region. It has also indicated its GPUaas global ambitions by announcing a collaboration with Nscale to use the AI cloud infrastructure specialist’s AMD and Nvidia GPU capacity in Europe in exchange for Nscale being able to able to tap into Singtel’s Nvidia H100 Tensor Core GPU capacity in the South-east Asia region for its customers’ workloads through an integration with Singtel’s orchestration platform.

For “as a service” options to show their worth, they will require a degree of cloud openness, but whether that openness should be mandated by regulatory guidance is still an issue being worked through in many jurisdictions. The adoption of GPUaaS in Asia implies that customers, at least at this point, are not too concerned about application lock-in or onerous data exit charges – both inhibitors for “as a service” options. Already, the evolving structure of the cloud business in South-east Asia and the partner relationships involved look to be on course to create a lively, competitive environment where data and applications can move about reasonably freely, enabling customers to chase down the provider best suited to the components of their applications, even though any GPUs serving a customer in one country may be physically located in another. However data sovereignty issues may prevent this in many instances, especially where government data is involved.

For instance, Malaysian telco Maxis is to use Paragon, Singtel’s 5G edge and cloud resource management platform, to offer its customers a range of services, including access to low-latency GPUaaS and storage, according to Singtel.

So the Paragon platform will be hosted on local server infrastructure in Malaysia to meet the data sovereignty and cybersecurity needs of Malaysian businesses.