Not all internet is equal: Choosing the Right APAC Internet Underlay for AI
Enterprises are upgrading their networks to deal with the widescale rollout of AI, which has put the quality of internet underlay in the spotlight. This task can be challenging in the APAC region, where consumer-focused ISPs and a complex regulatory landscape can impact enterprise application performance.
As AI moves from pilot to production, companies may need to upgrade their networks to ensure they have the necessary bandwidth, peering, and latency to support demanding applications, such as agentic AI or AI-enabled features of enterprise applications like Salesforce or HubSpot.
According to exclusive new research from Omdia, over two-thirds of current AI users expect to quadruple their AI bandwidth within the next 18 months. These bandwidth upgrades are being implemented in both backbone and enterprise networks. For the latter, companies are acquiring links of up to 1 Gbps into their offices and other facilities, compared to legacy connections of around 50 Mbps.
Choosing the more appropriate networks
This rapid scaling of enterprise networks has been made possible by the increasing use of internet as an underlay, which gives companies access to bandwidth more easily and cheaply compared to traditional MPLS networks. However, not all internet is the same, and with the increasing use of AI-enabled applications, enterprises need to choose their underlay networks carefully to assure the end-user experience.
While overlay networks based on SD-WAN can effectively knit together internet networks, they can’t overcome performance issues created by unsuitable internet links. In particular, critical sites running bespoke AI deployments, such as software development or production optimisation, typically require carrier-grade international connectivity to ensure optimal performance. The situation for branch offices is more nuanced, and in many cases, multinationals can utilise locally sourced internet, particularly in regions where local peering with cloud services is available, thereby avoiding the need for international traffic.
Global variation in internet services
Choosing networks in APAC can be challenging because of a complex digital regulatory landscape. In fact, the 22 APAC governments have adopted over 1,500 digital laws and regulations in the last two decades, with more than 300 currently in consultation. When building their APAC underlay networks, enterprises need to pay particular attention to regulations governing cross-border data transfer, cybersecurity and AI to ensure compliance.
For example, some countries, such as Vietnam and Indonesia, have limited in-country cloud peering, which makes international internet bandwidth essential for enterprises wanting high bandwidth access to enterprise cloud services. To achieve this, many enterprises route traffic via regional connectivity hubs, such as Tokyo, Hong Kong and Singapore.
In China, three main carriers dominate connectivity. Direct routes into China are expensive and the default routing often goes via the US, even for trac from Hong Kong to mainland China. Enterprises will be better o‑ checking for indirect paths and optimising accordingly.
Business focus required
In addition, because APAC ISPs are focused mainly on the consumer market, enterprise trac can also experience higher latency and jitter. This is because routing decisions prioritise consumer content delivery rather than business-critical trac. In fact, ISPs often peer with major content providers but not necessarily with business-focused cloud services such as Security Service Edge (SSE) or Microsoft Azure.
Enterprises should look for enterprise-focused internet services offering high reliability and performance for business-critical trac. Consumer-focused networks rarely implement quality of service (QoS) or trac engineering for enterprise-grade requirements. This will affect the performance of enterprise-focused AI-enabled and AI applications along with VoIP, video conferencing, and real-time collaboration tools.
It is also important to build redundancy into the network by using multiple ISPs and diverse paths to avoid consumer-centric bottlenecks. Enterprises should regularly test latency and paths to detect inefficient routes. In some cases, private connectivity options, such as MPLS, dedicated links, or cloud interconnects, are a better choice for critical workloads.
In fact, according to Omdia research, 60% of enterprises plan to rebalance their connectivity from the public internet to private connectivity. This is because uncertain performance is unacceptable for some digital workloads, prompting enterprises to reinvest in dedicated connectivity.
Questions of data sovereignty
Another issue that enterprises need to consider in their network choice is sovereignty. As Omdia points out in its recent research, “The need for sovereignty applies not only when data is stored or processed. Sovereignty also matters the moment data moves. Routing policy determines where data moves – and on the public internet, this can be difficult to control.”
Data sovereignty is of particular concern for sectors such as banking and defense, where strict regulations govern where data can transit and even which network technologies are used. Enterprises in these sectors must ensure that their service providers deliver physical route disclosure in the form of a Keyhole Markup Language (KMZ) file and provide verifiable compliance guarantees.
Preparing your network for the AI-first age
Choosing the most appropriate underlay for your network to support the AI-first age is not just about nding the cheapest networks. In the APAC region, in particular, consumer-focused ISPs with limited business cloud peering and international routes can cause issues with AI-enabled and business applications, such as video conferencing.
Omdia recommends that enterprises look for service providers that apply application classification rules to traffic routing to secure the best performance. For example, can they send high-sensitivity inference trac directly to the cloud using private connections, while other applications use the public internet?
In many cases, the best way to mitigate performance issues is to work with a regional Tier 1 provider, such as Telstra International, which can provide global reach, shorter paths, richer peering options, and single-hop direct peering with other service providers, hyperscalers, content delivery networks (CDNs) and trac exchanges. In addition, Telstra International owns and operates subsea and backbone infrastructure, delivering greater flexibility and control to optimise IP and Internet networks. This enables superior resilience and performance, setting us apart from providers that rely solely on third-party networks.
To find out more about Telstra International’s network offers in APAC and globally, please visit our adaptive networks page.