How AI workloads will test APAC mobile networks

• APAC mobile networks support text AI, but heavier workloads expose upload and latency gaps.
• Singapore leads on latency, Malaysia is consistent, but cloud routing and jitter remain issues.

AI workloads are changing how mobile network performance is measured, according to Ookla’s latest research on 5G network readiness. The report assesses 22 markets, including nine in Asia Pacific: South Korea, Japan, India, Indonesia, Australia, Singapore, Malaysia, Thailand, and the Philippines.

Ookla’s framework looks beyond download speed and coverage. It measures upload capacity, multi-server latency, loaded latency, cloud infrastructure latency, and cloud jitter across text-based large language models, conversational voice AI, multimodal and AR vision, AI-generated video, and agentic AI.

The report said most 5G markets can support today’s text-based AI workloads under normal conditions. Ookla said the gaps are wider for workloads that require higher upload capacity, lower latency, or more stable cloud connectivity.

Upload capacity becomes a key test

Ookla identifies upload capacity as one of the main constraints. AI workloads send more traffic upstream than traditional mobile use cases, while most deployed 5G networks were designed around downlink-heavy demand.

Ookla estimates that text LLM traffic runs at about a 29/71 uplink-to-downlink split. Conversational voice AI and agentic AI are closer to 50/50, while AR and multimodal vision can push uplink share toward 40%.

In APAC, Indonesia has the highest 5G upload share in Ookla’s dataset at 23.9%, with median upload speed of 26.38 Mbps. Thailand follows at 14.2%, while Malaysia records 11.6%, Japan 10.5%, the Philippines 8.8%, Singapore 7.9%, India 7.5%, South Korea 7.5%, and Australia 7.3%.

South Korea records one of the strongest absolute upload results, with median upload speed of 45.27 Mbps. Malaysia also performs well on uploadwith median upload speed of 34.78 Mbps.

Singapore records 30.25 Mbps, while Thailand and Indonesia sit above the 20 Mbps target Ookla uses for AR and multimodal AI. Ookla benchmarks workloads against absolute upload speed rather than upload share.

Ookla notes that a market with low upload share can still meet upload targets if total throughput is high. Fewer than half of the operators in the full 86-operator dataset meet the 20 Mbps upload target for AR and multimodal AI.

The same report said all markets meet the minimum upload requirement for text LLMs, voice AI, AI-generated video, and agentic AI at the median. Ookla also compares operators against combined upload and latency requirements.

Latency separates APAC markets

On latency, Singapore records the lowest result in Ooklas dataset. Its multi-server latency stands at 24.6 ms, making it the only market Ookla measured that meets the minimum threshold of under 30 ms for AR and multimodal vision.

Malaysia follows at 33.0 ms, ahead of Australia at 33.7 ms, Thailand at 37.7 ms, and Indonesia at 38.4 ms. The Philippines records 40.2 ms, Japan 46.3 ms, India 51.6 ms, and South Korea 53.0 ms.

Ookla’s threshold table places text LLM and agentic AI latency targets at under 50 ms. Conversational voice AI has a lower target of under 40 ms, while AR and multimodal vision require under 10 ms as a target and under 30 ms as a minimum.

In Malaysia, all six operators meet both the text LLM and voice AI latency targets. Singapore’s three major operators sit in the lowest-latency zone and also meet both targets.

In South Korea, operators deliver some of the highest upload speeds in the dataset outside the UAE, but two of the three sit above the 50 ms text LLM latency threshold, according to Ookla.

Network performance changes under load. Ookla uses loaded latency to measure latency when a connection is fully utilised. Its degradation ratio compares loaded latency with normal multi-server latency.

Indonesia records one of the lowest degradation ratios in the dataset at 3.7x, while India records 4.0x. The Philippines sits at 5.2x, Japan at 5.3x, South Korea at 5.5x, Australia at 6.8x, Malaysia at 7.9x, Singapore at 9.2x, and Thailand at 11.4x.

Singapore has the lowest normal latency in the dataset, but its degradation ratio rises to 9.2x under load. Ookla attributed the increase to dense urban demand competing for cell resources during peak periods.

Thailand records the highest degradation ratio in the dataset at 11.4x, with median loaded latency of 960.3 ms. Ookla said the result points to backhaul and cell-load management constraints rather than coverage or spectrum gaps.

Cloud routing adds another layer

Ookla also measures cloud infrastructure latency and jitter for AI workloads. These metrics cover the path from the operator network edge to cloud endpoints, where inference often runs, rather than the mobile network alone.

Across APAC, cloud-provider choice changes latency by tens of milliseconds in several markets. Australia records 69.3 ms latency to AWS but 165.9 ms to Oracle Cloud Infrastructure, a 96.6 ms gap within one market.

In Southeast Asia, OCI is the furthest away in latency terms across Indonesia, Malaysia, Thailand, and the Philippines when compared with AWS, Azure, and Google Cloud.

Malaysia records cloud latency of 77 ms to AWS, 86 ms to Azure, 86 ms to Google Cloud, and 111 ms to OCI. The Philippines records 104 ms to AWS, 106 ms to Azure, 88 ms to Google Cloud, and 133 ms to OCI.

Singapore’s cloud latency is lower than most of Southeast Asia but still varies by provider. Ookla records 74 ms to AWS, 75 ms to Azure, 73 ms to Google Cloud, and 96 ms to OCI.

East Asia has lower cloud latency than Southeast Asia in the dataset. South Korea records 40 ms to AWS, 39 ms to Azure, 48 ms to Google Cloud, and 43 ms to OCI, while Japan records 46 ms, 46 ms, 62 ms, and 54 ms, respectively.

Ookla also measures cloud infrastructure jitter for real-time AI workloads. It defines jitter as the average difference between consecutive cloud latency measurements, with higher values indicating less stable timing.

The Philippines records the highest 90th percentile cloud jitter in the dataset at 34.9 ms. Malaysia follows at 33.1 ms, while Indonesia records 27.9 ms, India 25.7 ms, Thailand 24.3 ms, Australia 24.2 ms, Japan 19.7 ms, Singapore 16.3 ms, and South Korea 13.4 ms.

Median jitter ranges from 3.9 ms in Singapore to 7.0 ms in Indonesia. Ookla said the separation appears at the 90th percentile, where worst-case jitter can run three to six times the median in most markets.

Ookla points to three network investment areas: rebalancing links toward upload, reducing latency, and treating cloud peering as network infrastructure. It also points to 5G Standalone, carrier aggregation, better peering, and edge inference as relevant technologies for mobile AI workloads.

TDD spectrum remains a practical constraint. In TDD, uplink and downlink share the same frequency band through different time slots, so increasing uplink allocation reduces downlink capacity unless operators coordinate changes across the market.

FDD spectrum avoids that trade-off by using separate frequency bands for uplink and downlink. Ookla said markets that combine TDD mid-band with FDD low-band, including Australia, tend to show more consistent uplink performance.

Ookla said multimodal AI workloads can place sustained demand on uplink capacity and cloud routing. These use cases can combine voice, images, video, and sensor data in the same session, making both mobile network performance and the cloud path relevant to the user experience.

Want to experience the full spectrum of enterprise technology innovation? Join TechEx in Amsterdam, California, and London. Covering AI, Big Data, Cyber Security, IoT, Digital Transformation, Intelligent Automation, Edge Computing, and Data Centres, TechEx brings together global leaders to share real-world use cases and in-depth insights. Click here for more information.