TECH

Ookla: AI creates new paradigms for 5G quality; speed no longer tells the whole story about the network

Ookla states that artificial intelligence has profoundly changed the definition of a good mobile network, rendering the traditional download speed metric insufficient for predicting real-world performance. According to the report, networks topping download speed rankings are not necessarily the best prepared for AI traffic, as the experience of these applications depends primarily on upload capacity, network behavior under load, and the consistency of the path to the cloud.

The document redefines the industry benchmark based on what AI actually demands from 5G networks, assessing where they are ready and where they still fall short.

Ookla explains that AI traffic is not homogeneous; text, conversational voice, multimodal systems, augmented reality (AR) vision, generated video, and agent activity all place different demands on the network—often in areas that download speed metrics have never captured. According to Ookla, the shift driven by AI relates less to raw capacity and more to the new traffic profile.

This traffic is upload-heavy, continuous, and prone to spikes, rather than being download-centric and session-based. The report also seeks to determine whether current 5G networks are prepared for these workloads, concluding that the answer depends on metrics that have historically received little attention.

Based on Speedtest Intelligence data collected in 2025—covering 22 markets and 86 operators—the study measures upload capacity, latency under load, and cloud-path quality, identifying where current 5G falls short of AI requirements. The key takeaway is that download speed is an unreliable indicator of technology readiness.

Markets such as Singapore, the United Arab Emirates, Malaysia, Finland, and Australia lead in baseline latency, despite not necessarily being the fastest in terms of download speeds, the document notes. The case of India appears to illustrate this divergence, as it misses the latency target for text-based AI despite ranking ninth in download speed among the markets studied.

The report concludes that networks are generally prepared for text-based AI but not for more demanding workloads. Eighteen of the 22 markets studied meet the latency target for text-based AI, and 13 meet the target for conversational voice, with Singapore and the United Arab Emirates leading the way.

The four markets that miss the text target are only slightly above the threshold. However, no market hits the latency target for augmented reality and multimodal vision. Only Singapore meets the more lenient 30 ms minimum, demonstrating that the most demanding modalities remain beyond the reach of current 5G technology.

Upload gaps for AI... The biggest gap identified in the study concerns upload speeds. 5G networks were built on the assumption that users consume more data than they produce, but AI flips this logic, Ookla explains. Text traffic currently operates with a split of approximately 29% upload and 71% download, whereas the workloads of conversational systems and AI agents approach a 50/50 split. Despite this, operators continue to dedicate only about 10% of capacity to uploads.

In more than half of the markets, this proportion has decreased since 2023, even amidst absolute increases in upload speeds. The report highlights that Indonesia leads in upload share, although it also recorded the largest drop.

Conversely, Germany is the only market to increase this share, thanks to targeted investments in the sector. In terms of absolute speed, the United Arab Emirates leads with 57.50 Mbps—more than four times the speed of any U.S. operator—while South Korea demonstrates the limitations of a single-band strategy. Latency remains stable under normal conditions but degrades significantly under load, with marked variations across markets. Degradation rates range from 3.7 times in the UK to 11.4 times in Thailand, where loaded latency reaches 960.3 ms.

The company warns that this metric can be misleading: Singapore has the lowest baseline latency but one of the highest degradation rates, whereas the United Arab Emirates has the lowest loaded latency—a figure more relevant to the speeds required for AI. Variation within each market is also significant, as illustrated by the UK, where different carriers exhibit widely differing latencies.

The path data takes to reach the cloud emerges as another critical factor. While upload and baseline latency end at the network edge, the remainder of the journey to the server where the model runs proves decisive. In markets like Australia, the gap between the fastest and slowest cloud providers reaches 96.6 ms—enough to compromise voice applications and AI agents.

"In Europe, the landscape is more uniform, with minimal differences between cloud providers; however, Brazil exhibits high and similar latencies across clouds due to infrastructure concentration and limited peering," states Ookla.

The report adds that jitter—the variation or fluctuation in data packet delivery delay across a network—is equally crucial. Although markets may appear similar in terms of median performance, significant differences emerge at the 90th percentile: South Korea, Norway, and Singapore demonstrate the most stable connections, whereas the Philippines and Malaysia show greater variability.

Ookla concludes that speed and stability are distinct attributes, and that the markets best positioned for real-time AI are those that maintain consistent timing.

mundophone