• Existing 5G networks can support text-centric AI but face latency challenges
• Augmented reality (AR), multimodal AI and physical AI use cases require better upload speeds and improved latency, according to Ookla
• Of the 22 markets analysed, Ookla found only Singapore met the minimum latency requirement to support AR 
• Major markets including India, South Korea, Spain and the US currently fail to meet target latency thresholds to consistently support text-based generative AI 

Most 5G networks are powerful enough to support text-centric AI, but performance benchmarks fall significantly short of the latencies required for emerging AI use cases, according to analysis from Ookla.

Text-based large language models (LLMs) require around 50 milliseconds (ms), which according to benchmarking research undertaken by Ookla, is achieved in 18 out of 22 markets it assessed, while 13 markets currently achieve the latencies required for conversational voice (under 40ms). 

However, future AI modalities, such as augmented reality (AR) vision, will need a minimum multiserver latency of less than 30ms, while the ultimate target multi-server latency to support such services will be sub-10ms – which isn’t available in any market at this time. Ookla notes that “minimum thresholds represent the level below which users are likely to notice degraded quality” while “target thresholds represent where consistently good performance becomes achievable”.

In its most recent report, Ookla found that the dominant AI traffic on mobile networks today is text-based LLMs, which generate a two-phase traffic pattern. Firstly, a user prompt travels upstream to a cloud inference server, then the response streams back token by token in irregular bursts determined by the server-side compute load, which can cause delays, similar to messaging.

Latency was also found to generally hold up under normal conditions but degrades sharply under load, and in an uneven way. Degradation ratios run from 3.7x to 11.4x across markets, and the gap between operators inside a single market is as wide as the gap between markets.

In the report, Ookla also argued that benchmarking needs to move beyond traditional measures. In judging AI readiness, Ookla said that performance indicators, such as upload capacity, latency and path to the cloud, are favoured over indicators such as download speeds, something for which the Accenture-owned company is well known (and trusted) for benchmarking.

The geographic breakdown shows that of the 22 markets benchmarked, only Singapore meets the minimum multiserver latency required for future modalities, such as AR vision, at 24.6ms (though that is still some way from achieving the target threshold of sub-10ms). 

At the other end of the scale, South Korea, India, the US and Spain all fail to achieve the target latency required for text-based LLMs of 50ms.

In its analysis, Ookla claims that several factors contribute to this gap. This includes the fact that time division duplex (TDD) mid-band spectrum delivers high throughputs but also adds round-trip overhead, leaving a device waiting for its uplink transmission slot before responding. Non-standalone architecture (NSA) is also having an impact because it may route through a 4G rather than a 5G core. As an example of this, South Korea, which deploys 5G almost entirely on C-band TDD and is predominantly NSA, records the second-highest download speeds but also the highest multiserver latency at 53ms.

Ookla used its Speedtest Intelligence data to measure performance across 86 operators in North America, Europe, Asia Pacific, the Middle East and Latin America. It found that the proportion of network capacity allocated to the uplink – calculated as median upload speed divided by the sum of median upload and download speeds – is not only small but shrinking in many markets. According to Ookla, fewer than half of the operators measured meet the 20Mbit/s upload speeds required for AR and multimodal AI.

It also warned that while latency holds up under normal conditions, it degrades sharply and unevenly under heavier loads. Finally, cloud latency and worst-case jitter vary as much as the operator network does, and within one market (Australia) the choice of cloud provider can swing latency by nearly 100ms, enough to decide whether real-time AI is viable. This is enough to push voice and agentic applications past perceptible delays, Ookla added.

Europe leads the dataset for low cloud infrastructure latency. This gives users in these markets the most latency headroom for inference, leaving more room for AI model processing before the delay degrades the user experience.  

In conclusion, Ookla says that changing patterns in AI workloads and the shift to multimodal interaction will require upgrades to 5G infrastructure but will also provide a use case for 5G standalone and network slicing.

The report states: “As these workloads move from early adoption to mainstream use, the ability to guarantee consistent latency for specific traffic classes will become a competitive differentiator for operators, and a practical requirement for enterprise deployments where service level commitments apply.”

Physical AI, such as robots and industrial automation, will represent “the most network-demanding wave” of AI.

“The operators investing in upload rebalancing, latency reduction and cloud peering optimisation today are laying the foundation for these workloads. The performance gaps this report identifies are already wide enough to shape competitive positioning – and they will widen as the AI modality mix shifts,” Ookla concluded.

- James Pearce, Contributing Editor, TelecomTV