Ookla Benchmarks 22 Countries for 5G Mobile Broadband in AI Workloads

Ookla (Accenture), which operates the popular broadband Speedtest.net benchmarking service, has today published a new report that evaluates metrics like 5G upload capacity, latency under load, and cloud infrastructure pathways across 22 countries in order to identify how well they’ll handle AI workloads.

The first thing to understand here is that AI readiness tenders to favour specific kinds of network performance, such as upload capacity, latency under load, and the path to the cloud. As such the leaderboard for AI is going to be a bit different from those that would otherwise normally be dominated by download performance, and the gap widens as adoption shifts toward heavier use cases like conversational voice and multimodal AI.

Overall, the report finds that existing 5G infrastructure generally supports text-centric AI, yet often falls short of the performance required for emerging modalities. For example, while latency (server response time) targets for text LLMs / Large Language Models (under 50 milliseconds) are achieved in 18 of 22 markets and conversational voice targets (under 40ms) in 13, no market currently reaches the sub-10ms requirement for AR (Alternative Reality).

Latency was also found to generally hold up under normal conditions, but degrades sharply under load, and unevenly. 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.

Upload allocation also turned out to have the widest and most persistent gap. The typical operator devotes only around 10% of throughput to the uplink, while fewer than half meet the 20Mbps target for AR and multimodal AI, and upload share has declined or held flat in 12 of 22 markets since 2023.

Finally, cloud latency and worst-case jitter vary as much as the operator network does, and within one market the choice of cloud provider can swing latency by nearly 100ms, enough to decide whether real-time AI is viable. The results for just the UK are mixed, which leaves plenty of room for improvement.

Ookla’s UK Findings on 5G for AI Performance

Baseline AI Readiness: The UK records a multi-server latency of 46.4ms. This means the market meets the target for text-based large language models (under 50ms), but misses the target for conversational voice AI (under 40ms).

Resilience Under Stress: The UK shines when the network is fully utilized, tying with Indonesia for the lowest latency degradation ratio in the dataset at just 3.7x.

The Operator Divide: Loaded latency varies sharply within the UK market, a 2.6x gap between the best and worst-performing operators sharing the same market.

Upload Capacity: The UK allocates 9.18% of its 5G throughput to upload, delivering a median absolute upload speed of 10.96Mbps. Furthermore, the UK’s upload share has contracted slightly by 0.26 percentage points between 2023 and 2025.

Path to the Cloud: Europe leads the dataset for low cloud infrastructure latency. The UK records strong routing times across the major hyperscalers: reaching AWS at 44.0ms, Azure at 48ms, Oracle Cloud Infrastructure at 48ms, and Google Cloud at 49ms.

We’ll publish a link to the full report once it’s available later this morning, although there’s clearly some work to do in a few areas of the UK’s 5G performance for AI purposes. The rapid roll-out of 5G Standalone (5G+) networks may well help matters, although adoption is currently still quite low and so that may take a bit of time to feed through to studies like this that use crowdsourced data.