Report: How European networks handle AI workloads
July 7, 2026
AI has changed what a good mobile network looks like, and the metric the industry has marketed for two decades — peak download speed — no longer predicts it. The networks that top the download charts are often not the ones best prepared for AI traffic. Whether an AI application feels instant or breaks depends in large part on how much a network can upload, how it holds up under load, and how consistently it reaches the cloud, and on those measures, different networks come out on top. A report from Ookla rebuilds the industry’s download-led scorecard around what AI actually asks of a network, and shows where today’s 5G mobile networks are ready and where they fall short.
AI traffic is not one thing explains Ookla. Text chat, conversational voice, multimodal and AR vision, generated video, and agentic activity each load the network differently, and most of them lean on parts of the network that download speed never tested. The change AI brings is less about raw capacity, which operators have expanded for years, than about the shape of the traffic — heavier on upload, always on, and bursty, rather than download-led and session-based.
Ookla asks whether today’s 5G mobile networks are ready for AI workloads — and finds that the answer depends on metrics that have drawn far less attention than download speed.
Using Speedtest Intelligence 5G data from 2025 across 22 markets and 86 operators in North America, Europe, Asia Pacific, the Middle East, and Latin America, it measures upload capacity, latency under load, and the quality of the path to the cloud, and shows where current 5G falls short of what AI actually demands.
Key Takeaways:
AI readiness follows a different order than download speed
The clearest finding in the data is that download speed is an unreliable guide to AI readiness. Different markets lead in terms of latency that AI applications depend on: Singapore, the UAE, Malaysia, Finland and Australia form the top tier on baseline responsiveness. India makes the point cleanly, missing the AI text latency target at 51.6 ms despite ranking ninth on download speed among the 22 markets studied. The metrics that decide AI performance — upload capacity, latency under load, and the path to the cloud — follow a different order, and the gap widens as adoption shifts toward heavier use cases like conversational voice and multimodal AI.
Networks are broadly ready for text-based AI, not for what comes next
Measured against the AI workload thresholds, today’s 5G is in reasonable shape for the traffic that dominates now and short of the bar for what is coming. Eighteen of 22 markets (see image above) meet the AI text target of under 50 ms on multi-server latency, and 13 meet the conversational voice target of under 40 ms, led by Singapore at 24.6 ms and the UAE at 31.1 ms. The four that miss the AI text target sit just outside it: South Korea at 53 ms, India at 51.6 ms, the US at 50.5 ms, and Spain at 50.2 ms. The harder ceiling is AR and multimodal vision, where no market reaches the sub-10 ms target and only Singapore clears even the looser 30 ms minimum. In aggregate, the most demanding modality remains beyond what current 5G delivers, even in the most advanced deployments.
Upload is the widest gap, and in most markets, it is growing
Mobile 5G networks were deployed around a safe assumption — people consume far more than they produce — and AI inverts it. AI text traffic already runs at roughly a 29/71 uplink-to-downlink split by volume, and conversational and agentic workloads move closer to 50/50, yet the typical operator still devotes only around 10 per cent of throughput to the uplink. In more than half of markets that share has slipped since 2023 even as absolute upload speeds rose. Indonesia leads the dataset on upload share at 23.9 per cent but recorded the largest decline, while Germany is the exception, the only market to raise its share, up 2.4 percentage points on the back of targeted spectrum, 5G SA, and carrier aggregation. The spread in absolute upload speed is wider still: e& UAE leads the entire dataset at 57.50 Mbps median upload, more than four times any US carrier, while the US sits at the bottom on allocation at 5.1 per cent, where T-Mobile’s 13.94 Mbps tops a tightly clustered field of US carriers. South Korea shows the limit of a single-band strategy, reaching the second-highest market upload at 45.27 Mbps on C-band TDD while holding one of the lowest upload shares at 7.5 per cent.
Latency holds up under normal conditions and breaks under load
Most networks clear the AI baseline at rest. Under full utilisation the picture changes, and unevenly. Degradation ratios, the multiple by which loaded latency exceeds the baseline, stretch from 3.7x in the UK to 11.4x in Thailand, where median loaded latency reaches 960.3 ms. The degradation ratio alone can mislead: Singapore posts the lowest baseline in the set at 24.6 ms but one of the highest degradation ratios at 9.2x, while the UAE pairs a mid-range ratio with the lowest median loaded latency of any market at 288.4 ms — the absolute value, not the ratio, is what shapes AI responsiveness on a congested cell. The spread inside a market can rival the spread between markets. In the UK, EE holds a best-case loaded latency of 119 ms while O2 records 305 ms on the same ground.
The path to the cloud has become its own constraint
Upload and latency stop at the network edge. The rest of the AI journey runs from there to the cloud where the model executes, and that segment is increasingly decisive. Across much of Asia Pacific, the most consequential choice for an AI deployment is not which operator to use but which cloud to reach: in Australia, the gap between the fastest and slowest cloud provider inside the same market is 96.6 ms, enough to push voice and agentic applications past perceptible delay. Europe shows the opposite, with Germany reaching AWS at 42.2 ms and just 2.7 ms separating its fastest and slowest cloud provider. Brazil stands apart entirely, with median cloud latency of 149.7 to 163.6 ms across all four providers, tied to infrastructure concentrated in São Paulo and thin direct peering. Jitter adds the final twist. At the median, markets look alike; at the 90th percentile, worst-case jitter runs three to six times higher, and the steadiest connections, in South Korea, Norway, and Singapore, are not the fastest, while the Philippines and Malaysia swing widest. Speed and stability are separate properties of the path to the cloud, and the markets best placed for real-time AI are the ones that hold their timing steady.
Other posts by :