Ookla: How Well Do 5G Networks Support AI Workloads?
While the mobile industry has been traditionally preoccupied with download speeds, the parameters that matter most to AI performance are the upload capacity, latency under load and the quality of the connection path to cloud servers, according to Ookla.
The network analytics firm’s latest analysis across 22 5G markets has found that while every market meets the minimum median upload speed requirement for text LLM, voice AI, AI-generated video, and agentic AI services, only 10 reach the 20Mbps target for Augmented Reality (AR) and multi-modal AI.
According to Ookla, today's 5G infrastructure can handle the latency demands of today’s AI applications – text LLM interaction, AI-generated video, and agentic AI – but falls short for emerging applications.
The firm says that 18 of the 22 markets meet the AI text target of under 50ms on multi-server latency, while 13 meet the conversational voice target of under 40ms. Singapore leads at 24.6ms, followed by the UAE at 31.1ms.
Four markets narrowly miss the AI text threshold: South Korea at 53.0ms, India at 51.6ms, the US at 50.5ms and Spain at 50.2ms. However, augmented reality and multimodal vision applications present far more challenging requirements. No market reaches the sub-10 millisecond target for AR and multi-modal version, and only Singapore clears the looser 30ms minimum.
Those in the 40-80ms range can support Voice AI under normal conditions, but with what the reports describes as “limited headroom”.
“Once cloud transit, inference processing, and server queuing are added to the network’s latency contribution, the cumulative delay in these markets approaches the boundary where users perceive degradation,” the report states.
Cloud connectivity emerges as a constraint
The connection path from network edge to cloud infrastructure has become a critical bottleneck for AI performance.
In Australia, the gap between the fastest and slowest cloud provider within the same market reaches 96.6ms, sufficient to push voice and agentic applications past perceptible delay thresholds.
Brazil faces particular challenges, with median cloud latency of 149.7 to 163.6ms across all four providers, linked to infrastructure concentrated in São Paulo and limited direct peering arrangements.
Upload capacity presents the widest gap
The report identifies upload capacity as the most significant weakness in current 5G deployments.
Mobile networks were built on the assumption that users consume far more data than they produce, but AI reverses this pattern.
AI text traffic already operates at roughly a 29 to 71 split between upload and download by volume, while conversational and agentic workloads approach a 50-50 split.
Despite this shift, typical operators still devote only around 10% of throughput to upload capacity. In more than half of markets, this share has declined or stayed constant since 2023 even as absolute upload speeds increased.
Indonesia leads the dataset on upload share at 23.9% but recorded the largest decline. Germany stands as the sole exception, increasing its share by 2.4 percentage points through targeted spectrum allocation, 5G standalone architecture and carrier aggregation.
The gap in absolute upload speeds varies dramatically. e& UAE leads at 57.50Mbps median upload speed, more than four times faster than any US carrier. The US sits at the bottom for upload allocation at 5.1%, with T-Mobile out in front with 13.94Mbps.
Performance degrades under network congestion
To further explore this, Ookla has calculated degradation ratios for these markets. These measure how much latency increases under full utilisation and range from 3.7 times baseline in the United Kingdom to 11.4 times in Thailand, where median loaded latency reaches 960.3ms.
The UAE pairs a mid-range degradation ratio with the lowest median loaded latency of any market at 288.4ms. Within individual markets, variation between operators can rival differences between countries. In the UK, EE maintains a loaded latency of 119ms while O2 records 305ms.
