Network testing and analytics firm Ookla has released a new study indicating that while today's 5G networks can support certain generative AI services, they remain insufficient for emerging applications such as augmented reality (AR), multimodal AI, and Physical AI. Key limitations—including network latency, uplink bandwidth, and performance under heavy traffic—suggest that global 5G infrastructure is still some distance away from fully meeting the demands of the AI era. According to Ookla, most existing 5G networks are already capable of supporting text-based large language models (LLMs) and conversational AI services. Text-based LLM applications typically require network latency below 50 milliseconds (ms), a benchmark already achieved in 18 of the 22 markets evaluated. Meanwhile, conversational voice AI generally requires latency below 40 ms, with 13 markets meeting this threshold.
However, significant gaps remain for more demanding, real-time AI applications. Ookla noted that AR and multimodal AI require end-to-end latency below 30 ms to deliver an acceptable user experience, while the long-term target should be under 10 ms. Among the 22 markets surveyed, Singapore was the only one to meet the 30 ms benchmark, recording an average latency of 24.6 ms. No market has yet achieved the 10 ms target.
The report also highlighted that several major markets have yet to meet even the baseline latency requirements for text-based LLMs. The United States, India, South Korea, and Spain currently struggle to consistently achieve the 50 ms threshold. South Korea, for example, ranks second globally in 5G download speeds, yet its continued reliance on 5G Non-Standalone (NSA) architecture and C-band time-division duplex (TDD) spectrum results in higher latency because uplink transmissions must wait for allocated time slots. As a result, multi-server latency testing recorded an average of 53 ms, demonstrating that high download speeds do not necessarily translate into low-latency performance.
Beyond latency, Ookla identified uplink bandwidth as an increasingly critical indicator of AI network readiness—potentially even more important than download speed. Fewer than half of the mobile operators evaluated worldwide currently provide the 20 Mbit/s uplink throughput required for AR and multimodal AI applications. Network congestion further compounds the challenge, with latency increasing by 3.7 to 11.4 times under heavy traffic conditions. In addition, connectivity differences among cloud service providers can significantly affect application performance; in Australia, for example, latency between different cloud platforms can vary by nearly 100 ms.
Looking ahead, Ookla expects Physical AI applications—including robotics, smart manufacturing, and industrial automation—to place the most stringent demands on mobile networks. To prepare for the AI era, the firm recommends that telecom operators accelerate the deployment of 5G Standalone (SA) networks, network slicing, uplink capacity optimization, and improvements to cloud connectivity architecture. These upgrades will be essential to delivering the low latency, high reliability, and real-time communication capabilities required by next-generation AI services.