India’s mobile networks have made significant progress in 5G performance, but they remain short of the latency and cloud infrastructure benchmarks needed to support the next generation of artificial intelligence (AI) applications, according to a new study by network intelligence firm Ookla.
The report argues that as AI shifts from text-based chatbots to voice assistants, autonomous agents and multimodal applications, traditional measures such as download speeds will become less important than latency, upload capacity and cloud connectivity. It also concludes that no mobile operator among the 86 studied across 22 markets is currently prepared for the most demanding AI workloads expected in the coming years.
For India, the findings present a mixed picture. While the country ranks ninth in median 5G download speeds among the markets analysed, it is one of only four markets that fail to meet the sub-50 millisecond latency threshold identified by Ookla for responsive text-based large language model (LLM) applications. India recorded a median multi-server latency of 51.6 milliseconds, marginally above the recommended benchmark.
According to Ookla, latency rather than bandwidth is increasingly becoming the defining factor for AI performance.
“The AI workloads dominating mobile networks today sit below the most demanding AI performance thresholds. That will not last,” the report said, pointing to the rapid emergence of multimodal AI applications that combine text, voice, images and video in a single interaction.
Unlike video streaming, AI workloads require continuous two-way communication between devices and cloud infrastructure, making network responsiveness far more critical.
Despite missing the baseline latency target, India’s network performs comparatively well under heavy traffic conditions.
Ookla measured a loaded latency degradation ratio of 4.0 times, indicating that network latency increases relatively modestly when connections are fully utilised. That places India ahead of several advanced markets where network performance deteriorates sharply during peak demand.
Singapore, for example, records one of the world’s lowest baseline latencies but experiences a degradation ratio of 9.2 times, while Thailand records the highest degradation ratio at 11.4 times.
Realistic assessment of AI readiness
The report notes that loaded latency provides a more realistic assessment of AI readiness because enterprise AI applications often operate continuously rather than intermittently.
Another area requiring improvement is upload capacity. India currently allocates 7.53% of total 5G throughput to uploads, delivering a median upload speed of 15.75 Mbps. While this remains below the 20 Mbps benchmark Ookla considers necessary for advanced AI applications, the country recorded the second-highest growth in upload allocation among all markets studied, increasing by 1.53 percentage points between 2023 and 2025.
According to the report, future AI applications—including autonomous agents, real-time translation, robotics and physical AI—will generate substantially more upstream traffic than today’s text-based services, making upload performance an increasingly strategic metric for operators.
The study also suggests that network quality alone will no longer determine AI performance. Latency between telecom networks and hyperscale cloud providers is emerging as an equally important factor because most AI inference is performed inside cloud data centres rather than on devices.
For users in India, median cloud latency ranges from 108 milliseconds on Microsoft Azure to 158 milliseconds on Oracle Cloud Infrastructure, with Amazon Web Services and Google Cloud positioned in between. The nearly 50-millisecond difference between providers can materially affect the responsiveness of conversational AI and enterprise AI applications, the report said.
“The most consequential infrastructure decision for AI deployment is not which mobile operator to use, but which cloud provider to use,” the report observed, noting that cloud routing and peering arrangements increasingly influence user experience.
Connection stability presents another challenge. India recorded a median cloud jitter of 6.7 milliseconds, but worst-case jitter reached 25.7 milliseconds at the 90th percentile. According to Ookla, such fluctuations can disrupt conversational AI sessions even when average latency appears acceptable because voice and autonomous AI applications require highly consistent response times.
The report argues that the telecom industry’s traditional focus on download speeds no longer reflects the requirements of AI-driven services.
Instead, operators will increasingly need to invest in standalone 5G (5G SA), fibre backhaul, uplink optimisation, direct cloud peering and network slicing to support emerging AI workloads.
It also points to AI-RAN, where radio infrastructure simultaneously supports mobile connectivity and edge AI inference, as a longer-term architectural shift already attracting investment from companies including Nokia, NVIDIA, Ericsson and T-Mobile.
Among the report’s most striking findings is that none of the 86 operators assessed worldwide currently meets the 10-millisecond latency target required for future multimodal AI applications.
While 65 operators satisfy latency requirements for text-based LLMs and 46 meet the threshold for conversational voice AI, the study concludes that the industry’s next challenge lies in preparing networks for AI systems that continuously exchange voice, video and sensor data.
For India, where telecom operators have invested heavily in nationwide 5G rollouts over the past three years, the report suggests that the next phase of competition may depend less on delivering higher download speeds and more on building network infrastructure capable of supporting AI-native applications and enterprise workloads.
