India’s 5G shines on speed, falls short on AI workloads

Ask an artificial intelligence (AI) assistant on your phone connected to mobile internet to read through a long PDF and write you a summary. Watch how long it takes for that document to leave your phone before any answer comes back. That upload gap, more than the download speed operators love to advertise, is where India’s 5G network is struggling most, according to a new report from connectivity intelligence provider Ookla.

In its report, titled “Beyond Download Speed: Benchmarking 5G Mobile Networks Against AI Workloads,” Ookla tested 86 operators across 22 markets against the network conditions that AI applications actually need to work well. India ranks ninth globally on headline 5G download speeds. On the two things that matter most for AI workloads, upload capacity and latency, it lands in the bottom tier.

Why upload matters
Every mobile network splits its capacity between two directions: downlink, which is data flowing to your phone, and uplink, which is data flowing from your phone back to the network. For years, networks were built around a roughly 90:10 split, 90 percent of capacity reserved for downloads and only 10 percent for uploads, because people mostly consumed content such as videos and web pages rather than sending large amounts of data out, according to the Ericsson Mobility Report of June 2026.

AI upends that math. When someone sends a prompt to a chatbot, especially one attached to a document, a photo or a long conversation, all of that content has to travel upstream before any answer comes back. Ericsson’s data shows text-based AI chat already runs closer to a 29:71 uplink-to-downlink split, meaning uploads now take up nearly three times the share they used to. Voice AI and AI agents push that further, closer to an even 50:50 split between what goes out and what comes in, and AI glasses or camera-based AI, which continuously stream what they see and hear back to the cloud, push uplink demand higher still.

India networks allocates 7.53 percent of its 5G throughput to upload, delivering a median upload speed of 15.75 Mbps, Ookla found. That falls short of the 20 Mbps target the report sets for AI modalities like augmented reality and multimodal vision, a bar only ten of the 22 markets studied managed to clear. There is a bright spot here. India’s upload share actually grew by 1.53 percentage points between 2023 and 2025, at a time when 12 of the 22 markets studied saw their upload share shrink or stay flat, Ookla noted. The improvement is real, it is just starting from a low base.

Part of the reason is technical. Much of India’s mid-band 5G spectrum uses Time Division Duplex (TDD), where uplink and downlink share the same frequency band and take turns using it. Giving uploads more room directly eats into download capacity, and every operator using the same band in a market has to coordinate its timing to avoid interference, so no single telco can simply decide to fix this alone.

Where the delays creep in
The other constraint is latency, or how quickly data makes a round trip. Ookla measured what it calls multi-server latency, the baseline responsiveness a network offers under normal conditions, and found India’s figure stands at 51.6 ms. That puts India in a group of just four markets, along with South Korea (53 ms), the US(50.5 ms) and Spain (50.2 ms), that miss the sub-50 ms mark Ookla sets as the target for text-based AI chat and AI agents. Eighteen of the 22 markets studied cleared that threshold. Ookla chose this particular metric to benchmark against AI requirements because, in its own words, “those thresholds describe what AI applications require under typical operating conditions, and multi-server latency is measured under those same conditions.”

For voice AI, which needs latency under 40 ms to sound natural rather than stilted, only 13 of the 22 markets qualify. For AR and multimodal vision, which needs latency under 10 ms, not a single market in the study clears the bar.

There is better news buried in the same report. Ookla also measured how much networks slow down under heavy load, a figure it calls the degradation ratio. India’s ratio comes in at 4.0x, among the best in the study and far better than markets like Thailand (11.4x) or Singapore (9.2x). So while India’s baseline latency is weak, its network does not fall apart badly once traffic gets heavy, which is not nothing.

The bottleneck beyond the network
There is a second delay most people never think about, the one that happens after data leaves the mobile network altogether and heads to the cloud servers where AI models actually run. This is not something telecom operators fully control, since it depends on where a cloud company’s data centres sit and how well an operator’s network connects to them, but it shapes how fast an AI reply comes back just as much as the mobile network does.

Ookla found India’s median latency to reach these cloud servers runs to 114 ms for Amazon Web Services, 109 ms for Microsoft Azure, 121 ms for Google Cloud and 158 ms for Oracle Cloud Infrastructure. Set against the rest of the world, that is a weak showing.

Regions like Europe and East Asia lead here, with South Korea reaching AWS in just 40 ms, Germany in 42 ms and the UK in 44 ms. This means users in these areas get more than double India’s headroom before an AI response even starts forming. But India also trails markets much closer to home. Singapore reaches AWS in 74 ms and Indonesia in 63 ms, both comfortably ahead of India despite facing similar distances to major cloud regions.

The gap points to something specific: India’s problem is not primarily geography, since Southeast Asian markets sitting at similar distances from cloud infrastructure do better, but the quality of interconnection between Indian networks and the hyperscalers’ edge locations.

Are network operators ready for the future?
For today’s most common AI use case, text-based chat, the answer is yes for now. Every one of the 22 markets studied, India included, meets the minimum bar for text LLMs, AI-generated video and AI agents at the median, Ookla found.

The concern is what happens next. Ericsson’s own scenario modelling projects that under a medium AI adoption scenario, additional AI traffic alone could make uplink demand three times higher by 2031 compared with 2025. Under a high-adoption scenario, that multiple rises to five times. Ookla’s report is even starker on the more demanding modalities ahead: not one of the 86 operators it studied anywhere in the world currently meets the target for multimodal AI.

Do operators need to redesign their networks
The industry itself seems to think so. At the Mobile AI Industry Summit held during MWC Shanghai in June this year, telecom executives reached what Tech Wire Asia described as an explicit consensus that improving uplink is now the single most urgent priority for mobile networks. Huawei used the event to unveil a solution it calls GigaUplink, built specifically to address this gap. In a separate keynote at the same event, reported by South Korean outlet The Elec, Huawei’s vice chairman David Wang listed “establishing sustainable and future-oriented spectrum planning and allocation” and “defining standards for AI-native core networks” among six priorities he expects to shape the industry over the next decade.

Ericsson made a similar case in a separate blog post this April, titled “Four moves for operators in the AI-native era.” It said AI traffic is “more uplink-heavy, more latency-sensitive, and more session-rich” than anything mobile networks were designed to carry, and that “best-effort” connectivity, the standard approach until now, is turning into a bottleneck on both user experience and operator revenue. Ookla’s own recommendation for operators is more concrete: deploy 5G Standalone architecture, turn on uplink carrier aggregation, and strike direct peering deals with cloud providers to cut down the latency gap.

Will India’s spectrum planning start favouring upload
There are early signs of this. TRAI released its recommendations for the country’s next spectrum auction on February 24, 2026, covering nine bands from 600 MHz to 26 GHz.

One recommendation speaks directly to the upload problem. TRAI has asked the Department of Telecommunications to explore using the 1427-1518 MHz band purely for what is called Supplementary Uplink, spectrum meant only to add upload capacity without cutting into download speeds elsewhere. It has also proposed setting aside some TDD spectrum for private and machine-to-machine networks, the kind of connections industrial AI systems depend on.

This falls short of the uplink-first approach Ookla’s report calls for. But treating upload as spectrum in its own right, rather than leftover capacity, is new. Whether it survives into the final auction terms remains to be seen. Business Standard