Spread the love

Post Views: 4

Ookla’s global benchmark finds that tomorrow’s mobile AI will depend less on flashy download figures and more on upload strength, latency under pressure and a clean route to the cloud.

For years, mobile operators have treated download speed as the grand final of network performance. The biggest number won the headlines, the advertising campaign and, with luck, the customer. It was simple, easily understood and wonderfully suited to glossy charts.

Artificial intelligence has now wandered into the field and changed the rules.

A new Ookla Research report, Beyond Download Speed: Benchmarking 5G Mobile Networks Against AI Workloads, argues that a fast download result is no longer a reliable measure of whether a 5G network can handle the next generation of AI services.

The study draws on 2025 Speedtest Intelligence data from 22 markets and 86 operators across North America, Europe, Asia Pacific, the Middle East and Latin America. It examines upload capacity, baseline latency, performance under load and the quality of the connection between mobile networks and cloud infrastructure.

Its conclusion is uncomfortable but useful: networks built mainly to deliver content downstream are being asked to support applications that constantly send information upstream, react in real time and reach cloud-based AI models without stumbling.

In other words, the old speed-test trophy cabinet may need dusting rather than expanding.

AI traffic changes the network equation

AI is not a single, tidy category of traffic. A text prompt, a spoken conversation, an augmented-reality image and an autonomous software agent place very different demands on a mobile connection.

Text-based AI is comparatively forgiving. A prompt travels to a cloud server, the model processes it, and the answer returns in bursts. Voice assistants require a quicker and steadier exchange. Multimodal applications must upload images, video or sensor data while maintaining an immediate response. Agentic AI can keep communicating in the background while it completes a chain of tasks.

That makes upload capacity, latency, loaded performance, cloud routing and jitter far more important than they were in the streaming era.

Download speed still matters, of course. Nobody is proposing a return to the cheerful patience of dial-up internet. It simply no longer tells the whole story.

Ookla’s benchmark clearly shows the new order. Eighteen of the 22 markets studied met the target of less than 50 milliseconds for text-based AI. Thirteen met the tougher conversational voice target of less than 40 milliseconds.

Singapore led at 24.6 milliseconds, followed by the United Arab Emirates at 31.1 milliseconds.

Four technologically advanced markets narrowly missed the text-AI target: Spain at 50.2 milliseconds, the United States at 50.5 milliseconds, India at 51.6 milliseconds and South Korea at 53 milliseconds.

India’s result is particularly revealing because it ranked ninth for 5G download speed among the markets examined. A strong download ranking, it seems, can arrive wearing a very convincing disguise.

Today’s chatbots pass; tomorrow’s vision systems do not

For the AI services most people use today, the findings are broadly reassuring. Text assistants should perform adequately across most measured markets under ordinary network conditions.

The trouble begins when the workload becomes more immediate, visual and interactive.

Ookla sets a target below 10 milliseconds for augmented-reality and multimodal vision applications, with less than 30 milliseconds treated as a minimum threshold.

No market reached the ideal target. Singapore was the only one to clear the broader minimum.

That gap matters because future travel, aviation, hospitality and mobility services are likely to use live visual AI rather than text alone.

A traveller may point a phone at a railway sign for instant interpretation. An airport navigation assistant may analyse its surroundings and guide a passenger towards a gate. A wearable device may continuously interpret objects, directions and location data.

These experiences cannot afford a long pause while the network gathers its thoughts.

The commercial lesson is plain. A network can feel perfectly competent when answering a typed question, yet become decidedly less clever when asked to see, listen and respond at once.

Upload is 5G’s in the luggage compartment

The widest structural weakness is the upload capacity.

Traditional mobile networks were designed around a sensible assumption: consumers download far more than they upload. Video streaming, web browsing and social feeds reinforced that model.

AI begins to overturn it.

Ookla estimates that text-based large language model traffic has an uplink-to-downlink split of roughly 29:71. Conversational voice and agentic AI traffic is closer to an even 50-50 split.

Yet many 5G networks still devote only a modest share of their throughput to uploads.

Indonesia recorded the highest upload share in the study at 23.9 per cent. Australia stood at 7.3 per cent, while the United States sat at the bottom of the dataset at 5.1 per cent.

Absolute upload speed can soften the effect of a low share, but it does not remove the strategic problem.

Only 10 of the 22 markets and fewer than half of the 86 individual operators assessed met the 20 Mbps upload benchmark for augmented-reality and multimodal AI services.

There were notable bright spots.

The UAE’s e& recorded a median upload speed of 57.50 Mbps, more than four times the result of any American carrier. South Korea posted a strong market median of 45.27 Mbps.

Germany was the only market to increase its upload share between 2023 and 2025, helped by targeted spectrum allocation, standalone 5G architecture and carrier aggregation.

Those results offer operators a useful reminder: the uplink is no longer the back door of the network. For many AI services, it is the primary entry point.

Congestion turns milliseconds into an eternity

A network’s best behaviour is not its only behaviour.

Under ordinary conditions, most markets delivered acceptable baseline latency. Under full use, performance deteriorated sharply and unevenly.

Ookla’s degradation ratios ranged from 3.7 times normal latency in the United Kingdom to 11.4 times in Thailand, where median loaded latency reached 960.3 milliseconds.

That is almost a full second before allowing for the AI model’s own processing time.

For a person waiting on a typed response, it is irritating. For a live voice exchange, a visual assistant or an automated safety system, it can destroy the experience.

The ratio itself does not tell the entire story.

Singapore had the fastest baseline latency but a comparatively high degradation ratio of 9.2x under load. The UAE recorded the lowest median loaded latency in the study, at 288.4 milliseconds, despite a middle-ranking degradation ratio.

Operator differences within a single country could also be substantial.

In the United Kingdom, EE recorded a loaded latency of 119 milliseconds, while O2 reached 305 milliseconds. The postcode and provider may therefore matter nearly as much as the flag on the national scoreboard.

For travel companies, this is more than a technical footnote.

Airports, convention centres, stadiums, cruise terminals and major tourism precincts create exactly the dense, high-demand conditions in which an otherwise capable network can lose its composure.

The cloud route can make or break the experience

The mobile network is only the first leg of an AI request.

Once data reaches the operator’s network edge, it must travel to the cloud infrastructure where the model usually runs. Ookla found that cloud-provider choice can produce striking differences within the same country.

In Australia, the gap between the fastest and slowest measured cloud providers was 96.6 milliseconds. The study recorded 69.3 milliseconds to Amazon Web Services and 165.9 milliseconds to Oracle Cloud Infrastructure.

That gap alone can push a voice or agentic AI service from responsive to noticeably delayed.

Germany offered the opposite picture. Its latency to AWS was 42.2 milliseconds, with only 2.7 milliseconds separating the fastest and slowest cloud options.

Brazil faced a much harder road, with latency ranging from 149.7 to 163.6 milliseconds across the four major cloud providers assessed.

Jitter, or variation in timing between consecutive measurements, adds another complication.

Median results appeared fairly similar, but the worst 10 per cent of connections were three to six times less stable in many markets. South Korea, Norway and Singapore were among the steadiest. The Philippines and Malaysia showed some of the widest swings.

Speed and stability, then, are separate virtues. A quick network that periodically hesitates can be more disruptive than a slightly slower connection that behaves predictably.

Australia is fast enough to compete, but not to relax

Australia’s baseline multi-server latency of 33.7 milliseconds placed it among the leading markets and comfortably inside Ookla’s targets for text-based and conversational voice AI.

That is encouraging.

However, Australia’s 7.3 per cent upload share, 6.8-times degradation ratio under load and 96.6-millisecond spread between cloud providers show that the national picture is not uniformly polished.

The implications extend well beyond telecommunications.

Airlines are exploring AI-supported operations and customer service. Hotels are introducing automated concierge systems. Travel advisers are using AI to construct and amend itineraries. Airports, attractions and transport operators increasingly rely on real-time mobile interactions.

These services will be judged by travellers in the simplest possible way: did they work immediately, or did they not?

An application may be intelligently designed, beautifully branded and backed by an expensive model. If the network cannot upload the input, control latency under pressure or reach the cloud efficiently, the result will still feel second-rate.

Customers seldom blame the routing architecture. They blame the app, the airline, the hotel, or whichever organisation had the misfortune of putting its logo on the screen.

A new scorecard for the AI era

Ookla’s findings point towards four broad investment priorities: stronger uplink capacity, lower latency under load, wider deployment of standalone 5G and advanced radio features, and better peering and cloud connectivity.

Operators will also need to benchmark their networks against real workloads rather than one heroic speed figure.

That means testing voice exchanges, image uploads, continuous agent activity and congested-cell conditions. It also means examining the entire journey from the handset to the cloud, rather than stopping at the mobile network’s edge.

The report should be read as a benchmark rather than a guarantee.

It uses median Speedtest Intelligence results collected during 2025. Individual experiences will vary by location, handset, spectrum, cell load, operator, cloud region, and application design. AI models also introduce their own processing delays, which fall outside the network measurements.

Even with those qualifications, the direction is unmistakable.

Download speed remains useful, but it has been demoted from judge to jury member.

The next phase of 5G competition will be won by networks that can send as confidently as they receive, remain calm when crowds arrive and reach the right cloud without taking the scenic route.

For an industry fond of boasting about speed, that is a timely reminder that the fastest headline is not always the smartest network.

Read the official Ookla research article and download the full Ookla research report.