For two decades, the mobile industry has marketed one number above all others: peak download speed. A new report from Ookla, the company behind Speedtest, argues that number no longer tells you whether a network is any good — and the United States is a clear example of the mismatch.
American carriers dedicate a smaller share of their networks to uploads than any other country Ookla studied, and U.S. text-based AI performance misses the report’s own readiness target, even as some of the same networks post solid download numbers. The bigger story, Ookla argues, is that download speed no longer predicts whether a network is ready for the AI tools people are actually starting to use — and by that measure, the U.S. has real work to do.
A Different Kind of Traffic
AI doesn’t use a network the way video or web browsing does. Ookla’s report, “Beyond Download Speed: Benchmarking 5G Mobile Networks Against AI Workloads,” draws on 2025 Speedtest Intelligence data covering 86 carriers across 22 countries. Instead of download speed, it measures upload capacity, how much a connection slows under real load, and how reliably it reaches the cloud — the three things that actually decide whether an AI app feels instant or breaks. Text chat, real-time voice, camera-based AI, generated video, and autonomous AI agents each stress a network differently, and most of those demands fall on parts of the network download testing never checked. Carriers haven’t run out of raw capacity, per the report; the traffic itself has changed shape, growing heavier on upload, constantly connected, and bursty.
Where the U.S. Actually Lands
The U.S. misses Ookla’s latency target for basic text-based AI, coming in at 50.5 milliseconds against a 50-millisecond bar — one of only four countries in the study to fall short. It also sits at the very bottom of all 22 countries in the share of network capacity devoted to uploads, at just 5.1 percent, though T-Mobile’s 13.94 Mbps median upload leads a tightly bunched field of U.S. carriers. For comparison, the UAE’s e& carrier posts a median upload speed more than four times faster than any American carrier. Under real-world network load, the three major U.S. carriers all showed meaningful slowdowns as well, in line with a broader pattern in the report: most networks perform fine at rest, and the real test is what happens once they’re busy.
Where Other Countries Stand
Singapore, the UAE, Malaysia, Finland, and Australia lead the pack on the baseline responsiveness AI depends on. India is the report’s clearest illustration of the broader mismatch: it ranks ninth of 22 countries on plain download speed, yet also misses the text-AI latency bar. Across the full study, 18 of 22 countries meet the target for text-based AI and 13 meet the tougher bar for real-time voice AI, led by Singapore and the UAE. Augmented reality and camera-based AI remains out of reach everywhere — no country meets that target, and only Singapore clears even a looser minimum.
What Ookla Says Needs to Change
Ookla’s central recommendation is that carriers start treating the path to the cloud as core network infrastructure, not an afterthought. Beyond that, the report points to wider deployment of standalone 5G, upload-focused carrier aggregation, and direct connections between carrier networks and major cloud providers to shorten and stabilize that final leg of the trip.
Ookla’s full report includes operator-by-operator benchmarks and a complete list of investment priorities for carriers, and is available as a free download directly from Ookla.
