Latency has long been a critical networking metric. That’s because latency is essentially the reaction time of your internet connection – it’s how quickly your gadget can get a response after you've sent out a request.
A snappy, low-latency connection might not make much of a difference if you’re just browsing the web. But it might make all the difference if you’re playing cloud-based video games. Consider a server-rendered 3D experience on a VR headset running at 60 frames per second: Any perceptible latency between moving in virtual reality and the screen updating could make you nauseous (not good).
But what does all this mean in the age of AI? A laggy connection between a user and a text-based AI chatbot might not cause much trouble. However, high latency between a manufacturing robot and its AI orchestrator could have significant repercussions.
What kinds of latency will AI applications need?
So now the question becomes: What kinds of latency will AI applications need? Here a Nokia study offers some insights. The firm surveyed over 1,000 U.S. technology decision-makers across high-growth industries to understand where AI adoption stands. According to the results, 72% said their AI applications will demand latency below 30 milliseconds.
Nokia’s survey didn’t delve into the details, such as the specific types of AI applications that might require 30 ms of latency. Nor did it quiz respondents about the different types of latency they may need (more on that later). However, the study does at least set a milestone for network operators and others involved in the design and operation of the internet – and that’s useful for an industry trying to anticipate what an AI future might look like.
All this brings up the next question: Where do U.S. providers stand in relation to the 30 ms latency milestone? Ookla Speedtest data can provide some insights.
In the fourth quarter of 2025, 59.2% of Speedtest samples from all U.S. mobile network operators provided minimum latency measurements below 30 ms. The rest of the samples clocked in above 30 ms.
Among fixed network operators in the U.S., even more samples registered latency below 30 ms.
This is no real surprise. Fixed networks, such as those running through fiber, typically have lower latency than mobile networks.
What should you make of all this?
First, it’s important to define “minimum latency,” which is the metric we’re using in the above charts. Minimum latency measures the best case latency for the user at the time they decide to take a Speedtest. Across 10 pings made prior to the download speed test, the lowest ping value is determined – this represents the “minimum latency.” These pings are made possible via the 15,000 servers around the world that host Speedtest server processes.
Minimum latency is different from “multi-server latency.” Multi-server latency is the metric that Ookla typically uses when discussing latency in general. That’s because multi-server latency measures latency to multiple connections and then takes an average. This is a better representation of a customer’s experience as their connection usually pulls media and other information from a variety of locations around the internet. Poor multi-server latency can show where improvement is needed in data peering arrangements and transit.
The difference between “minimum latency” and “multi-server latency,” is significant. U.S. mobile operators notched minimum latency of 27 ms in the fourth quarter of 2025, whereas their multi-server latency was 51 ms. On the fixed side of things, U.S. operators recorded minimum latency of 12 ms in the fourth quarter of 2025 and multi-server latency 25 ms.
Other ways to measure latency
A final note here: There are plenty more ways to measure latency beyond just the “minimum” and “multi-server” variety. For example, loaded latency measures the round-trip time delay of a connection, as measured when the connection’s bandwidth is fully saturated by another flow. Meaning, it’s how a connection performs when there’s other data being uploaded and downloaded at the same time. In the fourth quarter of 2025, U.S. mobile operators registered a whopping 683 ms in a measurement of the “interquartile mean” of loaded latency tests performed to the primary server during the download portion of a Speedtest. (The interquartile mean is the average of the middle 50% of a dataset.) This loaded latency measurement helps to show how much network traffic can impact latency.
Ookla also measures CDN latency (responsiveness of your network when accessing resources hosted via Content Delivery Networks) and cloud infrastructure latency (responsiveness of a connection to popular cloud service providers like AWS, Google Cloud, and Oracle).
And don’t get me started on the difference between latency and jitter!
AI requirements
Finally, it’s worth noting that AI applications themselves may require a wide range of differing networking capabilities, including latency. For example, Cloudflare notes that AI actors like Tilly Norwood might not need snappy connections with low latency if they’re being developed for a scene in an upcoming film. But they might need much lower latency if they’re being deployed in a real-time, interactive ad campaign or an augmented reality experience. (This topic leads us into the concept of edge computing, where Tilly might “live” geographically closer to users – but that’s beyond the scope of this article.)
The bottom line here is that U.S. telecom networks, both mobile and fixed, do show progress toward the 30 ms latency milestone, especially when measured by "minimum latency." However, the specific demands of diverse AI applications – and the types of latency that might match up with those applications – remain largely unknown. Thus, a 30 ms target for AI latency is a useful milestone, but it masks a complex landscape of differing needs and measurement challenges that the industry is still figuring out.
Mike Dano is a Lead Industry Analyst on Ookla’s analyst team, covering telecommunications markets across North and South America, as well as global technology trends. Prior to joining Ookla, Mike spent 25 years as a journalist reporting on the telecom industry for publications including RCR Wireless News, Fierce Network, and Light Reading.
Opinions from industry experts, analysts or our editorial staff do not represent the opinions of Fierce Network.