Timestamps Make Up AI Network Metrics, But What About the Clock Underneath Them?

Every AI network metric is a difference between clocks in different places, and the GNSS signal that those clocks depend on is the easiest thing in the chain to overpower.

What you’ll learn:

  • Why every AI network metric reduces to a subtraction between clocks that must agree.
  • How much timing margin 5G already spends, and how little 6G positioning leaves.
  • Why a spoofed site is more dangerous than a jammed one, and what to test instead.

In network-metrics parlance, time to first token (TTFT) is the elapsed time from when a client sends a prompt request to the arrival of the first output token. TTFT gets filed under application performance. Underneath the hood, it's a subtraction.

You take the moment a request left the client, the moment the first output token arrived, and if you want to know how much of the gap belonged to the network rather than the model serving stack. You need a third mark at the network boundary. Three timestamps, three places, three clocks.

The number that comes out is worth something only if those clocks agree. When they don't, nothing announces it. You get a plausible number that's wrong, and you get it every time you run the measurement.

6G Narrows Already Tight Timing Tolerances

Cell phase synchronization for time-division duplex (TDD) is specified at 3 µs between base stations, which works out to ±1.5 µs each. The time reference at the input to a base station must be better than ±1.1 µs, which leaves roughly 400 ns for the radio itself. If you switch on features that need tighter alignment, the budget collapses: TS 38.104 puts time alignment error between 65 and 260 ns depending on the feature, and only for colocated sites.

Positioning is where it gets uncomfortable. One meter of positioning error corresponds to about 3 ns of synchronization error. ITU hasn't published the final IMT-2030 accuracy figures yet, but nobody is arguing for looser than a meter. So, the time tolerance across a single system moves from microseconds to nanoseconds depending on which function you ask, and every one of those functions draws on the same reference.

The Reference is the Softest Input in the Network Timing Chain

Most of that time traces back to GNSS. In RF terms, that means the foundation of the network's timing is a signal arriving from roughly 20,000 km away at a received power below the noise floor. Nothing else in the chain is easier to overpower.

And it’s being overpowered, deliberately and at scale. Aviation has the best public record. European flights encountering GNSS interference rose from about 200 a day in the first quarter of 2024 to roughly 900 a day in the second, and EUROCONTROL has estimated that up to 38% of European en-route traffic passes through regions with intermittent interference. EASA has now revised its safety bulletin on GNSS outages four times, most recently in July 2026.

That's aviation rather than telecom, and the distinction matters. Nobody publishes an equivalent daily count for base stations. But interference doesn't select by receiver type. A cell site inside one of those footprints sits in the same RF environment as the aircraft overhead.

The Failure Mode is Drift, and Drift Doesn't Cause Alarm

When GNSS goes away, the base station doesn't stop. It enters holdover and the local oscillator carries the time forward. That degradation is gradual. An OCXO buys minutes to hours inside a 1.5-µs budget and a rubidium oscillator buys hours, and neither buys days. Through all of it, the site keeps serving traffic and keeps stamping measurements, and those stamps get quietly less true.

The same applies to an AI model running outside the conditions it was trained on. Both keep producing confident output after the thing that made the output trustworthy has gone. Neither raises a hand.

Spoofing is worse than jamming here. A jammed receiver loses lock and knows it. A spoofer hands the receiver a signal that looks correct, so it stays confident and disciplines the oscillator to a lie. The jammed site knows it's in holdover. The spoofed site thinks it's locked, and everything downstream believes it.

What Happens When the Timing is Modeled?

3GPP's Release 18 study on AI/ML for the air interface picked positioning as one of its three use cases, including fingerprinting approaches that learn a mapping from radio measurements to location. That training data is timestamped and position-labeled. If the reference was degraded while it was collected, the model learns a distorted map. If the reference degrades later, the fingerprint stops matching, and the model has no way to tell whether the environment changed or its own clock moved.

The assurance function meant to catch that degradation reads from the same reference. Check and checked share a failure mode. Correlated failures are what break a redundancy argument, and this one runs underneath the whole stack.

Treat the Clock as an Input Under Test

Two things follow. For one, test the reference the way you'd test any other RF input: controlled GNSS simulation with jamming and spoofing profiles, measured receiver behavior, and a holdover curve characterized against your real budget rather than a datasheet figure. Then, measure what the functions above it do while it degrades. Spirent (a Keysight company) builds GNSS simulators, so treat that recommendation with the suspicion it deserves.

Then make assurance carry the uncertainty. A timestamp is a measurement, and it has an error bar like any other measurement. If a metric is quoted in milliseconds and the clocks might disagree by a microsecond, you're fine and you can say so. If you're asserting position at nanosecond alignment, you aren't, and the assurance record should say which case it's in.

The 6G functional freeze lands in December 2028. At some point before then, the industry decides whether the time reference is an assumption or a tested input. The spoofers improve in the meantime, so test against the ones you'll meet in 2029 rather than those in the field today.

About the Author

Steve Douglas

Steve Douglas

Head of Market Strategy, Spirent Communications

Stephen Douglas is Head of Market Strategy for Spirent Communications (a Keysight company), helping to define technical direction, new innovative solutions and market-leading disruptive technologies. With over 20 years’ experience in telecommunications, Stephen has been at the cutting edge of next generation technologies and has worked across the industry with service providers, start-ups, and network equipment and device manufacturers, helping drive innovation and transformation across the 5G ecosystem.

Stephen is an ardent believer in connected technology and strives to challenge, blur and break down the silos that prevent innovation and business success.

Sign up for our eNewsletters
Get the latest news and updates