Given all the hype surrounding AI, it must have driven internet traffic through the roof, right? Well, think again.
According to a new whitepaper from Data Centre Alliance member Ian Bitterlin, internet traffic through the Amsterdam Internet Exchange has not shown an obvious AI-era surge, despite the enormous investment now flowing into artificial intelligence infrastructure.
The whitepaper took a look at historical traffic recorded by AMS-IX between 2001 and 2026, and compared three distinct periods: the long-term growth of the internet, the Covid-19 pandemic and the subsequent boom in AI investment.
Bitterlin notes in the whitepaper that the Covid-19 pandemic produced a clear step-change in traffic, but there has thus been no equivalent acceleration since generative AI began driving much greater investment in compute infrastructure. In fact, he found that annual traffic growth through AMS-IX fell to 3.8% in 2025, while monthly traffic plateaued at around 3.1 exabytes during the opening months of 2026.
Of course, that doesn’t automatically mean that demand for AI infrastructure is weak, but it does raise the question as to when, where and how the growing use of AI will become visible in the wider network traffic.
Covid provides a useful comparison
According to data from AMS-IX, monthly traffic at the exchange rose from around 690TB in June 2001 to 1.64 million TB by December 2019. That’s a significant jump, but not entirely unexpected given the growing power of the internet over the almost two decade period.
The more abrupt change came in March 2020, when lockdown restrictions across the world pushed everyone online. Bitterlin said during that period he saw a roughly 20% jump in traffic, which proves that a significant change in digital behaviour can produce a visible change in exchange traffic.
And yet, the first years of the AI boom have not produced anything comparable. Bitterlin argues that part of the explanation may lie in the difference between AI training and inference.
Training large models requires enormous amounts of compute, but that does not necessarily translate into a corresponding increase in traffic through public internet exchanges. Additionally, much of the largest AI training infrastructure has been concentrated outside Europe, so it wouldn’t have even gone near AMS-IX.
Meanwhile, inference – the everyday use of trained AI models by businesses and consumers – is potentially more relevant to future internet traffic, but Bitterlin questions whether widespread use of AI has yet reached the scale implied by the infrastructure investment now taking place.
AMS-IX can’t tell the whole story
Of course, there needs to be a word of caution to the figures. AMS-IX is one internet exchange and, while Bitterlin argues that Amsterdam can provide a useful proxy for wider European trends, its traffic does not represent all network activity across the continent.
Additionally, internet exchange traffic isn’t entirely the same thing as total data centre activity. AI workloads can generate substantial data movement inside individual facilities, between tightly connected GPU clusters and across private hyperscale and cloud networks without that traffic necessarily appearing at an exchange such as AMS-IX.
Relatively weak growth at the exchange therefore can’t, by itself, show that AI demand is overstated or that infrastructure is being overbuilt. But, what the figures do challenge is the simpler assumption that rapidly increasing AI compute demand should already be producing an equally dramatic rise in conventional internet traffic.
For now, the contrast remains notable: Covid produced a clear jump in AMS-IX traffic, while the first years of the AI infrastructure boom have not. Whether that changes as AI inference becomes more widely adopted may offer a clearer indication of how quickly infrastructure investment is translating into everyday network demand.
You can read Ian Bitterlin’s full whitepaper here.
