Summary

  • AMS-IX's quality page states a network-availability aim of at least 99.99% with no measurement point and no window; the per-port objective of at least 99,99% per year is the one printed figure that carries both a unit and a period.
  • Scheduled maintenance, member-caused malfunction and force majeure are cut out of the definition of service failure, so excluded hours leave before the percentage is formed — and across the six pages read on 21 November 2026, no per-incident record carrying both endpoints and a stated cause appears.

The exchange's quality page says it plainly: "The aim of the AMS-IX Network Operations Centre is to have a network availability of at least 99.99%." (AMS-IX Quality Statement) Aim — that is the publisher's own word, and the sentence stops there. No measurement point. No averaging window. No statement of whether the figure describes the peering LAN, a named platform, a named site or a single member's port. The companion page, Quality Statement 2, repeats the aim and the same three exclusions and, in the text I retrieved, leaves out the per-port table altogether.

Where the document does get specific, it gets specific about ports. The same page sets an individual service availability objective per port at "at least 99,99% per year", printed with a comma because that is how a Dutch page prints a decimal. This number has a unit and a window, so for once the arithmetic is available — and it is mine, not AMS-IX's. One hundredth of a per cent of a 525,600-minute year is 52.56 minutes, roughly 52 minutes 34 seconds. That is the whole annual allowance a single port's objective buys, and it is the figure a member could actually hold the exchange to.

Then the definition. AMS-IX considers both an interruption of service and a deterioration of service to be service failure, which is a wider net than most operators write down. Three categories are cut out of it: scheduled maintenance; violations of AMS-IX regulations by members that cause the exchange to malfunction; and force majeure. Read the mechanics rather than the sentiment. The exclusions sit on the definition of failure, which means excluded intervals are removed before the percentage is formed, not subtracted afterwards from a published number.

How large that removal is, none of the six pages I read on 21 November 2026 prints, and I will not invent an estimate for it.

The only measurement geometry any of those pages describes is per-port. Probes sit at each access router that members connect to, measuring connection performance across the AMS-IX infrastructure towards other probes. It is a sensible design, and it is a design that produces port-scale readings. Nothing on the pages says the network-level aim is computed from those probes, and nothing says it is computed from anything else.

The commercial answer exists and stays off the page. The Resources page offers members the option to extend the regular port service with a carrier-grade service level agreement, and prints this: "It is a carrier-grade SLA with service credits up to 100% of the monthly fees upon underperformance." The option covers initial provisioning and the daily availability of the port, and it is offered at all AMS-IX points of presence in the Amsterdam metro. It prints no availability percentage, no credit schedule, no measurement point and no definition of unavailability. So the penalty-backed instrument is real, priced in the only unit the page gives — up to a month's fee — and published one signature away from anyone who wants to compare the commitment with the aim.

The public dashboard is a different object on a different clock. status.ams-ix.net displays "Uptime over the past 90 days" with a link to historical uptime. When I read it on 21 November 2026, every listed component showed as Operational and the text also carried scheduled maintenance notices. What the page's method is, what it excludes, and whether the displayed percentage is per component or an aggregate of them are not evident from the text I retrieved — a statement about what the page displays, not a claim that no method exists anywhere. Its window is also not the quality page's window: ninety days, not a year. The same one hundredth of a per cent over 129,600 minutes is 12.96 minutes, my arithmetic again. Two printed percentages a reader will instinctively read as one number, and two different durations behind them.

The annual report supplies the tone the dashboard cannot: "Results: service availability close to 100% in all locations, zero SLA claims for outages and happy users." (Chair's statement, 2024) Close to 100% in all locations, with no window, no measurement point and no exclusions stated in that sentence. And zero SLA claims measures claims, not availability. It tells you how many members invoked the optional agreement. It does not tell you how many minutes the fabric was down.

One more sentence deserves its own paragraph, because it is the only published promise about telling anyone anything. Scheduled maintenance, the quality page says, is always announced to the tech-l mailing list with at least 72 hours' notice. The list's information page did not come back to me in readable form on 21 November 2026, so what that archive holds and who may read it are not established here; the 72-hour sentence rests on the quality page's own words. Note also what the promise covers: notice of planned work, not a record of what the planned work did.

Put the pieces in order and the shape of the record becomes visible. AMS-IX publishes the aim, the objectives, the exclusions and the notice commitment. On the six pages read on 21 November 2026 it does not publish the measurement point or the window of the network-level aim, and it does not publish a per-incident entry carrying a start, an end, a stated cause and the affected scope — the one artefact that would let a member take a single bad afternoon and test the percentage against it.

None of those pages states whether the 99.99% aim binds a member who has not bought the optional agreement, and none says whether so-called emergency scheduled maintenance falls inside the scheduled-maintenance exclusion.

BTW has been around this exchange before: its route-steering communities, its route servers, its fifteen buildings and one cable, its 15.822 Tb/s headline. Those pieces asked what a published number counts. The subject of this one is the role BTW records as AMS-IX NOC — and nobody has yet asked where the availability figure is measured.

What would change the classification is unglamorous and cheap. Name the measurement point and the window beside the network-level aim, so the percentage becomes an interval a stranger can convert. Publish the interval record per incident, so a member can pick a case. Print the service level agreement's percentage, credit schedule and definition of unavailability, so the priced commitment can be read by the people paying for it. Any one of the three turns a stated aim into a testable one. None of the six pages read on 21 November 2026 does any of them.

A percentage with no measuring point is not a measurement. It is an aim with the checking end cut off — and the exchange is one published interval record away from making it something a member could argue with.