Most data center operators say low latency. Few give you a number you can check. The words on operators’ marketing pages are “ultra-low latency,” “low-latency paths,” and “excellent latency.” The number, the destination, the method, and the date are missing. That is not an accident. A latency number without those four things is not a fact. It is a slogan. This post explains why, and commits to doing it properly.
Start with how latency is measured. Latency is measured in milliseconds, as the time elapsed between sending a data packet and receiving the acknowledgment that it arrived. Ping is the simplest tool for measuring it, because it sends a small packet to a target address and measures how long the response takes. Distance sets a floor, because light in fiber travels at roughly two-thirds of its speed in a vacuum, about 200,000 km per second. New York to London is about 5,500 km in a straight line, so the floor is roughly 27 ms one way and 55 ms round trip, before any routing detour or processing delay. Each network hop adds delay, because data passes through routers, switches, and nodes that inspect, process, and forward each packet. Congestion adds variable delay, because overloaded infrastructure makes packets queue while waiting to be processed. A single number hides all of that. Two round-trip values can describe different paths, different loads, and different times of day.
Now look at what operators actually publish. Equinix’s SCX.ai case study carries the statistic “Sub-10 ms ultra-low latency connectivity” and says its sovereign AI infrastructure delivers ultra-low latency connectivity to support real-time applications by keeping data, model training, and inferencing onshore. Iron Mountain describes Frankfurt as providing “excellent latency across Europe” and “optimal latency across Europe,” without a method on that page. Iron Mountain also publishes numbers for Frankfurt: 14 ms to London, 11 ms to Paris, 28 ms to Madrid, and 14 ms to Amsterdam. No measurement method, date, or percentile is stated on that page. CoreSite says its ecosystem “facilitates low-latency access to a variety of carriers and financial services providers” and lists a customer challenge as providing ultra low-latency connectivity to exchanges and the trading floor. The same CoreSite story reports “sub 10ms round trip connectivity,” a threshold with no named endpoints or date. CoreSite states a customer “achieves less than two milliseconds of latency to public clouds,” where the destination is generic public clouds, not a named endpoint. DartPoints says “Lower latency, higher throughput, and more consistent inference performance across your footprint” and lists “Low-latency paths for inference and real-time workloads,” with no millisecond number on either page. The pattern is consistent. On the pages we reviewed from these four operators, the figures that exist are thresholds, generic destinations, or city-to-city numbers with no stated method, date, or percentile. Iron Mountain comes closest, naming both ends of each path, but none of these pages gives a named origin, a named destination, a stated method, and a date together.
That matters because inference, unlike training, is latency sensitive. Iron Mountain states that training is latency insensitive, because milliseconds of external latency do not affect the outcome, so training facilities can go where power and land are cheapest. Inference is the opposite. Consumer-facing AI generally targets sub-50 ms response times, which means inference infrastructure must sit close to population centers. Iron Mountain’s comparison table lists latency as “Not critical” for training and critical, at under 50 ms, for inference. Users experience the latency of the entire application, not only the model, because each step in the request path contributes to total response time. A few milliseconds on the network, multiplied across a chain of calls, is the difference between a product that feels instant and one that feels slow. That is why inference has to sit close to users, and why the number has to be real.
So here is how a latency figure should be measured and reported, and how we will do it. A useful test is a repeatable experiment, and the record should include source host, destination, interface, protocol, packet size, interval, test window, system time, route, and workload state. Without that context, two round-trip values may describe different paths and different conditions. A single ping of 20 to 40 ms does not prove an application is healthy, so repeated tests should baseline the minimum off-peak round-trip time, then compare median, percentiles, jitter, loss, and behavior under load. For each window, record minimum, maximum, median, selected percentiles, loss, and jitter, and do not report the arithmetic mean as the only result. The rule is simple, and it is stated well in one API provider’s benchmarking documentation: a latency number without context is noise, so record client environment, server region, network path, sample count, measurement method, percentile breakdown, and date and API version. Never promise “global low latency.” Publish your measured numbers with their environment attached.
That is our commitment. We will measure latency from our sites to named hubs, and we will publish the result with the origin, the destination, the method, the percentile, and the date attached. We will not publish a number we have not measured. We will not publish a threshold and call it a measurement. The number that matters is the one you can check, and the check is the method. When the measurement is done, we will share it under NDA with the power and fiber facts for each site, where a tenant can use it. Until then, we will not claim a millisecond figure, because an unmeasured number is worse than no number. It is a promise we cannot keep.
Sources
- Operator pages were viewed in November 2026 and may have changed since.
- v2cloud, What Is Latency: https://v2cloud.com/glossary/what-is-latency
- Equinix, SCX.ai delivers 4x faster sovereign AI inferencing: https://www.equinix.com/insights/case-studies/scxai
- Iron Mountain, Frankfurt data center: https://www.ironmountain.com/data-centers/locations/emea/frankfurt
- Iron Mountain, AI training vs. AI inference data centers: https://resources.ironmountain.com/blogs-and-articles/d/data-centers-ai-training-vs-ai-inference-data-centers-whats-the-difference-and-why-does-it-matter
- CoreSite, customer stories on low-latency connectivity: https://www.coresite.com/customer-stories/architecting-unified-experiences-from-cloud-migration-to-cybersecurity-for-enterprise-clients and https://www.coresite.com/customer-stories/financial-services-provider
- DartPoints, technology and financial services pages: https://dartpoints.com/technology-service-provider/ and https://dartpoints.com/industries/financial-services/
- EdgeNext, Edge AI Inference: Why AI Is Moving Closer to Users: https://www.edgenext.com/resources/blog/edge-ai-inference-why-ai-is-moving-closer-to-users
- ArpHost, Network Latency Testing: How to Measure and Fix Lag Fast: https://arphost.com/network-latency-testing/
- Michitai, Latency documentation: https://api.michitai.com/docs/latency.php
