Latency is a geographic problem first
Light travels through fibre at roughly two-thirds the speed of light in a vacuum. That physics sets a floor on latency that no network operator can engineer around: New York to London is at minimum about 35ms round-trip, and New York to Singapore is at minimum about 90ms. The practical question is not how to defeat physics but how to avoid adding unnecessary distance through routing detours.
A routing detour happens when your transit provider does not have a direct exchange peering with the destination network and has to hand your traffic to an upstream carrier, which then hands it to another, adding geography and hops. The more direct peering a transit provider has with networks in your target regions, the fewer detours your traffic takes.
Exchange presence geography, not just exchange count
A transit provider present at 30 internet exchanges in Western Europe and 2 in Asia is not a good choice for traffic destined for Tokyo or Singapore, even though its total exchange count looks impressive. When evaluating providers for low-latency requirements, ask where their exchange presence is concentrated, not just how many exchanges they use overall.
The major regional exchange clusters to ask about: North America (NYIIX, Equinix, DE-CIX New York), Western Europe (AMS-IX, DE-CIX Frankfurt, LINX London, DECIX Paris), Asia-Pacific (JPNAP, HKIX, SGIX), and Latin America (PTT Metro). A provider with dense presence in the cluster closest to your traffic destinations will deliver structurally better latency than one with global breadth but shallow regional presence.
What to ask about in SLA terms
Standard transit SLAs specify uptime and sometimes a global average latency threshold. Global averages are nearly useless for latency-sensitive applications because they smooth over the regional variance that matters. The clauses worth negotiating are:
- Round-trip latency between your handoff and named exchange points or regions, measured over five-minute samples
- Packet loss threshold (0.1 percent or lower is achievable; anything above 1 percent is a problem for real-time traffic)
- Jitter threshold (relevant for voice and video; often absent from standard contracts)
- MTTR for routing issues, distinct from physical circuit outages
Most carriers will negotiate these clauses if you ask. The willingness to commit to regional latency numbers is itself a signal about how well a provider understands its own network.
Testing before you sign
The only reliable test is a live trial from your actual handoff location to your actual destination networks. Ask every shortlisted provider for a 30-day trial circuit. Run automated traceroutes and round-trip measurements every five minutes to ten key destinations. Run the tests across all hours of the day including peak congestion periods; latency on a lightly loaded network looks identical to an optimally routed one.
Compare hop counts between providers to the same destination. A lower hop count to a specific destination is a structural advantage that persists across load conditions. A provider that takes 14 hops to reach a Tokyo exchange will not catch up to a provider that takes 9 hops when congestion is low.
When to run two providers for latency, not just redundancy
For most buyers, dual transit is primarily about redundancy: one provider backs up the other during outages. For latency-sensitive workloads, dual transit can also be a performance strategy. If Provider A has shorter paths to North American destinations and Provider B has shorter paths to Asian destinations, BGP policy can route each traffic class to the provider with the better path. This requires a network team capable of managing BGP policy, but for real-time global applications the latency improvement can be significant.