This chapter is the one that turns networking from a technical subject into a commercial one, and it is the part most engineers never learn. It matters because it explains why latency between two cities is what it is, why bandwidth costs what it costs, and why your traffic sometimes takes a route that appears geographically absurd.

• Transit — I pay you, and you carry my traffic to everywhere. You announce my prefixes to the whole internet. This is the only relationship that gives full reachability, which is why every network needs at least one transit provider or must be large enough to peer with everyone.
• Peering — we exchange traffic between our own customers only, usually for free, because it saves us both money on transit. I will not carry your traffic to a third party. Most peering is settlement-free and is agreed on a handshake plus a short document.
• Customer — the inverse of transit. You pay me.
• A tier 1 network is defined precisely: one that reaches the entire internet purely through peering, buying transit from nobody. There are roughly a dozen and the membership changes slowly.

The BGP consequence is direct and worth stating in an interview, because it explains routing behaviour that otherwise looks irrational: a network will always prefer a customer route (it earns money), then a peer route (it is free), then a transit route (it costs money) — regardless of AS path length. That preference is implemented with local preference, which sits at position 2 in the decision process, above AS path at position 4. Money outranks distance, structurally.

An IXP is a shared layer-2 fabric in a data centre. You buy one port, connect once, and can then peer with hundreds of other members over that single port — typically via a route server, so you do not need hundreds of individual agreements. You pay for the port per month, not for the traffic, which means the marginal cost of an additional gigabit is zero.

This is why content networks build out to exchanges aggressively, and why the map of major IXPs is essentially the map of where the internet is dense: Frankfurt, Amsterdam, London, Singapore, and in India, Mumbai and Chennai. India's exchange landscape matured substantially in the 2020s — DE-CIX India, Extreme IX, NIXI and others — which is a large part of why domestic Indian latency and transit pricing improved so much over that decade. Before that, a meaningful share of traffic between two Indian networks physically transited Singapore, which added around 60 ms to a domestic connection for purely commercial reasons.

Wholesale IP transit has fallen by orders of magnitude over two decades — from hundreds of dollars per Mbps per month in the early 2000s to low single-digit dollars, and well below a dollar at very large commitments in the densest markets. Two consequences follow, and they are the useful part:

• Bandwidth is no longer the expensive part of a data centre. Power is. If you are reasoning about infrastructure cost and you start with bandwidth, you are starting in the wrong place.
• Cloud egress pricing bears no relation to the underlying cost of transit. The gap between what a hyperscaler pays for a gigabit and what it charges for one is very large, which is why egress is a margin line and why the entire ‘data gravity’ argument — it is cheap to put data in and expensive to take it out — exists as a commercial strategy rather than a technical constraint.

Transit is billed on the 95th percentile: throughput is sampled every five minutes for a month, the samples are sorted, the top 5% are discarded, and you pay on the highest remaining sample. Five percent of a month is about 36 hours. That is a real, exploitable allowance — schedule bulk transfers, backups and replication into deliberate short bursts rather than smearing them across the month, and you can move a meaningful amount of data for nothing.

Around 99% of intercontinental traffic goes through submarine fibre. Light in fibre travels at roughly two-thirds of the speed of light in vacuum, which sets a hard floor on latency that no amount of engineering removes.

Two operational points follow. First, when you place infrastructure, you are choosing a latency floor you cannot later optimise away — which is why “we will make it faster later” fails for anything cross-ocean, and why edge termination exists. Second, submarine cables have chokepoints: the Red Sea corridor carries a large share of Europe–Asia traffic, and cable cuts there have repeatedly degraded Indian and Gulf connectivity for weeks. Route diversity at the cable level is a real consideration for anyone running intercontinental services.