Peering and Transit: The Economics Behind Internet Traffic
Nobody owns the internet, and nobody runs it, but everybody pays for some of it. The money moves through two arrangements — transit, which you buy, and peering, which you usually do not — and the tension between them explains why Netflix ships free hardware to internet providers, why Toronto has its own exchange point, and why two large carriers once cut each other off and partitioned part of the internet for three days.
Transit is a purchase. Peering is a swap.
Transit is what a small network buys from a large one: the right to send traffic to, and receive traffic from, everywhere else. Cloudflare’s learning centre describes it as a paid arrangement that gives the paying network access to the rest of the internet. It is billed on capacity, usually on the 95th-percentile of measured usage, and it is the default position for anyone who cannot get what they need for free.
Peering is a direct link between two networks for the exclusive purpose of exchanging traffic between themselves and their own customers — not the whole internet. Cloudflare’s figure is that peering is free in over 99% of cases, and the free form has a name: settlement-free peering.
Paid peering is the awkward middle. It looks like peering — a direct link, limited to each party’s own customers — but money changes hands, typically because one side has more leverage: a large access network with subscribers the other side needs to reach.
| Transit | Settlement-free peering | Paid peering | |
|---|---|---|---|
| What you reach | The entire internet | The peer and its customers only | The peer and its customers only |
| Who pays | The buyer, on committed or 95th-percentile capacity | Nobody; each side covers its own port and cross-connect | One side pays the other |
| Typical motivation | You need reachability you cannot get any other way | Both sides save transit money and remove a hop | One side wants direct capacity into an access network badly enough to pay |
| Failure mode | You can escalate — you are a customer | Either side can walk away; congested ports can sit unupgraded | Renegotiation leverage sits with whoever has the eyeballs |
A tier-1 network is the endpoint of that logic: a network that reaches every destination on the internet through settlement-free peering alone, buying transit from nobody. The label describes a settlement position, not speed, size or quality, and it is asserted more often than it is verified.
Internet exchange points
Peering with fifty networks one cable at a time is unworkable. An internet exchange point solves it with a shared switching fabric: everyone buys one port, plugs in once, and can then peer with anyone else on the fabric — bilaterally, or through a route server that automates the arrangement with many networks at once.
The large European exchanges are the reference points. AMS-IX in Amsterdam reported a 2025 peak of 15.095 Tb/s across 902 connected networks, moving 35.66 exabytes over the year. DE-CIX states more than 18 Tb/s of peak traffic in Frankfurt alone, and over 4,600 connected networks across more than 60 locations worldwide. LINX in London reports roughly 900 members and over 950 autonomous system numbers connecting from more than 80 countries.
Canada’s own is TorIX, the Toronto Internet Exchange Community, a not-for-profit with over 250 connected organisations. Its PeeringDB record lists 233 peers across 279 connections and 15.5 Tbps of connected capacity, spread over eight Toronto facilities. The reason it matters to Canadians is unglamorous: without a domestic exchange, traffic between two Canadian networks has historically had a habit of touring the United States and coming back, adding both latency and a jurisdiction.
Why the content companies stopped buying transit
If you are a video service, transit is the worst possible way to deliver your traffic. You pay per bit, you have no control over the congested link three networks away, and every viewer’s stream crosses the same handful of interconnections at the same time in the evening. So the large content networks did two things.
First, they built backbones — private long-haul capacity connecting their own data centres to exchange points and directly to access networks, so that traffic touches the public internet only at the last handoff.
Second, they moved the content inside the access networks. Netflix’s Open Connect programme supplies caching appliances to qualifying internet providers at no charge. The published specifications are modest and specific: a storage appliance is a 2U server holding up to 120 TB and serving around 200 Gbps at roughly 400 W, while a lower-cost global appliance holds up to 60 TB and serves around 80 Gbps at about 250 W. Netflix says it works with over a thousand internet providers and also runs appliances in more than sixty of its own data centres, with an open peering policy at exchange points.
Google runs the same play as Google Global Cache. Google’s own documentation describes GGC as a way for providers to serve certain Google content from within their own networks, and states that typically between 70% and 90% of cacheable traffic can be served from a GGC node. Google supplies the hardware; the provider supplies rack space, power, a network connection and a BGP session.
The economics are simple enough that both sides usually agree. The content company avoids transit and inter-network congestion. The access network avoids hauling the same popular video across its expensive links thousands of times, because after the first request it is stored down the road. The cost is structural: an ever-larger share of consumer traffic never crosses the public internet at all.
Two disputes worth naming
Cogent and Level 3, October 2005. Level 3 had warned Cogent in July that it would end their settlement-free arrangement unless Cogent paid, arguing it carried a disproportionate share of the traffic. On Wednesday 5 October 2005 Level 3 disconnected the peering, and customers of each network could not reach parts of the other. Level 3 restored the link on Friday 7 October, describing it as a free connection re-established because users had been affected “through no fault of their own”, but only until 9 November. On 28 October 2005 the two announced a modified agreement: settlement-free exchange with specific payments if certain obligations were not met, advance written notice to customers on termination, and a guaranteed transition period. The threatened November disconnection never happened. The precedent that stuck was the last bit — that walking away from peering should come with notice, because third parties get hurt.
Netflix and Comcast, 2014. After a period of degraded Netflix performance for Comcast subscribers, the two announced a direct interconnection agreement in February 2014, with Netflix paying. On 20 March 2014 Netflix’s then chief executive Reed Hastings published “Internet Tolls And The Case For Strong Net Neutrality”, arguing that strong net neutrality should prevent providers charging for the interconnection needed to deliver good service. Netflix went on to sign similar arrangements with other large American providers that year.
Be careful about the causal story, though. Researchers at Princeton’s Center for Information Technology Policy examined the measurements in 2015 and concluded that congestion appeared to be occurring both at the interconnection points and inside transit providers’ own networks, and that the available techniques could not definitively separate the two. Both sides had a version of events that the data did not settle.
The practical version
- Peering is not a discount on transit. It only reaches the peer and its customers, so it complements transit rather than replacing it, unless you are one of the very few networks that can reach everything this way.
- If you run a network with real traffic, look up your local exchange. PeeringDB is the industry’s shared directory of who is where and who peers openly.
- Cheap transit and good transit are different products. Price per megabit tells you nothing about how congested the seller’s interconnections are at 9pm.
- For Canadian networks, domestic exchange presence is a latency decision and a jurisdiction decision. Traffic that stays in Toronto is faster and stays in Canada.
- Treat “tier 1” as a claim, not a specification. What matters to your users is where a network interconnects and whether those ports are upgraded before they fill.