Around 2021 I spent months exploring internet connectivity — peering, IXPs, transit economics — as the core domain for a startup. The startup failed for the least technical reason possible: the team’s priorities misaligned. The research survived in a folder. This post is those old exploration notes, lightly edited, plus the questions I’d want answered before going back in.
The startup, half a decade back
Around 2021 I co-founded a startup in the connectivity domain. The thesis was simple to state and hard to prove: the internet’s real margin hides at the interconnection layer. Underneath every consumer subscription is a stack of networks handing traffic to each other, and in most markets that stack still runs on expensive, paid transit when it could run on cheap, mutual peering.
I went deep — transit pricing data, IXP models, CDN economics, the private backbones of the hyperscalers, submarine cable routes. The notes in this post are what came out of that exploration.
The startup didn’t die of competition or technology risk. It died the way most startups die: the team’s priorities drifted out of alignment, and the domain expertise never got paired with the go-to-market muscle it needed. If you’ve lived that, you know it’s the most expensive thing on earth to admit in public.
So consider this the admission — and the post the failure paid for.
What the notes argued
Peering vs. transit: the cost structure
| Metric | Peering (settlement-free) | Transit (paid) | Impact |
|---|---|---|---|
| Cost (per Mbps/mo) | $0.05–$0.30 | $0.50–$2.00 | ~90% cheaper |
| Latency | <20 ms at a local IXP | 100–300 ms via transit | up to 10x lower |
| Reliability | direct paths, 99.99% class | transit hops, 99.9% class | fewer outages |
The rule of thumb in the notes: if a network is paying more than $0.30/Mbps for transit, it’s overpaying by an order of magnitude. Peering is free or near-free to carry traffic directly with a counterparty where flows balance.
The alternatives — and why the notes dismissed them
Transit (the expensive default). Buy capacity from a Tier 1 and let them carry you anywhere. It works, it’s instant, and it’s the most expensive and least controllable option. You inherit someone else’s routing decisions.
CDNs (the band-aid). Caching works beautifully for one-way static content — video files, web pages. It does nothing for two-way, latency-sensitive traffic: video calls, gaming, live collaboration. A CDN is a mirror, not a peer.
Private networks (the vendor toll road). AWS Direct Connect, Azure ExpressRoute — dedicated pipes that skip the public internet. Cheaper per bit than transit and better behaved, but they lock you into one vendor’s egress fees, and they still terminate inside someone else’s walled garden. The notes’ verdict: the only neutral, scalable, low-cost interconnect is IXP peering.
The case studies the notes leaned on
- Netflix. Open Connect is the canonical example: an access ISP and a content giant replacing paid transit with free peering and appliances inside the ISP’s network. The notes estimated the swing in the hundreds of millions of dollars a year — traffic that once flowed over paid transit flows over free interconnects.
- Jio. Built dense in-country interconnection and its own submarine cable capacity, pulling interconnect costs down toward cents per Mbps. Peering economics became a pricing weapon — measurably cheaper plans than rivals paying transit rates.
- AWS (ironically). Even AWS — the company selling the toll road — charges a fraction for Direct Connect of what public-internet transit costs. When the vendor’s own cheaper alternative starts at one-tenth the transit price, that’s the market confessing what the correct price is.
A note on these numbers: they’re assembled from public reporting and industry folklore circa 2021 — order-of-magnitude estimates, not audited figures. I’ve kept the shapes and dropped false precision. The shapes have held for two decades: transit prices fall, peering stays cheap, the gaps in underserved markets persist.
The peering paradox
Peering follows Metcalfe’s law: the more networks connect directly, the cheaper and faster the whole thing gets — for everyone, including the newcomers. Settlement-free peering only stays settlement-free as long as traffic roughly balances, which is exactly why IXPs exist: neutral rooms where balanced flows can find each other.
What’s changed since 2021 — and why the thesis still breathes
- Mature markets solved this; underserved ones didn’t. North America and Western Europe have transit at commodity prices. In parts of South/Southeast Asia, Africa, and MENA, access networks still pay several times more, and IXP coverage is thin. The >$0.30/Mbps band is still out there.
- The traffic mix changed. The notes’ best argument against CDNs — “caches can’t fix two-way traffic” — got a second life: AI. Inference and agent traffic is latency-sensitive, request/response, and non-cacheable. Edge peering paired with inference placement is the same argument with a new payload.
- Cables moved the map. 2Africa, Jio’s IAX/IEX and friends reshaped regional economics. The 2010s built cloud regions; the 2020s are building cable + exchange + edge combos that determine who pays what for the next decade.
The asks
I’m reopening this space deliberately, with better scar tissue. If any of the below is you, I’d genuinely like to talk.
- Revival opportunities. Where would you push interconnection economics in 2026? My current hypothesis is edge peering + AI inference exchange — the “CDN argument” re-run for non-cacheable AI traffic. What am I missing, and what part of this thesis is now dead weight?
- A new team. I’m looking for people strong in the three places the first attempt was thin: peering ops and IX relations, network engineering, and go-to-market for infrastructure. Advisor-shaped part-timers are as welcome as co-founder-shaped full-timers.
- Region validation. If you operate in — or have networks in — a market where access networks still pay multiples of $0.30/Mbps transit, I’d like to validate the demand side. I have a short validation checklist (transit quotes, IXP maturity, cache-fill rates, regulatory posture) and I’ll trade findings.
If any of this overlaps with your world, my inbox on LinkedIn is open — linkedin.com/in/jetty.
Further reading
- Netflix Open Connect — how the content-to-ISP peering model works in practice
- PeeringDB — the crowdsourced database of who peers where
- CAIDA AS Relationships — the academic view of the same web of interconnection
- DE-CIX — exchange-operator write-ups on edge peering and the app-hub model