They’ve written about it before: https://blog.cloudflare.com/aws-egregious-egress
For larger providers, bandwidth cost drops tremendously, especially if you’re well connected as transit is much cheaper and if you are really large or a network provider you may even be routing between your own facilities or in some cases from one customer to another and every large scale isp is going to want a “direct link” to your facility (a peering relationship). Those costs are astronomically small at scale for bandwidth.
The ISP or similar then turns around and sells a sustained network throughout as GB transferred, which isn’t how wholesale bandwidth is sold at all. So the get to charge for the data the pipe moves while they only pay for the connection itself — the markup added to this process is considerable.
For someone operated a global CDN, which is basically what they do, they have racks of storage and computer collocated all over the world and optimize the living crap out of their network to reduce their costs and make it run on as many peering relationships as possible. It’s an expensive and complex business to set up, but once it’s set up you get a fairly good and consistent return out of it.
The reason for this article is related to the nature of that business: it’s the issue of liability.
When you have policies where you protect your clients from downsides and excessive use on the network, you suddenly have to assume the role of paying attention to what’s on the network and policing it’s contents. That’s not possible with a massive system like this generally, so they push the liability down to the customer and discount the mistakes that come up. That’s why things are set up like this… this kind of stuff isn’t their business at all really. They are looking for the customers that convert and pay, which is very profitable, and the free tier is often thought of as a sustainable cost if you are large enough scale, as it substitutes for the rather massive expense of marketing and sales which is one of the largest expenses in a bandwidth focused business. CAC is the free tier.
There also competitors, but the benefits of scale are tremendous in terms of cost efficiency. A large provider might be paying just a very small fraction of a penny or less (even “free”) compared to what a small provider is paying. So that’s why you end up with fewer competitors because it truly is a business that benefits from economies of scale.
There are other smarter people on here who can correct any mistakes I’ve made or provide better pricing or whatever, but that’s the more in depth answer.
Those big providers have pretty much normalized high fees and convinced people that's what it costs, the reality is any normal provider like Hetzner for example gives you tons of bandwidth for essentially zero cost included with your servers.
The invisible hand of the free market has come to fix that, *but you have to opt into the hand by shopping around.* If you don't, you don't get its benefits! You have to willingly take the choice to move to cheaper providers instead of overpriced ones.
Hetzner Cloud: $1/TB (20TB free) Digital Ocean: $10/TB (few TB free depending on server size) AWS: $90/TB (0.1TB free, used to be 0.001TB free) Netlify: $550/TB (0.1TB or 1TB free)
If you move up from $5/month VPSes, to real dedicated servers, you are now spending a lot more money and therefore you get more free perks. A huge number of providers exist that will give you unlimited or unlimited† bandwidth depending on how much you spend. Renting a powerful server with unlimited 1Gbps should cost a few hundred to several hundred dollars per month, and a powerful server with unlimited 10Gbps (i.e. 3000TB/month) should cost a few thousand dollars per month. You can even get some with 100Gbps (for tens of thousands).
Also consider asking your local ISPs and datacenters. If you live in a central area, you can probably get a comparable connection to a nearby datacenter if not straight to your office, for a comparable price. Data center connections are their bread and butter and they should be able to give you a quote quite rapidly; to your office will be a more custom thing.
Recently I got a quote for AMS-IX peering in Berlin, i.e. a peering in Amsterdam plus a link from Amsterdam to Berlin, about a 600km distance. That would cost 950 euros per month. If 1Gbps, it would cost 300 euros per month. Even though it's not really got anything to do with internet access (transit), I include this number to give some indication of the "true" cost of "raw" bandwidth.
"your margin is my opportunity"
(from 2014, so it might be super outdated)
As a business for Cloudflare?
Cloudflare in 2014 blogged about how they work relentlessly to bring down bandwidth costs by peering aggressively where possible [2] (which apparently means $0 for unlimited bandwidth [3]). And where they can't / don't [4], egress is 5x (est) the ingress (one pays for the higher among the two), but this creates an opportunity for an arbitrage and give away DDoS protection for free.
This is pretty similar to Amazon's free-shipping offer for Prime customers despite it being one of the biggest loss makers to their retail business. Prime basically has since forced Amazon to bring down costs through building expensive and vast distribution & logistics network that spawns the globe. Doing so was a considerable drain on the resources in the short-run, but in the long run, it has become an unbreachable moat around its largest business.
Analysts like Ben Thompson (stratechery.com) and Matthew Eash (hhhypergrowth.com) have written in detail about Cloudflare's modus operandii over the years, with both agreeing that Cloudflare's model is so brilliantly disruptive that even Clayton Christensen would be proud of it.
Here's how it works:
1) I have a big network and I exchange traffic with another big network. Think of "eyeball" networks like last-mile ISPs (Comcast, mobile providers, etc) where a substantial portion of end-user traffic is going to handfuls of well known networks - Cloudflare, AWS, Netflix, etc.
2) Comcast and Cloudflare say "Hey, I send you X TB/PB/etc and you send me X TB/PB/etc. We both currently pay another provider to route that traffic between us. Let's not do that."
3) In locations where it makes sense they basically throw a cable across datacenters, POPs, internet exchanges, etc. The cost for this is typically extremely low - it's basically a port on a switch/router on each side and MAYBE a "cross connect fee" from the facility. This is usually billed in the tens of dollars/mo if at all. It takes very little time/effort to configure this but of course the details are more complex - multiple ports, multiple facilities, etc.
4) Both sides start routing traffic between their networks over their new shiny direct cables and extremely high speed ports. Faster throughput, lower latency, improved reliability, frees up bandwidth to the transit provider they were using previously, and most importantly the cost of bandwidth between the two networks goes to zero.
This is all well known and publicly available because it's visible in the global routing table(s). Cloudflare, for example[0].
All of the large providers do this and AWS, etc charging in bandwidth per GB (especially at their rates) is more-or-less pure profit.
I have a theory that AWS, etc capitalize on people not really understanding this anymore. AWS is 20 years old - that's an entire generation of CTO/CIOs on down that are completely unfamiliar with these details and think $0.10/GB or whatever is "just what bandwidth costs". It is not.