The back half of 2024 was rough for smaller companies with impressive global networks.
While there are still a number of niche players who have built their own networks (shout out to pure play delivery folks like KeyCDN, CDN77, and CacheFly; network-focused hosting companies like NetActuate, EdgeUno, and Latitude; and 2nd tier clouds Digital Ocean, Servers.com, Vultr, and OVH), the question must be asked: what gives?
Most recent failures (StackPath, Edgio, etc) simply didn’t generate enough market traction to overcome the weight of expensive infrastructure footprints. Aside from Equinix Metal, which never offered a delivery product, most edge platforms started with a CDN as the “no brainer” way to monetize the distributed network they were building.
Unfortunately, core CDN products have a 20 year history of commoditization, framed against a backdrop of rising underlying costs. As Matt Levine, one of the smartest folks in the business, recently wrote:
As bandwidth costs bottom out, other operational expenses are becoming more prominent. Power, labor, cross-connects, and compute resources are now the dominant factors in CDN pricing, and these costs are experiencing steady inflationary pressures.
Unlike traditional SaaS or other industries, if you’re delivering bits for money, the attack vectors and scale of the internet simply require you to get big quickly in order to even offer a basic service. And while you don’t need thousands of POPs to be relevant, you do need 15-30 key global locations with substantial routing, server, transit, peering and direct connectivity capacity. That’s a lot of CAPEX, but also ongoing OPEX and inherent upgrade cycles.
Outside of the hyperscale networks, the most recognized global edge networks are Cloudflare, Akamai, and Fastly. Each has moved heavily into security over the last few years and built up a platform story. Here is some copy I took from their respective websites:
These are each amazing global platforms with hundreds of TB’s of connected capacity. But the story in the market suggests that all is not equal — Fastly is trading at 1.9x revenue, Akamai at 3.63x, and Cloudflare at 22.5x. One of these is not like the other! While Akamai is a more mature business, it’s clear that growth is what drives stock value:
It’s clear that Cloudflare’s platform experience, hybrid sales motion, and rapidly expanding product portfolio has created a compelling engine that customers want.
Because of the underlying economics and competitive market dynamics, new cloud companies haven’t even tried to build their own networks.
There are a growing number of platforms that help developers (or even non-developers like me) deploy and manage apps at scale without worrying about infrastructure, let alone networks. These are modern versions of the grand daddy of “PaaS” ( ♥️Heroku ♥️) and they’re creating an entire generation of folks whose default deployment target is “everywhere” vs “US-EAST”.
Vercel is probably the biggest of the lot, having raised $250M in fresh capital to expand its AI and security features, which includes a WAF and a firewall service alongside reduced bandwidth pricing (down from a shocking $0.40/GB to a still incredible $0.15/GB).
The difference here is that Vercel doesn’t own any infrastructure. It’s built on top of AWS! And for customers, the value is abstraction, automation, and platform features — not cheaper bandwidth or differentiated network access, routes or performance.
If you’re not competing on moving bits, do you really need to own your network, or can you just borrow one? It seems the answer for this class of startups is to borrow! Of all my favorite platforms, only Fly.io runs a good chunk of its own infrastructure. Go, Kurt, go!
What this tells me — aside from the fact that AWS and GCP are fantastic businesses — is that software has largely caught up to the “run anywhere” promise of cloud native, and that even the most infrastructure-focused developers don’t think below Layer 4.
With public clouds, smart teams like these can leverage common tooling (Kubernetes or other advanced schedulers like Nomad, distributed databases, standard observability stacks, service meshes, etc) to build and run their own global clouds. They may not have a great cost basis or full control under the hood, but they don’t have to build a massive global footprint to deliver value to clients. As GEICO and Basecamp recently demonstrated, one can certainly buy gear further down the road when the numbers make sense.
Time will tell if this class of companies — which are directing capital towards platform features and customer acquisition instead of their own infrastructure — fares better than the network builders of the previous wave.
Although it's harder to see, there is an entire ecosystem of providers working on the private side of the internet: Equinix Fabric, Digital Realty Service Fabric, PCCW Console Connect, Packet Fabric (for now!), and Megaport. Traditional telcos like Lumen and Zayo are also playing a bit above transit and transport by packaging up their IP offerings.
These folks operate in network OSI layers 0 through 3 and offer physical cables (cross connects) or virtual cables (VLANS-as-a-service). Many now provide Layer 3 services via cloud routers, but their bread and butter is private interconnection and routing traffic between diverse networks.
This lower layer is profitable, but with a catch: you need to either have “owner economics” on the underlying telco or ensure that interconnection happens exclusively in buildings that you operate.
While these services continue to grow, the experience caters to a network/I.T. buyer that is aging out, or at least becoming more niche. It’s unlikely that any of the “deploy everywhere” platforms even know about these services, which aren’t designed for them. Instead, they were designed for folks who build, deploy and manage physical infrastructure and networks.
In the next decade, one might ask “if a cross connect fails, and a developer isn’t there to notice, did it really fail?”
It’s obvious that the artificial intelligence craze stimulated by ChatGPT and broader generative AI has kicked off a huge infrastructure buildout.
While much of the focus has been on GPU capacity and the associated space/power/cooling to support those servers, it’s clearly driving network demand as well. Telco provider Lumen saw its fortunes reverse after years of eroding margins and a stagnant stock price. Announcing $5B in new orders driven by AI companies seems to have a positive impact!
In terms of AI specific clouds, CoreWeave is the furthest ahead in building their own network, showing up at all the expected Equinix peering locations in the United States as well as some other interconnection points.
On the other hand, Lambda and Crusoe both have limited footprints. Ori, an AI company that offers both public and private clouds, leverages a new purpose-built connectivity platform called Stelia. We expect that the data and network heavy nature of AI workloads will encourage more examples like this to pop up.
In the meantime, Cloudflare is busy deploying GPUs to support its serverless Workers AI product, with more than 150 locations live so far.
It’s always been madness to build and run a global network…until it isn’t (looking at you Netflix, Apple, Roblox, VISA, Walmart, Meta, etc). In order to succeed, we’re taking some lessons from the recent failures and successes.
We believe that the gap between network “have” and “have nots” is growing alongside a Cambrian explosion in complexity driven by AI, distributed systems, public cloud, wireless, and localized regulation.
Our mission is to level the playing field in a way that application and DevOps teams love. More “Plaid for Networks” than “A-Side” to “Z-Side”.
Wish us luck!
