Finally! As of last week, this blog runs on Datum Cloud. Two years and about $6M of hard-won personal and venture money later, I have a “built from scratch” modern cloud on which to host my blog.
Most people use a static site generator and a free tier. I guess you could say I do cloud hosting the hard way. :-)
I love hosting, and I cannot lie
Datum is my third infrastructure company, after Voxel and Packet. You’d think that by now I could do this stuff in my sleep!
In fact, what I’ve learned from doing this “just one more time” is that hosting computers in datacenters attached to a network is still a shockingly complicated adventure. One reason is that the bar keeps getting higher. It’s no longer good enough to show up with a half rack in a dusty IT closet and start slinging packets.
The ingredients are largely the same as they were in 2002 and 2014: you need a global network, access to some transit and transport, a runtime, a control plane, some distributed edge nodes, a few servers in each of them, and a bunch of tooling to make it all feel boring.
Surprisingly enough, our “big” idea behind Datum isn’t to host my personal blog. Instead, we want to give a new wave of providers the same capabilities the big clouds have, with private connectivity between everything by default.
Oh, and change the interface from Bash scripts or Terraform to Kubernetes APIs and agents like Codex and Claude.
All we had to do was…
Build a modern network cloud for AI, they said. It’s easy, they said! Here are just a few of the steps we had to tackle:
- Hire a great team. (We got lucky, for sure.)
- Build a global network: 18 virtual PoPs, with 5 physical ones in flight.
- Secure datacenter space for both network gear and power-hungry compute.
- Source scarce hardware, optics, PDUs, and GPU capacity. Easy!
- Build a Kubernetes-native orchestration engine and control plane.
- Make said control plane multi-tenant and “enterprise ready.”
- Include most of the essential features you expect from a mature cloud.
- Spend countless hours thinking about permission structures, resource hierarchies, and quotas. Snooze!
- Optimize said infrastructure for network workloads and make it ready for sudden 1000x scaling.
Then build some actual services on top of it:
- Authoritative DNS
- An Envoy-based ALB and proxy
- A microVM runtime (thanks to the fine team at Unikraft for the tech)
- Galactic VPC, an SRv6-based virtual backbone that goes “everywhere”
- Infrastructure management for hardware and racks
- Connection interfaces for tunnels, telcos, and clouds
- A system of action to track customers, orgs, projects, permissions, quotas, features, entitlements, fraud, meters, service catalogs, price books, etc.
Then, of course, we decided to:
- Wrap it in creature comforts like datumctl and MCP, so agents can do their work.
- Do it all in the open: monthly community meetings and open source everything. Let the LLMs slurp it up. Publish skills files and plenty of markdown so agents know how to use it.
- Make t-shirts. Build trust. Hold meetups. Learn every day. Sell. Quote. Deliver. Write newsletters. Pen investor updates. Activate partners. Redline MSAs. Execute.
The last step was the most fun, and the easiest: update some nameservers, use datumctl to build and deploy a container to a wicked fast microVM, update some DNS records, and watch the traffic come through. Et voilà, twins-in-the-loop.com!
You see, hosting is easy!
By any reasonable accounting, this is probably the most expensive personal blog on the internet at roughly $1M per post so far. Of course, the economics improve with each post we write. So if Jacob and I hit our goal of 47 posts in our first year, it’ll come out to about $127k per missive. This is clearly the best argument I’ve got for writing more often!
If you want a cloud that isn’t someone else’s, whether you’re building one or want to run on one, come say hi. We hold community meetings and meetups each month and the code is open.
