A few weeks ago I wrote about what building Skyforge taught me about network emulation and digital twins. If you haven’t read it, you can find it here:
The response was much larger than I expected, but one conversation afterwards stuck with me.
An colleague mentioned they were starting to hear a different question from customers. It wasn’t really about digital twins anymore. It was about AI.
If tools like Claude can generate software this quickly, doesn’t that fundamentally change the build-versus-buy equation?
It’s a fair question, and honestly, it’s one I’ve been asking myself as well.
After all, I spent the last six months building Skyforge with Claude sitting beside me. I don’t mean that as a marketing statement about AI. I mean it literally. Large parts of the platform, from Go services and Kubernetes manifests to GitOps workflows, deployment automation, documentation, and countless little pieces of glue code started life as conversations instead of blank files.
There is absolutely no question that Claude changed how quickly I could build software.
Looking back, though, I don’t think it changed where most of the engineering effort actually went.
AI Changed the Cost of Building Software
If you’d asked me a year ago what slows software projects down, I probably would have answered “writing software.”
Today I don’t think that’s true anymore.
Once Skyforge’s overall architecture settled down, writing new code became surprisingly inexpensive. Claude could generate a new service, refactor an API, write tests, or build out another Kubernetes workflow much faster than I could have done manually. That didn’t eliminate the need for engineering judgment, but it dramatically reduced the amount of time spent translating ideas into working code.
I honestly don’t think Skyforge would exist in its current form if I’d been writing everything manually.
That’s a real shift, and I don’t think we’re going back.
The Bottleneck Moved
What surprised me was how quickly writing code stopped being the hard part.
Once a feature existed, I still had to integrate it with everything else. I still had to decide whether it belonged in the architecture. I still had to understand how it behaved under failure conditions, document it, test it, maintain it, and occasionally redesign it after learning something new.
One thing that became obvious fairly early was that every new network operating system effectively became another product that needed ongoing maintenance. New images needed testing. Existing deployment logic needed validation. Readiness checks evolved. Supporting another platform wasn’t simply another feature. it created another long-term responsibility.
Claude made writing that code dramatically easier.
It didn’t make owning it dramatically easier.
That’s Where Build vs. Buy Changed for Me
Before Skyforge, I thought build-versus-buy was primarily about development cost.
How many engineers would it take?
How long would it take?
Could we realistically build it?
Today I think those are still important questions, but they aren’t the first questions I’d ask.
The better question is whether you want to own the software after it’s built.
Owning software means maintaining it. Testing upgrades. Validating dependencies. Responding to bugs. Improving documentation. Supporting users. Making architectural decisions years after the original implementation shipped.
AI helps with every one of those activities.
It doesn’t eliminate any of them.
Looking Back
Building Skyforge changed the way I think about AI, but probably not in the way I expected.
AI absolutely changes the economics of writing software. I can build things today that simply wouldn’t have been practical for me to tackle a few years ago.
What it didn’t change is the responsibility that comes with operating software over time.
That’s the lesson I wasn’t expecting.
If someone asked whether they could build a platform like Skyforge today, my answer would probably be yes.
If they asked whether AI makes that decision trivial, my answer would be no.
Building the first version has never been easier.
Building something people trust, depend on, and expect to work every day is still the hard part.
That’s the part I think gets lost in a lot of the AI discussion, and it’s probably the biggest thing Skyforge taught me over the last six months.
