The Big AI Divide

There’s a story that made the rounds last year that goes something like this: Grab Firecracker, wire up a few REST calls, ask an agent to scaffold a scheduler and an API, and by Sunday night you have something that boots microVMs on demand. Have an agent code a dashboard in front of it and it demos beautifully

And that’s true! The barrier to entry in virtualization has been significantly reduced. What used to be months, or more, of plumbing is now potentially a weekend or week project. Firecracker already exposes a clean REST API, and gluing a control plane onto an API is exactly the kind of work agents are good at.

Functionality got cheap…

Across the stack, the engineering cycles behind functionality — the features, the endpoints, the glue that makes a platform do a new thing — have been compressed; point an agent at a well-specified feature and the cost of shipping it has fallen off a cliff.

…production-grade did not

So what’s the problem? Basically, its everything else. What hasn’t moved is building one of these platforms to be production-grade, where scale means millions of VMs, reliability means five nines, and performance is measured at p50 and p90 rather than “it worked on my laptop”. Here, a weekend project need not apply.

And that’s before the unglamorous, non-negotiable parts of operating critical infrastructure like incident response and 24/7 on-call, SOC 2 Type II and HIPAA certification, change management, capacity planning, and disaster recovery.

Just as hard as it ever was

Put plainly: It is just as hard (or very nearly as hard) to build production-grade cloud infrastructure today as it was before AI. The difficulty moved to the three places AI can’t easily shortcut:

1. Performance: Performance is easy with a handful of VMs. It gets unreasonably hard when you’re running millions of them, across wildly different workloads, each one expecting the same latency whether the fleet is idle or experiencing a spike.

2. Reliability: Reliability for infrastructure has to be close to absolute. Five nines availability is roughly five minutes of downtime across the entire year. Every VM should run flawlessly, and the platform beneath them has to stay up and be resilient to failures.

3. Scale: Scale is when the properties that make a single microVM special have to keep holding when you’re launching millions a day. Keeping those numbers true for one VM is a demo. Keeping them true across the whole fleet is a product.

Where AI DOES move the needle for us

None of this means AI is a bystander in how we build. Beyond careful, human-in-the-loop development, there are two places where it has become completely indispensable:

Battle-testing at scale: We use it to batter and pummel the platform — generating adversarial load, pathological workload mixes, and stress conditions that would have taken much longer to hand-write. This is now a permanent part of our development cycle, and it has measurably improved the quality of our releases.

Incident triage. When something breaks, it’s fast at proposing what the root cause of a problem might be. It doesn’t always get it right, but curating a handful of plausible hypotheses is quick, and when it does land on the cause, it measurably shortens the path to a fix.

The lesson isn’t that AI changed nothing — it clearly changed the cheap half of the problem, and we’re grateful for it. It’s that it left the expensive half standing. Performance at scale, five-nines reliability, and the operational discipline behind them are still earned the hard way. That half is the one we do.