Sam Lambert posted: “Thanks to AI we are about to enter the golden age of infrastructure.” I’d already been writing Space Cowboys when I saw it. He’s right. And things are about to get very interesting.
The Cost of Software Is Going to Zero
Every week I see another demo of an AI agent building a full app from a prompt. A landing page. A CRM. A blog. A dashboard. The quality is getting good enough to ship. Not good enough for every case, not yet, but good enough for a lot of them.
When software is cheap to produce, the things that are easy to build become impossible to defend. A hosted blog platform, a basic CRM, a project management tool. If an AI can generate a 20% solution that fits your exact needs, why pay for someone else’s 80% solution with features you’ll never use? And if the barrier to building drops far enough, you don’t get one competitor. You get hundreds. All building the same thing, all racing to the bottom.
This is the red ocean. Bloody, crowded, commoditized. Every SaaS product that’s primarily a CRUD app with a nice UI is about to discover how shallow its moat really was.
The Foundations Nobody Bothered to Build
While everyone was chasing the application layer, the infrastructure underneath got DoorDashed. Companies outsourced it, abstracted it, and forgot how to cook.
At Cloudera I spent years thinking about HDFS at scale. Rack locality. How the top-of-rack switch was interconnected. We’d stripe disk sizes and manufacturers across machines to avoid correlated failures. Backblaze has written about this for years. Same model, same batch, same age, they tend to fail together. You don’t lose a rack all at once. You lose drives across multiple machines in the same week, and if you didn’t plan for it, your replication math stops working. Two copies of a block on the rack, one copy off. Every decision was deliberate because the physics of the hardware dictated the architecture of the software.
The people who would catch this stuff, who would look at the architecture and say “this is insane, let me show you why,” they still exist. They work at Google, AWS, Backblaze, PlanetScale, Oxide, Unikraft. They’re not gone. The problem is we stopped making new ones. They weren’t invisible. They were treated like plumbing. Treated like infrastructure. The industry quit investing in that pipeline fifteen years ago, and the bill is coming due.
Where the Moat Actually Is
If AI makes the easy things trivially easy, the only defensible position is doing the hard things.
Think about what early VMware engineers had to solve. The x86 architecture had seventeen instructions that were sensitive but didn’t trap when executed outside ring 0. They just silently did the wrong thing. You couldn’t use the standard trap-and-emulate approach because the hardware wouldn’t tell you when the guest OS was doing something dangerous. VMware’s answer was binary translation, scanning kernel code before execution and replacing those instructions with safe call-outs, while simultaneously maintaining shadow page tables that intercepted every memory operation. All of this on hardware that was never designed to support any of it.
Or think about why Google put atomic clocks in their datacenters for Spanner. Clock synchronization sounds like a solved problem until you need to guarantee causal ordering across continents. NTP gives you 100-250ms of skew. That means a transaction on one node can get a timestamp that puts it before a transaction that causally happened after it. The whole database returns impossible states. Google’s insight was that you could turn a coordination problem into a latency problem: give every node an atomic clock, bound the uncertainty to under 7ms, and make every write wait out that interval before reporting success. The hardware isn’t a luxury. It is the algorithm.
Or look at what my friend Richard Crowley did at PlanetScale. They ripped out network-attached storage and went back to local NVMe drives. 220,000 IOPS instead of 40,000. P99 latency cut in half. Lower cost. The cloud’s abstraction was actively making their database worse, and the only reason they knew that was because they had engineers who understood what was happening underneath it. That’s the moat. Not the code. The judgment to look at EBS and say “no, we need the actual hardware.”
Those people are rare. Like gold.
And those people are about to inherit the earth. This is where Lambert’s golden age comes from. The demand for real infrastructure is about to spike, and the supply of people who can build it hasn’t grown in fifteen years.
The Meek Shall Inherit the Earth
The red ocean is here. Commodity software is going to drown in AI-generated alternatives. Margins collapse. Moats evaporate. If your entire product is a CRUD app with a nice UI, good luck.
The golden age is also here. The people with deep systems knowledge, the ones we spent fifteen years calling cost centers, are about to become the scarcest and most valuable engineers in the industry. The companies that kept them, that built around them, are sitting on a competitive advantage that no amount of AI tooling can replicate.
AI makes learning easier. It should. Use it. The pipeline of systems engineers needs to grow again and AI can help with that. Experience is something else entirely. The difference between someone who has read about building a datacenter and someone who has built one is not information. It’s judgment. It’s the scar tissue from fifteen years of production outages and hardware failures and architectural decisions that aged well because someone understood the physics. You cannot prompt your way to that.
Buckle up, buttercup.