I keep seeing people talk about AI with extreme positions. There’s almost no middle ground left. Every take has to be one of these:

  • It’s going to end the world.
  • Its datacenters are going to destroy the world.
  • It’s the best thing ever.
  • It’s going to take all of our jobs.
  • It’s going to invent new jobs.
  • We won’t need jobs anymore.

Doom on one side. Utopia on the other. Both treat AI like magic or fate. And both let people off the hook from doing the actual work in front of us.

I keep coming back to History repeats. Humans forget.

We said the exact same things about server virtualization. Then, a decade later, we turned around and said the exact same things about Kubernetes.

We’ve seen this movie already

When server virtualization started taking over datacenters in the mid-2000s, the panic was real.

If your job was racking physical servers and installing operating systems by hand, virtualization looked like the end of your career. One person could spin up fifty virtual machines before lunch. What were the rest of us supposed to do?

We even had the exact same datacenter energy panic. In 2007, the EPA sent a report to Congress warning that datacenter power consumption had doubled in five years and would nearly double again by 2011. The grid wasn’t going to survive it.

What actually happened?

Neither extreme came true. Between 2010 and 2018, global datacenter compute instances grew by 550%, while total datacenter electricity use grew by just 6% (according to a 2020 study in Science). Why? Because virtualization and better hardware decoupled compute from physical boxes. Nobody wanted to pay the power bill for idle silicon. Economics forced efficiency.

The moment you read this, GPUs and LLMs will both be the least efficient they will ever be.

And the jobs didn’t disappear. Traditional rack-and-stack sysadmin roles flattened out, sure (BLS data backs that up). But making compute cheaper didn’t shrink tech. It meant we built way more software. That wave created modern cloud engineering, SRE, and DevOps.

Then came containers and Kubernetes. And we forgot everything we had just learned.

Once again, everyone ran to the edges. Either Kubernetes was going to automate operations completely out of existence, or YAML was going to burn every engineering org to the ground. For a few years, it was messy. People shoved tiny stateless apps onto giant clusters they didn’t understand. People got burned.

Then we learned what it was good for. We built guardrails. Now it’s just boring plumbing.

For a few years it was messy. For a few years, it will be messy.

The nuance

Is generative AI the exact same thing as virtualization or Kubernetes?

No. This is more revolutionary.

Virtualization abstracted hardware. Kubernetes abstracted infrastructure operations. Generative AI abstracts cognitive grunt work: drafting, summarizing, pattern matching, translating intent into syntax. Because it touches how we think and write, the emotional reaction is ten times louder.

And the physical buildout is massive. The IEA estimates datacenters sit at roughly 1.5% of global electricity use today, and AI is pushing that number up fast. That’s a real engineering problem that needs real solutions, not hand-waving.

But if you step outside the tech bubble for five seconds, the numbers tell a very different story about where we actually are.

According to consumer data from PNC Bank (reported by CBS News), only 2% of U.S. households are paying for a monthly AI subscription like ChatGPT Plus or Claude Pro. Google’s subscription coupling other services make it a little harder to add to the math. But we all can agree there’s no revolution in the suburbs just yet.

Compare that to a couple of other numbers from the same report:

  • 25% of U.S. households pay for streaming services.
  • 5% of U.S. households spend money on sports betting apps.

More than twice as many households are paying to bet on parlays as are paying twenty bucks a month for AI. Even inside the enterprise, as I wrote about last month, we aren’t seeing autonomous software factories replacing teams overnight.

We’re barely at the starting line. And while this wave is bigger, it’s going to follow a very similar path to the ones before it:

  1. Hype and panic: Loud predictions of utopia or ruin.
  2. Messy overuse: Shoving LLMs into every problem and racking up huge token bills and comprehension debt.
  3. Economic gravity: The bill comes due. Just like virtualization bent the 2007 power curve, economics is already pushing us toward smaller models, hybrid local setups, and leaner harnesses.
  4. Boring utility: It becomes standard equipment for people who know their craft.

So what?

Generative AI is a great tool. I use it every day. I spend my days at Google Cloud helping customers wire it into real workflows, and my nights running it across my own codebases with Leaf.

But it’s a tool.

I don’t know what the labor market looks like in 2040, and neither do the people writing manifestos on social media. Arguing about absolutes doesn’t help anyone ship working software on Monday.

We need to learn how to use it. We need to learn what it’s good for. And just as importantly, we need to learn what it’s bad at.

That’s what I’m focused on. Building tighter feedback loops. Defining clear contracts between human judgement and machine mechanics. Knowing when to hand work to an agent and when a twenty-line script is the smarter call.

History repeats. Humans forget. Ignore the noise at the extremes, pick up the tool, and figure out what it can actually do.