It’s a F**king Mainframe.

This post was inspired by long time industry colleague Steve Chambers who is a solid follow across the board. Thanks Steve, for letting me take your context and run with it.

Every time I hear somebody talk about LLMs like we’ve entered some completely unprecedented phase of computing, I have to resist the urge to throw a chair.

Not because AI isn’t important.
It is.
Not because the hardware isn’t different.
It is.
Not because the economics aren’t changing.
They are.

But because underneath all the GPU worship, the prompt-engineering theater, and the endless parade of people pretending they just discovered centralized compute, we are watching the industry reinvent a bunch of old ideas and act like none of us were here the first time.

It’s a f**king mainframe.

Not literally, obviously. I know. Save your emails. I am aware an H100 cluster is not an IBM zSeries and that tokens are not COBOL transactions and that no one is submitting punch cards to Claude. Calm down.

What I’m saying is that the operating model is starting to rhyme in ways that should make old infrastructure people feel deeply uncomfortable, or deeply validated, depending on their mood and blood pressure.

Because look at what we’re building…

We are centralizing expensive compute behind tightly controlled interfaces. We are scheduling jobs. We are rationing access. We are enforcing guardrails. We are wrapping users in layers of abstraction so they don’t interact directly with the machinery. We are creating privileged operator classes who understand how to coax useful work out of a scarce, high-value system that most people are not trusted to touch directly.

Ya know… Mainframe shit.

And the funniest part is watching younger generations in tech talk about this like it’s some radical break from the past when half of what they’re celebrating is just time-sharing with better marketing and a leather jacket.

We’ve seen this movie and can spoil the ending.

When compute gets expensive enough, important enough, and scarce enough, the industry stops pretending everybody gets root and starts building hierarchy again.

That is exactly what’s happening here.

The personal computing era, the client-server era, the virtualization era, even the cloud era to some extent, all trained people to think the natural arc of computing is toward more direct access, more self-service, more democratization, more elasticity, more “just click the button and get a sandbox.” And sure, some of that is real. Some of it stuck. But those eras also tricked people into forgetting that centralization is not dead. It just waits for the economics to justify a comeback.

AI did that in record time.

Because the minute compute became both wildly valuable and painfully expensive again, we started dragging all the old patterns back out of the basement. Not by name, of course. God forbid we use old words for old concepts. That would deprive half the industry of an excuse to hold a conference. So instead of JCL batch jobs, now we have prompt chains, YAML workflows, and agent orchestration. Instead of time-sharing, we have rate limits, quota tiers, tokens-per-minute ceilings, and inference schedulers deciding who gets what and when. Instead of RACF and ACF2, we have RBAC, policy layers, content filters, guardrails, model permissions, and fifteen meetings about who is allowed to ask the machine certain questions in production. Instead of operators and sysprogs, we now have prompt engineers, AI platform teams, model ops people, and a new priesthood of weirdos whose entire job is standing between the users and the expensive machine translating intent into acceptable ritual.

Same religion. New vestments.

And before somebody gets cute and says, “well actually, this is distributed,” yes, parts of it are. Of course they are. That is not the point.

The point is not whether the hardware is centralized in one box or spread across a cluster or delivered through an API endpoint fronting a hyperscaler region full of scorched-earth capex. The point is that the control plane is centralizing. The trust model is centralizing. The governance model is centralizing. The economic gatekeeping is centralizing. The user relationship to compute is centralizing.

You are not the operator. You are the requestor.

That is a very different world than the one a lot of people think they’re living in. And honestly, this should not surprise anyone who has spent real time around infrastructure. Because this is what always happens when the machine matters more than the user’s illusion of freedom. Think about it… when the underlying resource is cheap and abundant, the industry sells decentralization, experimentation, self-service, and creative chaos. Everybody gets a sandbox. Everybody gets admin rights. Everybody gets to spin up nonsense until finance notices.

When the resource becomes strategic, scarce, or expensive, the adults come back into the room…

Suddenly there are policies.
Suddenly there are queues.
Suddenly there are quotas.

Suddenly there are approved workflows, approved interfaces, approved use cases, approved abstractions, approved operators, and approved language for talking to the machine.

That isn’t innovation failing. That’s economics asserting itself. And AI is an economics story every bit as much as it is a technology story. Actually more, if we’re being honest.

This whole sector is trying very hard to sell the fantasy that intelligence is becoming ambient, infinite, and frictionless, while the actual infrastructure underneath it is screaming the exact opposite. Power constraints, memory bottlenecks, scheduling pressure, bandwidth limitations, inference cost, latency tradeoffs, GPU contention, utilization nightmares, and every other ugly physical reality are forcing the system toward control.

Not freedom. Control.

Again: F’ing MAINFRAME.

This is also why I laugh every time somebody says prompt engineering is fake. Maybe the title won’t survive. Fine. Most titles don’t. But the function absolutely survives. There will always be a class of people whose job is to understand how to structure requests, workflows, permissions, and execution context so the expensive centralized system produces reliable value for the organization. We can call them prompt engineers. We can call them AI operators. We can call them platform people. We can call them sysprogs with emotional support hoodies.

I don’t particularly care.

The function is real, because the machine is real, and the rules around the machine are getting more rigid by the month. That’s not a temporary glitch. That’s the shape of the thing. And honestly, that should be good news for a lot of old-school infrastructure people. Because one of the dumbest narratives floating around right now is that AI somehow invalidated decades of experience in systems, networking, storage, scheduling, multi-tenancy, governance, and resource control. As if those instincts suddenly stopped mattering because the front end got conversational and the cap table got weird.

Bullshit. If anything, those instincts matter more now.

The people who understand contention, queuing, access models, workload isolation, cost discipline, observability, and the political reality of shared infrastructure are looking at AI and seeing exactly what it is: not magic, not sentience, not a civilizational break from computing history, but a violently expensive shared system that is rapidly accumulating layers of policy, abstraction, and ritual around it.

That should sound familiar. Because it is.

And that is the part too many people are missing. This is not some moment where the old guard gets swept into the sea because a new acronym showed up wearing a leather jacket. This is a moment where a lot of the people who built the modern datacenter should be realizing that they are suddenly useful in a very specific, very high-leverage way.

Not because everything stayed the same. But, because enough of the hard parts stayed the same. Scarcity stayed the same. Contention stayed the same. Tradeoffs stayed the same. Architecture still matters. Schedulers still matter. Memory still matters. Power still matters. Cost still matters. And the second any of those things matter at scale, the people who understand infrastructure stop being legacy baggage and start being a competitive advantage. That’s the hopeful part of this, at least to me.

AI is not just creating a new class of software. It is dragging a whole generation of infrastructure problems back into the center of the conversation. It is forcing people to care again about the machinery, the constraints, the economics, and the ugly realities underneath the glossy demo. That is not bad news for operators. That is the moment operators become important again.

So yes, I’m going to keep saying it.
It’s a f**king mainframe.
Not because that is a perfect one-to-one comparison.
It isn’t.

Because it is the most useful insultingly-simple way to explain what is happening to people who still think AI is primarily a UX story. It is not primarily a UX story. It is an infrastructure control story wearing a conversational mask. And once you understand that, the whole thing starts to look a lot less mysterious.

The nouns changed. The instincts didn’t.

Which means a whole lot of people who thought they were being left behind may actually be standing exactly where the next wave needs them.

And that, to me, is a hell of a lot more interesting than the hype.

/Nick