Who the Data Centers Are For

TL;DR

The risk I called “someday” arrived this month. In Renting the Frontier I priced out leaving the frontier and called vendor dependence a slow, structural risk. Three weeks of June turned it into a today problem.
The most capable model got pulled from the public and routed to institutions. Claude Fable shipped, the government killed it three days later, and its locked-down sibling was cleared for more than a hundred institutions and agencies. Read that order of events twice.
The flat-rate party is structurally over. The CEO selling you a $200 plan is on record that even $800 billion in revenue could bankrupt him. On Friday night, my own usage window quietly dropped by about half.
So I’m diversifying off a single vendor. Down to one Claude account, adding Codex, routing tasks through OpenRouter, and testing Chinese models that cost a tenth as much and pass many of the same benchmarks.
The job changed again, and it’ll keep changing. You can’t do this work passively anymore. Every three to six months the ground moves, and keeping up is part of the job now, not a side quest.
Ask who the data centers are for. We subsidize them in tax breaks and power bills, and we aren’t being sold the same tools the institutions get. That gap is the whole story.

Six months ago, burning tokens was the flex. Power users showed off their consumption, leaderboards ranked it, and nobody was telling you to use less.

Now the same behavior gets a different name. Tokenmaxxing. It’s almost a slur. Last week CNBC ran a piece about users moving “from tokenmaxxing to efficiency,” which is a polite way of saying the all-you-can-eat era is being walked back. We went from celebrated to scolded in two quarters. And the people who changed their minds weren’t us. We’re using these tools the same way we always did. The economics underneath them changed.

I’ve been circling this for two months. In Stop Burning Tokens I started treating AI capacity like a budget you allocate instead of a tap you leave running. In Renting the Frontier I priced out walking away, decided the real risk wasn’t cost but control, and then admitted I hadn’t actually moved. This is the post where I move. Because in about three weeks, June took that someday risk and made it a today one.

The Month the Frontier Showed Its Hand

Start with the executive order, because the popular version of it is wrong and the real version is worse.

The story going around is that the government paused AI development for ninety days. It didn’t. What President Trump signed on June 2 was Executive Order 14409, a cybersecurity order that asks frontier developers to voluntarily hand the government up to thirty days of early access before a major release. It explicitly creates no license and no pre-clearance. Here’s the part worth sitting with. There really was a ninety-day version, an earlier draft, and it got cut to a voluntary thirty after Musk, Zuckerberg, and Sacks reportedly called the President to argue it would slow the United States against China. So the honest headline isn’t “the government paused AI.” It’s that the mildest gesture toward oversight got filed down by the largest players in the room. The board favors the people already winning, and they’ll lobby to keep it that way.

Then Fable. Anthropic shipped Claude Fable on June 9, the most capable model it had ever released to the public, alongside a more locked-down sibling called Mythos. Three days later, on June 12, the models were disabled for every customer after a government export-control order citing national security. Gone in seventy-two hours.

Watch where it went next. On June 26, Mythos was cleared for limited release to more than a hundred United States institutions and agencies, and Fable was reported to be returning shortly after. That’s the sequence, start to finish. The frontier shipped to the public, got pulled from the public, and came back for institutions first. If you want to know who a scarce thing is for, watch where it goes when there isn’t enough to go around.

And then the money. In a February interview on Dwarkesh Patel’s podcast, Anthropic’s CEO laid out the math of his own business with unusual candor. He wasn’t bragging. He was warning. If revenue lands at $800 billion instead of a trillion, he said, there’s no hedge on Earth that stops the company from going bankrupt against the compute it’s buying. Be off by a single year on the growth curve and it’s ruinous. The $50 billion figure you may have seen isn’t a profit target, it’s the infrastructure bill. Sit with that. The person selling you a subscription is telling you, out loud, that the numbers underneath it are brutal. So when your plan gets tighter, that isn’t a bug. It’s the model working as designed.

The Party Was Never Going to Last

I’m a heavy user, and I don’t pretend otherwise. Maxed out week over week, what I run would cost ten to twenty times my $200 at list API prices. The meter shows me the value of what I’m drawing down, and it dwarfs what I pay. I’m not going to act wounded about that. I was getting a deal that was never going to hold, and I knew it. I wrote as much a month ago.

I’ll even give them the fair version. List API prices aren’t what it actually costs Anthropic to serve a token. The real cost is a fraction of the sticker, so a $200 plan isn’t the act of charity the scariest arbitrage numbers make it look like. Fine. But here’s the thing that matters. They’re tightening anyway. If the flat rate were comfortable, they wouldn’t be quietly shrinking it. They’re shrinking it because the buildout underneath has to get paid for, and the consumer plan was never the customer that pays for it.

You can watch it happen in real time. This past Friday at 11pm, my plan changed. The note dressed it up as more access to the smaller Sonnet and Haiku models and a “simplified” plan. No real detail. But I can see the dollar-equivalent of work a four-hour window now buys me, and it’s about half of what the same window bought me on Friday afternoon. Same money, half the work, announced at eleven o’clock on a Friday night, which is exactly when you ship the thing you’d rather nobody read closely. The same stretch of days that cleared the frontier model for a hundred institutions is the stretch that halved my window. I won’t tell you those two facts are the same decision. I’m telling you they point the same direction.

So I’m Spreading the Bets

Here’s the part that’s actually useful. Less critic, more consultant.

A month ago I wrote about juggling three Claude accounts and flipping between them as each one hit its limit. I’m going the other way now. Down to one account, kept fresh, and the rest of the work spread across vendors so no single one of them holds my whole workflow.

I’m adding Codex. OpenAI built it to mirror Claude’s pricing almost line for line, $20, $100, and $200 tiers, which tells you these two are fighting for the exact same desk. Having a real second coding agent isn’t redundancy, it’s leverage. I’m routing more work through OpenRouter and paying by the task, sending the routine, high-volume jobs to cheap models and saving the expensive frontier for the work that genuinely needs it. That’s the same right-size-the-model discipline from Stop Burning Tokens, just spread across providers instead of inside one.

And I’m taking the Chinese models seriously, which a year ago I’d have waved off. They cost roughly a tenth of what the American frontier costs. One recent comparison clocked an hour of coding at about ten dollars on Claude and under fifty cents on DeepSeek. They don’t match the very top of the American frontier. There’s still a real gap of several points on the hard coding benchmarks. But they pass many of the same tests, and for a lot of ordinary work I honestly can’t tell the difference. At a tenth of the price, “good enough for most of it” is a serious offer.

The one I keep turning over is a co-op. I did the self-hosting math in Renting the Frontier, and for one person it doesn’t pay. A single machine sits idle between your keystrokes, which is what wrecks the economics. But a handful of engineers splitting one good box, or a rented node, is a different animal, because batched inference is built to serve many people from one machine at once. It isn’t quite realistic for me today. It also isn’t science fiction, and the price of the hardware drifts our way every quarter. The point of all of this isn’t to leave the frontier. It’s to never again be a hostage to it.

This Is the Job Now

Step back from the tactics, because there’s a bigger shift underneath them.

For almost the entire history of this work, you could learn a stack and ride it for years. Pick a language, get good, and your knowledge held its value for a long time. That’s over. The tools, the pricing, the limits, the best model for a given task, the whole posture of how you work, can change completely every three to six months. The thing that was right in the spring is wrong by the fall.

Which means staying current stopped being a side quest. It’s part of the job now. You can’t put your head down and do the work the way it was done last quarter and expect to still be good at it. I don’t love this. It’s genuinely tiring. I said as much last time. Set a timer and make yourself look up before the field laps you. But pretending it isn’t true is the one move that guarantees you fall behind. We’re in a job, for the first time I can remember, where the work itself rewrites its own description twice a year.

So follow where the scarce thing goes, because it tells you who all of this is being built for.

When Fable got pulled, it came back for institutions first. When the window tightens, it tightens on the consumer plan, not on the enterprise contract. The CEO’s own math says the flat rate can’t subsidize the buildout forever, so it won’t. None of this is a conspiracy. It’s a business model, stated out loud, and the consumer was never the customer it was built around. OpenAI’s enterprise revenue is already catching up to its consumer revenue. The frontier is turning toward institutions because that’s who can pay for what it now costs.

And we’re helping pay for it whether we use it or not. Federal policy now fast-tracks data center permitting and dangles loans, grants, and tax breaks at the biggest projects, by executive order. On the grid that serves the mid-Atlantic, data center demand helped drive a 76% jump in capacity costs, and household power bills are climbing across the country. Some of that increase is bad market design rather than the server farms themselves, and I want to be fair about that. But the direction isn’t in dispute. We’re subsidizing this in tax breaks and in our electricity bills, and the thing we’re subsidizing isn’t being built to put a frontier model in your hands.

That’s the part I can’t shake. It’s hard to disrupt anyone with a bow and arrow when the established players have rifles, and the rifles are precisely the part of the market that isn’t for sale to you. You can follow that worry a long way downstream, to what a widening gap does to a society and how that tends to get managed, and I do follow it there in my darker moments. But I don’t need the bleak version of the story to act on the plain one. A world where the best tools belong to the people who already won, paid for in part with public money, is a world that gets less fair, not more. That’s enough to take seriously on its own.

So I’m not waiting around to find out how the rest of it goes. I’m spreading my bets, keeping a hand in the cheap and open tools, and treating independence as something worth paying a small tax for. Not because the frontier isn’t genuinely useful. It is. But because I’d rather not be a hostage to a tool that was never really built for me.

Ask who the data centers are for. Then build like the answer is “not you,” because right now, it mostly isn’t. And honestly? I’m still figuring out the rest of it.

David Kerr is the founder of Kerrberry Systems. He builds custom software for businesses that don’t want to be a hostage to one vendor. Find him on LinkedIn or GitHub.