An Americanized Caste System

Building Walls to Support A Bubble

The AI bubble needs a caste system. Not in the way most people talk about bubbles, where someone points at a chart and says “this can’t last.” The AI bubble needs stratification the way a levee needs sandbags. Without it, the water gets in and the whole thing comes apart. That’s what the last three weeks have shown us. Not that the government is regulating AI for safety. Not that the tech companies are worried about security. But that both parties, for entirely different reasons, need to make sure you don’t get the same product that their friends get. And that the upstart competitors who might give it to you anyway get kneecapped before they can.

Two weeks ago the Commerce Department, under Secretary Howard Lutnick (who probably uses Grok), forced Anthropic to pull Fable 5, its newest frontier model. The first of the long-talked-about Mythos class. Anthropic said it was a “misunderstanding” regarding the model’s ability to be jailbroken. Amazon apparently ratted out Anthropic to the federal government. And Anthropic, having already angered an administration that doesn’t forget slights and doesn’t forgive them either, was in no position to fight. Fable was pulled. Pete Hegseth (definitely a Grok user) rejoiced on X. Then came the negotiations. And now, as of June 27, Mythos 5 has been partially restored but only to a set of companies on what Commerce is calling “Annex A”, a government-approved list of U.S. organizations that “operate and defend critical infrastructure.” Fable 5 itself? Still gone. Still restricted. Criminal and civil penalties if you try to get around it.

Then, the day before Mythos was partially restored, OpenAI dropped GPT-5.6 and its Sol, Terra, and Luna models. The flagship, Sol, is by all accounts a massive leap. And it is available to roughly twenty companies whose participation has been approved by the government. Not by OpenAI. By the government. Sam Altman, in an internal memo obtained by Axios, said the government would be “approving access customer by customer during this preview period.” He added that he hopes to release Sol broadly in “a couple of weeks.” In a blog post, OpenAI wrote that it doesn’t believe “this kind of government access process should become the long-term default.” What it didn’t do was refuse the process. It bent the knee, put out the requisite statement of mild displeasure, and kept it moving. The kabuki is almost ritualized at this point. A frontier model comes out. The government steps in. The AI company hems and haws but complies. Everyone moves on and waits for the next one.

Now, the stated reason for all of this is security. On June 2, Trump signed an executive order titled “Promoting Advanced Artificial Intelligence Innovation and Security” which, on paper, establishes a “voluntary framework” for pre-release engagement on frontier models. Voluntary. The order even includes language saying it shall not be construed to authorize mandatory licensing or preclearance. But “voluntary” is doing a lot of work in that sentence. The designation thresholds for what counts as a “covered frontier model” will be set through a classified NSA process. Developers won’t know where the line is until they’ve crossed it. And as Fable and Sol have demonstrated, the “voluntary” framework has a way of becoming functionally mandatory when the alternative is having your product pulled by the same people who tried to blacklist Anthropic on First Amendment grounds and got enjoined by a federal judge for it.

So that’s the picture on the government side. An administration that picks winners and losers as a matter of identity, not ideology, now has a mechanism to decide who gets frontier AI and who waits. The precedent is set. The pattern is established. If you’re in the circle, you get the product. If you’re not, you get a blog post explaining why it’s coming soon.

Then there is the other side of this arrangement. Because the tech companies aren’t just victims here. They have their own reasons to like this system, even when they publicly complain about it.

The thing the closed labs need you to not think too hard about is that the open-source models are catching up at a fast pace. OpenRouter data from late 2025 shows open-source models already accounting for roughly a third of all token usage. DeepSeek V4, released in March 2026, matches or exceeds GPT-4o on seven of twelve standard benchmarks. Meta’s Llama 4, Alibaba’s Qwen 3.5, and a handful of others are closing the gap on the most advanced closed models. According to a Hugging Face analysis, 89% of enterprises are now using open-source AI models, reporting 25% higher ROI than closed alternatives.

Eighty-nine percent of enterprises and twenty-five percent higher ROI. And the AI ecosystem is being built on the product with the lower ROI.

Now imagine you’re OpenAI. You’re valued at $920 billion. You’re projected to lose $14 billion this year and don’t expect profitability until 2029 or 2030. Your $200-a-month Pro subscription could represent $14,000 in API-equivalent usage if a user actually maxes it out. You’re heading for an IPO. And a set of free or near-free competitors is approaching your capability level with a fraction of the overhead. If those competitors reach parity, and they’re getting close, the pricing model that justifies your existence collapses. Not your technology. Your technology can stay and be useful for the future. Your business mode though, that goes bye-bye. The thing that makes a $920 billion valuation seem sane instead of delusional goes away. And it takes your shareholders with it.

So what do the closed labs do? They aim to kill the competition. .

Many of the closed labs have their own initiatives to examine open-weight models and flag “security risks.” They’ve banded together in the Frontier Model Forum to study security concerns in open-weight systems and to raise the alarm about Chinese models training on data from the closed labs. OpenAI told Congress in February 2026 that DeepSeek was engaged in “ongoing efforts” to exploit capabilities developed by U.S. labs, claiming to have detected accounts associated with DeepSeek employees using obfuscated methods to distill OpenAI’s models. Anthropic followed up, reporting “industrial-scale” campaigns by three Chinese labs generating over 16 million exchanges with Claude through about 24,000 fake accounts.

Set aside for a moment the sheer hypocrisy of companies that built their entire businesses on scraping the open internet now complaining about someone else using their data. The question now is, what happens when a small cabal of industry leaders gets legitimacy to look at rivals’ products and tell a receptive government they’re security risks? To a government that has already shown, repeatedly, that it is more than happy to throw its hefty weight around?

Those open-weight models wouldn’t stand a chance. And if they go, so does the competition.

Here’s why that matters beyond the tech industry. The AI bubble isn’t just about inflated valuations. It’s about artificial scarcity preserving those valuations.

In June 2026 alone, the Nasdaq fell over 4% in a single session. Semiconductor stocks shed more than $1.3 trillion in market value. The Philly Semiconductor Index dropped 7.9% in one day. Then on June 24, the Nasdaq fell another 2.2% as a global tech sell-off deepened. Combined capital expenditure across Microsoft, Alphabet, Amazon, and Meta has now exceeded $452 billion in 2026, and investors are starting to ask a question that should have been asked two years ago. Where is the revenue?

A National Bureau of Economic Research survey of 6,000 executives found that while 70% are using AI, 80% report no impact on company productivity. An MIT study found that 95% of businesses invested in AI have not made a profit from it. The Bank of England has warned of a global market correction due to possible overvaluation of AI firms. The Shiller price-to-earnings ratio for the U.S. market has exceeded 40 for the first time since the dot-com crash.

These aren’t fringe concerns. Fortune is openly comparing 2026 to 1999. And if you read Part 1 of this newsletter’s series on circular funding, you know that the revenue the big tech companies are reporting is, in many cases, money they loaned to each other and booked as sales on the return trip.

So the tech companies and the administration have different but complementary reasons to keep this system in place. The tech companies need stratification to protect pricing. If open-source reaches parity and anyone can run a frontier-class model on commodity hardware, there is no justification for a $200-a-month subscription, no justification for a $920 billion valuation, no justification for the $452 billion in capex. The house of cards needs the moat, and the moat is access control.

The administration needs stratification because gatekeeping is leverage. Deciding who gets frontier AI is a new form of patronage. It is, in the way we’ve said before, crony capitalism with a neural network. The government doesn’t care whether the models are safe. If it cared about safety, GPT-5.5 would be restricted alongside Fable 5, not sitting on the open market while the administration selectively punishes a company whose CEO wouldn’t take a meeting with the Pentagon on their terms. The security rationale is the means, not the motive.

Neither party is explicitly coordinating with the other. They don’t have to. Their interests align naturally. The tech companies get regulatory cover for keeping competitors out. The administration gets to hand a new class of economic spoils to its friends. And the result is a tiered system where access to the most powerful AI is determined by your relationship to power, not your ability to use it.

We’ve been told for years that AI would cause huge economic disruptions. Eliminate jobs. Reshape industries. And in the face of that, some people started building businesses with AI. They saw its usefulness, leveraged it, created space for themselves in an economy the elites had long touted as being transformed by the greatest technology since the internet itself. They used the tool.

And the reaction of the elite class and the administration? Shut it down. Restrict the best products to the billion-dollar companies in the circle. And if the masses try to move to the open-weight models, the ones that are free and closing the gap? Use the same argument they made for Fable and Sol. Too dangerous and too much of a threat.

There is one thing that threatens both parties’ interests at the same time. A capable, open-weight model that anyone can run. That’s the exit door from the high token costs.. It’s the thing that collapses the pricing model, renders the gatekeeping irrelevant, and makes the patronage system toothless. And both the tech companies and the administration know it. That’s why the open-weight models are the real target. Not because they’re dangerous. Because they’re a threat to a system that needs you locked in.

There is often a need for regulation in modern society. Water should be clean. Cars shouldn’t explode. Seat belts are good. And AI is likely to need some regulation based on what it can do and the impact it will have. But what’s happening here is not regulation for the general wellbeing. This is regulation to keep a bubble inflated and a patronage system fed. It balances the needs of the two parties. The tech companies whose valuations depend on artificial scarcity and an administration that treats every lever of government as a tool for rewarding loyalty. The general public, whether they’re building with AI, paying $200 a month for a subscription, or have no involvement at all, is irrelevant to the arrangement.

An Americanized caste system. Americanized in that a regular caste system would see no movement. In America, if you have enough money and grease the right palms and egos, you have the potential to move up. But for the greater number of people, they’ll be stuck. Their jobs will be displaced by the tool the elites love. And if they want to use that tool, learn it and grow from it, they’ll have a ceiling. The best products will be behind a wall that protects the product, the well-connected, and the bubble they created. Maybe you’ll be lucky enough to be in the club or find a way in. Unless the open-weight models survive and people embrace the exit door before it gets locked.