The gate of the Vinaya sect: the plaque faces out, the precepts face in.

Prologue: The night the rations were cut

February 27, 2026. Washington.

A list was updated. It is called the supply-chain risk list. The last time anyone remembered it, every name on it belonged to an adversary: companies judged to be on the other side, untrustworthy, to be cut off.

That day, an American company appeared on it for the first time.

Anthropic. San Francisco. The makers of Claude.

The order came fast: every federal agency stops using it immediately, six months to wind down, no loose ends. The justification, in public remarks, leaned on heavy words. Arrogant. Letting ideology stand in front of American soldiers.

Seventy-two hours later, the sky over Tehran lit up.

The joint U.S.–Israeli strikes had begun. According to The Wall Street Journal and Axios, Central Command fought that campaign using Claude: intelligence assessment, target identification, simulated battle scenarios.

Using it. The very thing that had just been blacklisted by its own government.

A company chose to be struck off by its own state rather than release the two vows in its hands. Seventy-two hours later, its sword flew anyway.

That is what this piece is about. The answer isn't whether the company got worse. It's where the sword is stuck.

1. Leaving Shaolin

Start with where he came from.

November 2014, Silicon Valley. A 31-year-old walks into a lab Baidu had just opened there, to work on speech recognition under Andrew Ng.

His name is Dario Amodei. Physics at Stanford, then a PhD in biophysics at Princeton, then a postdoc at Stanford's medical school studying cancer biomarkers.

In other words, he came into AI sideways.

The lab was Baidu's Silicon Valley AI Lab, founded in May 2014. Ng had a $100 million budget and a room full of GPUs, and he was recruiting worldwide. Dario was spotted by the lab's technical lead, Greg Diamos — by one account, hired on the spot.

He stayed a year. The project was Deep Speech 2, an end-to-end speech system that handled both English and Mandarin. The paper made MIT Technology Review's list of ten breakthrough technologies for 2016. Somewhere in there he also picked up a "Star Employee of the Quarter" award, Q2 2015.

Later, a story circulated on the Chinese internet: that he had been "an intern at Baidu," that he had a bad time, and that this is why he resents China.

The story fails on the facts. He was not an intern; he was a research scientist. He reported directly to Andrew Ng. That stint is the longest single entry on his LinkedIn, and it is all results. When he talks about those years, he talks about "a $100 million budget," "a lot of GPUs," "research conditions with almost no cost constraints."

And here is the giveaway. The "Baidu intern" meme actually belongs to Yan Junjie, founder of MiniMax, who really was an intern at Baidu. A story about A got transplanted onto B, and then used to explain B's behavior. A very ordinary narrative drift.

But what Dario actually took away from Baidu matters more than an award.

It was there that he first felt something: stack more layers, add more data, add more compute, and performance climbs. Smoothly. Predictably.

He didn't dare push the thought. Maybe it only works for speech.

In 2017, looking at GPT-1's training results at OpenAI, he understood: that wasn't a law of speech. That was the master principle of the whole discipline.

Today we call it the scaling law.

So the thing he would later build began, in a real sense, in China. He wasn't driven out of the Central Plains. He learned there that mass produces power — and that is what took him to Shaolin.

Shaolin, in this telling, is OpenAI — the mountain gate raised after that 2015 gathering at the Rosewood. The feud is covered in an earlier piece in this series. Here, one line is enough: seven people walked down the mountain in the winter of 2020.

The jianghu version is that they fell out over whether to make money.

No.

They fought over something else: who gets to rewrite the rules.

In November 2023, Shaolin did something that shook the whole jianghu. The abbot was removed by the board. Five days later he was back, and the board had been reconstituted overnight.

Outsiders read that as gossip. Inside the Vinaya sect, they read it as a verdict.

It proved one thing: the rules can be rewritten by one man, in five days.

So they made a decision. Stop writing vows into people's hearts. Write them into the deed.

They never stopped believing in the ideal. They stopped believing in an ideal with no deed behind it.

2. Vows carved into the deed

In 2023, alongside its Series C, Anthropic rewrote its charter.

They built four things.

A field of merit: a charter that lets you not maximize profit

A Delaware public benefit corporation. A normal company's directors owe their duty to shareholders' money. A PBC's directors are legally permitted to weigh shareholders' interests, the interests of those materially affected, and the public benefit written into the charter — "the responsible development and maintenance of advanced AI for the long-term benefit of humanity."

Put plainly: it is allowed to not maximize shareholder value.

A precepts hall: five people with no shares

The Long-Term Benefit Trust. The company issued a single Class T share, held only by the trust. Five trustees, all financially disinterested, barred from holding stock, paid $400,000 a year. They elect and remove four of the seven board seats.

An ancestral tablet: one share that holds half the world

The seven co-founders pooled into a Founder LLC holding a single Class F share: 50.1% of all voting power, locked, undilutable. Class A shares in public hands can never add up to more than 49.9%.

Giving it away: pledging eighty percent

Before the IPO, roughly 5% of diluted equity — some 846 million Class N shares — was moved into a public benefit fund. And each of the seven founders signed a pledge to give away 80% of their personal equity.

Other sects swear to heaven. This one nailed its vows into paper.

The jianghu calls this idealism. I think the word is too light.

This is not idealism. This is an autopsy report.

A way of living, revised inch by inch against a death they had personally watched. The death was this: a nonprofit becomes a tech giant's subsidiary inside four years; a mission statement loses to a board vote; a promise not to commercialize gets overwritten by the phrase "competitive necessity."

So they stopped trusting "we agreed." They trust only "it can't be changed."

But there is one line missing from that deed.

Who are the five trustees?

A former Federal Reserve chair, Ben Bernanke. The head of a national-security think tank, Richard Fontaine of CNAS. A global-health executive, Neil Buddy Shah of the Clinton Health Access Initiative. A Harvard experimental physicist, Christopher Stubbs. And one seat, empty.

Five seats in the precepts hall: four occupied, one empty.

Four people spanning macroeconomics, national security, public health, and basic science. On paper, comprehensive.

Now close the door and ask: is any of them from outside that circle?

No.

So in the phrase "the long-term benefit of humanity," the humanity in question is institutionally assumed to mean "the people we know."

There's no malice in that. It's a social radius. A network built from San Francisco, Washington and the Ivy League picks the people it believes best represent the species.

And that is exactly what determined what the expulsion order would look like three years later.

3. The incense ledger

Keeping vows costs money.

The jianghu doesn't like admitting this. It happens to be true.

When Anthropic was founded in 2021, it had $124 million. Of the $580 million raised in April 2022, $500 million came from an exchange called FTX. Months later FTX collapsed and its founder went to prison, and that money became a liability with a fuse.

A sect whose central vow was that AI must not harm humanity nearly ran out of money because of a fraudster's cash.

Two patrons saved it. Amazon, $4 billion, September 2023. Google, $2 billion, the following month. Both are simultaneously shareholders, compute suppliers, and sales channels.

That is the material basis of the Vinaya sect's discipline: someone else pays for the incense.

But the incense is bought on credit. By the time of the S-1, the ledger looked like this:

Patron	Commitment	Runs through
Google	at least $114.5B	2033
Amazon	at least $110B	2036
Microsoft	at least $44.9B	2033
xAI	up to $120.5B	2029
Broadcom (equipment leases)	about $161.2B	five-year term

Committed spending runs to 2045.

In the first half of 2026 it booked $16.29 billion — its best year ever, on paper. It also booked an enormous loss.

That isn't moral criticism. It's physics.

When a sect must double every year, must win the newest generation of compute, must tell a growth story under the market's gaze — the price of every no rises at a visible rate.

So: which vow is cheapest?

The prospectus contains a number almost nobody noticed: roughly 5% to 7% of revenue came from customers in violation of its terms.

Those customers are the ones later expelled from the mountain gate.

I'm not going to write this up as a scandal. Five to seven percent means giving up that market hurts on the page but doesn't kill you. It lands exactly on the affordable side of the line.

That sentence is true of everyone, including genuinely honest people.

4. The expulsion order

September 4, 2025. An expulsion order is nailed to the Central Plains gate.

What makes it brutal isn't what it forbids. It's what it inspects.

It doesn't check where you use the model. It checks who paid for you: wherever you're incorporated, if more than half your equity is held — directly or indirectly — by companies from unsupported regions, you're out, including overseas subsidiaries.

In the announcement's own words: aligning the terms of service with real-world risk.

On February 23 of the following year, it named three companies: DeepSeek, Moonshot, and MiniMax. The charge: roughly 24,000 fraudulent accounts and more than 16 million exchanges used to extract its techniques piece by piece.

To be clear: every one of those numbers comes from Anthropic's own disclosure. No third party in the jianghu has audited the ledger.

There are three ways to read this, and you need all three.

First: the court's mood

The announcement's actual language — "adversarial nations," "authoritarian regions," "advancing democratic interests," "U.S. and allied strategic interests" — is not the language of whether AI will harm people. It's the language of state against state. In 2026 Washington, an AI company that stays vague on this has a very bad afternoon in a Senate hearing.

Second: protecting the techniques

Was the extraction real? Very likely — model distillation is no secret in this jianghu. But notice something: anti-distillation is simultaneously safety and business. At this exact point the two overlap completely.

And the place where the two overlap completely is the place most worth suspicion. Not because it must be false, but because you cannot use its sincerity to verify it.

Third: the doctrine extending itself

Almost nobody says this one, and I think it's the truest. The sect has exactly one root precept: the deployer must be visible and constrainable. A deployer whose ultimate controller you can't trace, who can legally be compelled to hand over data, whom you can't restrain beforehand or hold accountable afterward — that is the limiting case of the precept.

Which is why the method looks like obsessiveness: pierce the cap table, chase the ultimate beneficiary, hold subsidiaries liable.

It isn't hatred. It's a compulsion about traceability of control. It's the grammar they know best.

And that precept has a fatal blind spot.

It only recognizes "can this party be watched." It has no cell in its table for "a Central Plains sect with a life of its own."

In that grammar there are exactly two kinds of actor: auditable, and non-auditable. And non-auditable is automatically an extension of the state. So a company struggling to survive in the gray zone must be folded into the court's hand.

That is the real source of the "unflattering picture of China." Not hostility. A missing cell in the table.

And something else is worth saying: the ruler only measures in one direction.

America also has laws that can compel companies to hand over data, and long-arm jurisdiction that reaches across borders. That reasoning has never once been applied to American firms. If the ruler were truly neutral, then every European bank, every South American government, every Global South customer running a U.S. model would be scored the same way.

They aren't.

So this is not neutral risk management. It has a hierarchy of trust baked into it.

5. The jobs it won't take

Back to February 2026.

Let's be clear about one thing first: this sect is not "hands off the blade." Its hands are deeper in than most.

In November 2024 it built a channel with Palantir and AWS, putting Claude inside U.S. intelligence and defense agencies. In June 2025 it shipped Claude Gov for classified government environments; by then, per Ars Technica, it was running at multiple national-security agencies.

It drew exactly two lines: no mass domestic surveillance in the United States, and no fully autonomous lethal weapons.

In January 2026, the War Department asked for both lines removed.

Talks deadlocked in mid-February. The department wanted "all lawful use" — you don't tell a buyer that a whole category is off the table. The ultimatum: sign by February 27, or the contract ends, and the Defense Production Act may be invoked to compel access without one.

On February 26, Dario refused publicly.

On February 27, Secretary of Defense Pete Hegseth designated the company a supply-chain risk.

The same day, the President ordered every federal agency to stop using it, with six months to wind down, warning of serious civil and criminal exposure for non-cooperation.

It was the first time this company was handled by its own country as an adversary.

Why did the court need its nod so badly?

Because it isn't an ordinary supplier. It's the one that says no.

The evidence arrived immediately. Hours after the designation, OpenAI announced a new framework agreement with the Defense Department — nominally keeping similar restrictions, with a carve-out for national-security necessity. xAI signed something similar the same day.

Same money. One took it. One didn't.

After February 26, several hundred employees at Google and OpenAI signed an open letter titled "We Will Not Be Divided," arguing that blacklisting one company was a way to force the rest into line.

In March, Anthropic sued the Defense Department. On March 26, federal judge Rita F. Lin granted a temporary injunction. One line from the order keeps getting quoted: the designation "appears to be classic First Amendment retaliation."

It won. A sect took its own government to court, and won.

At this point you have to concede: the vows are real. This isn't branding, isn't PR. It was bought with real money and the entire federal budget.

But this is also where you can see exactly how the lines were drawn.

It refused "fully autonomous lethal weapons" — machines deciding who dies. It did not refuse "target identification" — a machine telling a human whether a coordinate is worth hitting.

The gap between those two lines is technically real, morally decisive, and commercially convenient.

Because it happens to sit exactly where the business survives.

This is not hypocrisy. A hypocrite doesn't lose a Pentagon contract over a clause, get blacklisted by his own government, and then go fight a case he might lose.

This is pricing. A sect weighing which clause can bend — bend this one, what's left of me? — and which cannot, because bending it means I am no longer myself.

It is not a saint. It is doing its accounting, and then calling the most accurate part of the arithmetic a matter of faith.

The reverse is also true: every saint who never did the arithmetic is dead by now.

6. The sword you can't pull out

February 28. The strikes on Iran.

Per The Wall Street Journal and Axios, Central Command used Claude during the campaign: intelligence assessment, target identification, simulated battle scenarios.

Some people called this an embarrassment. I don't think it is. It's the coldest fact in this piece, and far more serious than embarrassment.

The sword was already in the armory.

The day the sword went into the gears, pulling it out became dearer than leaving it.

Months earlier, per The Wall Street Journal, the U.S. military used it in the operation that captured Maduro. Eighty-three people died in that one, two of them civilians.

Once a tool is part of the command system, replacing it takes months. At that point "the supplier refused" stops being a switch.

It becomes a fine.

You can be fined, struck off, sued, threatened with criminal liability. The system runs anyway, because pulling the sword out costs more than leaving it in.

Which is why "the brake of the AI industry" strikes me as a category error.

Brakes go on cars. Not on roads.

Anthropic can slow itself down. It cannot make OpenAI go slower, cannot stop the Pentagon from using it, cannot stop another company from signing the contract it refused three months later.

It can brake its own car. Everyone is driving on the same road.

7. Opening the gate

October 26, 2026. Anthropic files its S-1.

Nasdaq, ticker ANT. Market expectations for the valuation run as high as $2 trillion.

If that number holds, it will be the largest IPO in history.

The most interesting part of the filing isn't the financials. It's the share structure. Five classes.

Class A to the public, one vote each. Class N to strategic partners like Amazon and Google, no votes. Class E to employees. Class F to the seven founders: 50.1%, locked. Class T to the Long-Term Benefit Trust: four board seats.

Economic rights, control rights, governance rights — separated three ways. On paper, the vows look impregnable.

The gate opens, the pilgrims come in, and the ledger is laid out on the desk.

That's not about share classes. It's about rhythm.

After listing, the thing that constrains it changes. It used to be the precepts — stopping points it wrote for itself. Now it's the quarterly report — every three months, the market asks why you weren't faster.

Quarterly reports are more diligent than precepts. A precept is recited once a year. A quarterly report is recited four times, each time with someone listening through a magnifying glass.

One thing should be recorded in its favor, though.

The filing states: Claude will stay ad-free.

In a year when every AI assistant is figuring out how to wedge advertising into the chat box, putting "no ads" into a legal document handed to investors is a small thing and a large thing. Because it means: some future day, when ad revenue becomes too big to ignore, it will have to answer for those words.

You can count on one hand the sects in this jianghu that leave money on the table.

In the same period, its transparency report carries another number. In the first half of 2026 it banned 11.4 million accounts. About 63,000 a day. Nearly eight times the previous half-year.

398,000 appeals. 42,000 overturned. 10.6%.

That number reads two ways.

One: it really is enforcing the rules it wrote.

Two: once "safety" becomes an industrial process handling sixty thousand cases a day, it stops being a moral judgment and becomes a production line.

Production lines make errors. A 10.6% correction rate means that for every ten people who cry foul, nine never get their account back.

8. The strictest branch, the coldest branch

A piece of old history.

Within Chinese Buddhism there was a school called Nanshan Lüzong — the Vinaya school of the southern mountain. Strictest in precepts, most meticulous in scholarship; the Commentary on the Four-Part Vinaya parsed the rules down to the millimetre. When Daoxuan founded it in the Tang dynasty, the brightest monks gathered.

A thousand years later, walk through the great temples of China and try to find a Vinaya ancestral hall thick with incense.

Why?

Because people want sudden awakening, not gradual stages. Chan Buddhism arrived with one line — no reliance on words, direct pointing at the mind — and swallowed centuries of monastic law whole. The Vinaya school stayed where it was, explaining, item by item: this you may not do, that you must think through.

The AI jianghu is no different.

The world wants stronger models, not slower ones. It wants "can this be done," not "should this be done." It wants launch-day benchmarks, not ninety pages of risk factors.

The Vinaya sect stands on the riverbank holding up a sign. The river is rising faster than the sign.

But it isn't meaningless. Three judgments.

First: it is not a hypocrite

A hypocrite doesn't lose a Pentagon contract over two clauses, get put on his own government's adversary list, and then sue. It happened. There's a ruling. There's a date. You can't argue with it.

Second: it is not a saint

Its vows are drawn where the business survives. Its ruler measures in one direction only. The "humanity" it invokes has an invisible boundary. All three are true, and all three are documented.

Third: it is the rarest kind of sect

When it could afford the cost, it paid. February 2026 was real, measurable, and settled on the spot. No other company in the industry has done that.

The real problem isn't the company. It's the position.

A company can brake itself. It cannot brake a racetrack. It can refuse once. It cannot make a weapon already wired into the system refuse a second time on its behalf.

So instead of asking how long the idealism can hold, ask the colder question:

You don't have to guess. Watch three things.

One: the next time a government demands a clause change, does it sue again — or invent a "technical exemption"?

Two: in its first annual report after listing, which grows faster — the safety and alignment budget, or the sales budget?

Three: who fills the sixth seat on the Long-Term Benefit Trust. If it's a second national-security institutionalist, the answer is written there.

A thousand years on, the Vinaya school is still here. The incense is thin, but it never went out.

What it held was not the world. It was itself. In this jianghu, that is already the hardest thing there is.

In closing

Seven points.

1. It institutionalized an ideal; it is not "an idealist." Idealists run on people, and disperse when the people leave. This one wrote the obsession into its cap table: a PBC charter that permits directors to not maximize shareholder value, a Class T share holding four board seats, seven founders pledging 80% of their equity. You won't find a second example on Wall Street.
2. Its "no" is real, but the location of the "no" was calculated. It refuses fully autonomous lethal weapons; it does not refuse target identification. That line is technically defensible, morally decisive, and commercially survivable. Not hypocrisy — pricing.
3. Keeping vows requires capital, and the capital comes from a world that doesn't. Amazon and Google pay the incense bill, which is what lets it decline dirty work. That structure isn't its fault. It's its fate.
4. Principles are easiest to hold where they are cheapest. Walking away from the Chinese market cost 5% to 7% of revenue. Affordable, so it was done. The genuinely hard vow is still ahead.
5. Its ruler only measures in one direction. The same "local law may compel disclosure" reasoning is never applied to American companies. That isn't a neutral judgment. A hierarchy of trust is baked in.
6. It can brake itself, not the racetrack. That night over Iran is the proof: it said no, got struck off, and the sword flew anyway. Once a technology is embedded, a single supplier's refusal stops being a switch and becomes a fine.
7. On listing day, the binding constraint shifts from "vows" to "quarterly reports." One is recited annually and runs on conviction. The other is recited four times a year and runs on the share price.