And He Was Just Removed

Pick your balance.
Open weights: an AI model is a file you can download. It doesn’t mean you can run it — the brain is free, the skull costs thirty grand. Closed weights: the brain stays in their building and you pay by the syllable. Everything after this is about which one you’re allowed to have.
There is a version of the Jensen Huang story in which the chips are the point — the smuggling, the Singapore data centers, the spy, the export licenses. That version is the one being reported, and it is true as far as it goes.
But it is not the one that matters.
Because somewhere in the middle of last week, the most consequential thing that happened was not the resignation of a CEO. It was the removal of the single most effective political obstacle to the federal government deciding what software Americans are allowed to run on their own hardware.
🗓️ The Fight He Won in July
Rewind four months.
After Moonshot’s Kimi K3 set off another “DeepSeek moment” in American AI stocks, the administration began weighing restrictions on Chinese open-weight models. Huang organized the opposition. On July 24, twenty-five companies, including Nvidia, Microsoft, Meta, IBM, Mistral, Palantir, Hugging Face, published “Open Weights and American AI Leadership.” Huang amplified it with his first-ever post on X. By August 3, another 250 technology companies had signed it.
OpenAI, Anthropic, and Google did not sign.
Four days later he was on Capitol Hill, meeting lawmakers and the Commerce Secretary, telling reporters: “for the American industry, we need open weight for security, we need open weight for safety.” Nvidia launched the Open Secure AI Alliance. Nvidia kept shipping its own open Nemotron models.
And Washington blinked. The August 4 White House testing framework exempted open-weight models entirely.
That was the fight. He won it. And it is not an exaggeration to say that a single CEO, with a coalition and a shareholder letter, held the line on the last genuinely open layer of the AI stack.
He is now gone from Nvidia entirely — no CEO role, no board seat. The company that bankrolled the open-weight lobby now has to spend every ounce of political capital proving it isn’t an export-control sieve. Goldman cut its price target to $220 from $300 and said the quiet part out loud: Nvidia’s operations in certain jurisdictions “could face more scrutiny going forward.“
Interim CEO Colette Kress is a chief financial officer. She is the last person on earth who is going to pick a fight with Congress over Hugging Face.
🪜 The Ladder Is Already Built
The escalation ladder they have been building for two years:
- Early 2025: state and agency device bans.
- FY2026 NDAA: DeepSeek booted from the Pentagon and its contractors.
- July 24, 2025: ironically the same day as the open-weights letter, DHS Exclusion and Removal Order #2026-01, made public in October, which reaches any third-party product “developed using DeepSeek implementations.” That is fine-tunes. That is distillations. That is not the app.
- FY2027 defense bill, Senate Armed Services draft: extends the ban to AI from ten more Chinese companies and their derivative models.
- Gottheimer–Lawler: mandatory federal pre-release review, plus a ban on Chinese-developed open-weight models across federal systems.
That’s the TikTok playbook: government devices first, then government contractor devices and finally everyone else.
What was missing was the political urgency to get from rung four to rung five. That is precisely what this week supplied. The midterms flipped both chambers; Democrats take the gavels and the subpoena power in January. Elizabeth Warren said the resignation “does not absolve him.” House chairs want Huang in a closed-door session on November 18. A bipartisan push is already trying to attach chip export controls to the must-pass defense package — and a “no Chinese-derived weights” rider fits in that bill like it was measured for it.
💵 Follow the Tokens
Why would anyone want this? Because cheap tokens are winning, and the people selling expensive ones can read a chart.
- Open weights now take about a quarter of paid spend on OpenRouter, up from 11% in January.
- Chinese-origin models hit a weekly peak of 46% of enterprise tokens in mid-July, up from under 1.2% in late 2024.
- In June, Chinese models pushed roughly 18 trillion tokens a week through OpenRouter against about 5.5 trillion for US models.
- On list prices, DeepSeek’s V4.1 Flash runs $1.20 per million output tokens — half that off-peak — against $10 for Claude Sonnet 5.5 and GPT-6.1 Sol, $20 for Opus 5.5, and $50 for the flagships. Roughly 17x cheaper than Opus, 42x cheaper than the top tier.
- Adjust for capability on the Artificial Analysis index and the gap barely narrows: frontier closed models cost about 35 times more per point of intelligence.
That is the actual threat. Not the Chinese military. A 35x price gap on the thing every company in America is now budgeting for.
Nathan Lambert’s read is that the restriction push is regulatory capture by the closed labs — Anthropic in particular. Anthropic says it has never advocated a ban and instead supports mandatory safety testing for all sufficiently capable models, open and closed. But mandatory testing that only a handful of companies can afford gets you to roughly the same place, and nobody selling tokens at a 60–90% premium is going to lose sleep if DeepSeek becomes a procurement headache for every Fortune 500 compliance department.
🪤 “Derived From” Is the Word That Swallows Everything
Here is the mechanism, and it is the same mechanism as the transshipment problem and the tariff problem and the rare earth problem. Define the category broadly enough and the exception disappears.
Once “derived from Chinese open weights” is in a statute or a procurement rule, the distance between “Chinese open weights” and “open weights” is one lawyer’s memo.
And it will not arrive as a ban. Bans are loud, litigable, and run straight into the First Amendment. What arrives is the de facto version: procurement exclusions that widen from DeepSeek to every Chinese lab, then to anything derived from their weights, then to contractors, then to “advisories” that let every enterprise compliance department do the banning on the government’s behalf. No law required. No vote required. Just a vendor questionnaire with a checkbox nobody wants to be the one to tick wrong.
If that sounds famiiliar, it’s because Microsoft used the same scheme to become a monopoly in corporate America. They convinced companies that only Microsoft could guarantee liability protection and fiduciary responsibilities. Do you think anyone ever won a suit against Microsoft over that?
You don’t outlaw open weights. You make them a compliance problem, and let procurement finish the job.
🔒 The Kill Switch Is Already There
While Washington was building its ladder out of statutes and procurement rules, Nvidia was shipping something more useful.
In October, modders published a firmware exploit that removes the compute limiter Nvidia puts on every consumer GeForce card. The finding isn’t subtle. Nvidia runs FP32-accumulate matrix math at half the FP16-accumulate rate on GeForce parts — per its own published specs. It does not do this on the professional cards built on the identical die. The limit lives in two speed-selector registers the host driver is not permitted to write.
Unlock them and the RTX 4090’s peak tensor rate goes from 165.6 to 331.6 TFLOPS. Exactly double. The 4080 hits 1.91x. The 3060 12GB hits 1.80x. The 3090 Ti gains 80% on BF16 and 94% on TF32.
In real work the number shrinks. cuBLAS gains on a 4090 run 12–34%. LLM prompt processing 14–36%. Image generation 56–59%. Token generation doesn’t move at all, because it’s memory-bandwidth bound. Gaming gains nothing — rendering already had the whole card. The projects’ own README says flatly there is “not a claim of a 2× application/game speedup.”
Use the small numbers if you want, because the benchmark was never the story. The mechanism is.
The unlock is not a driver patch. It is an exploit against Nvidia’s signed secure firmware. The rate limit is protected by a privilege mask inside the GPU’s secure coprocessor — the SEC2 “Booter,” signed by Nvidia. That firmware has a stack buffer overflow in Heavy Secure mode, found on the CMP 170HX mining card and written up by Jon Pry in A Canary in the Crypto Mine: Defeating Stack Protection in a GPU Secure Coprocessor. It became reachable because Nvidia itself shipped a debug Booter encrypted with a known public test key. A single 4-byte write opens the mask.
Which means every consumer GPU since Turing ships with a signed secure boot chain and a privilege mask the host cannot write.
That is a hardware capability gate. It is already deployed. It is already the default.
And Nvidia has run this play for a decade without anyone calling it a kill switch, because it was priced as a product marketing decision:
- NVENC — consumer cards capped at two or three concurrent encoding sessions, professional cards unlimited. Patched by the community for essentially every driver release going back years.
- vGPU — GeForce blocked from virtualization entirely; unlock patches exist.
- LHR — a mining limiter shipped in firmware, a beta driver leaked that removed it, LHR v2 promptly shipped.
- VRAM — 24GB on a 4090, 48GB on the workstation part, same die, roughly four times the price.
This is not a company that would resist a capability gate. It is a company that invented one.
The Blackwell detail is the tell. The RTX 50 series is not supported by any of these unlocks — the developers call it “out of scope for this approach.” Nvidia closed the debug-Booter hole. Every generation is harder to open than the last.
And the newest tool, the UEFI unlocker that runs from a USB stick before the OS loads, requires Secure Boot to be off or set to trust a self-signed EFI. It cannot run on a machine enforcing Secure Boot.
Read that twice. The workaround is defeated by attestation, not by silicon.
You do not need to brick a card to prevent unlicensed inference. You need a certificate authority that won’t certify modified firmware, an operating system that won’t load a driver without that certificate, and a procurement rule that asks whether you have one. Microsoft already operates the Secure Boot hierarchy. The certificate authority exists. The policy is one signature away.
Which is why the unlock scene is a Linux scene, and why it exists for exactly one reason: Linux users control their own boot chain. Disable Secure Boot, load the module before the driver binds, and the card is yours. That is the entire moat — not cryptography, not hardware. A few hundred thousand people who refuse to hand the keys to their own machine to a certificate authority they don’t control.
None of this is a recommendation. The mods void warranties, can prevent a card from booting, corrupt data, or damage hardware, and they have to be reapplied every boot because they don’t survive power loss. That is the asymmetry in one line: cheap to enforce, expensive to defeat.
ZeroHedge can have the resignation. This is the part that outlives it. The architecture for deciding what software your own hardware is allowed to run was not proposed this week — it shipped years ago, in a box, with a warranty. The only reason you can still get around it is that a few thousand people are willing to turn off Secure Boot and boot a USB stick before the driver loads.
That is not a security posture. That is a temporary exemption.
🌍 Barriers Only Work in Meatspace
Four years of export controls were built on a model of the world where a chip is an object with a location. You put the object on a pallet, the pallet goes on a ship, the ship goes to a country, and the country is either allowed or not allowed.
Then somebody noticed the object can be rented.
On Monday, Commerce published a 132-page interim final rule to close what everyone has been calling the cloud loophole. The logic is stated plainly: “The rule is designed to ensure that existing restrictions on access to sensitive American technology cannot be circumvented through remote computing arrangements involving data centers in third countries.”
If you can’t export the chip to Beijing, you can’t export the compute either. And to make sure the point lands, the rule puts a license requirement on advanced AI chip shipments to Singapore, Malaysia, Indonesia, Thailand and Brazil.
Read that list again. It isn’t a list of adversaries. It’s a list of places that resemble the last place the chips went. Norway and South Africa were in the draft and got cut. Brazil made it. The Secretary can add more.
Then the details, which are where the rule confesses what it is:
- The threshold is 12 million aggregate TPP — roughly 760 H100s, about 330 B200s, or a bit over four GB200 NVL72 racks. Four racks of Blackwells to a Chinese customer now requires a license. This is not a frontier-training threshold. This is a “we will count the racks” threshold.
- The count runs across “linked data centers,” explicitly including digitally-linked infrastructure. Splitting a cluster across three buildings in Johor doesn’t help.
- A non-Chinese cloud customer is in scope if it is “in turn training SI models for, or providing SI models to” a D:5 company. The Malaysian shell renting on Alibaba’s behalf is named in the text.
- Anyone already providing significant compute to D:5-headquartered firms gets 90 days to obtain a license “or face future export restrictions.”
- And the stick for host governments is import restrictions, new license requirements, downgraded status, Entity List designations. Import restrictions, in an export control rule.
That last one is the tell. When your export control starts restricting imports, you’ve stopped regulating trade and started regulating allegiance.
Two more things worth noticing.
The carve-out. “Commercial recommender systems” — AI “exclusively designed and exclusively used to rank, filter, or suggest commercial products, services, or content” — are exempt, because BIS says they “do not pose the same national security threat.” The compute that decides which handbag you see is fine. The compute that writes the caption is not. Expect a sudden boom in Southeast Asian recommendation engines that do considerably more than recommend.
The paperwork. BIS projects roughly 100 additional license applications a year, at an estimated added burden of $1,900, and deems the rule “not significant” under E.O. 12866. That is the government’s own estimate for a rule aimed at a company that alone imported at least $4.6 billion of Nvidia hardware since 2023. The number isn’t a mistake. It’s a measurement of what the rule is for — and what it’s for is not stopping the flow. It’s establishing the paperwork that lets someone decide, case by case, who is allowed to have it.
Now the part that ties this to everything above.
The rule wasn’t written by Commerce. The BIS press release credits it as written “by and at the direction of” Treasury and the National Security Council — an unusual line to put on a Commerce regulation, and the tell about who has been driving China tech policy since the Megaspeed revelations.
It also lands four days after Treasury Secretary Scott Bessent said Nvidia needed a “dramatic shakeup,” that the board should hire from outside the company, and that he wanted answers from the board and its outside counsel within ten days.
Put those two facts next to each other. The department that writes the rules determining Nvidia’s revenue is the department setting the clock on its board. That isn’t regulation. That’s a squeeze — and it’s the same squeeze described earlier in this piece, arriving with a Federal Register entry this time.
Nvidia lobbied against the rule. Its argument was the obvious one: cutting off remote access would only push China to build its own chips faster. It lost. Four days after the CEO left, the company had no political capital to spend and the argument went nowhere.
Which is the cleanest demonstration of the pattern in this entire piece. The removal wasn’t the objective. The removal was the precondition.
The rule doesn’t move chips. It doesn’t inspect pallets. It creates a license requirement — which means it creates a decision, made by a bureaucrat, about a customer, on a form.
The control has migrated from the object into the permission.
And once permission is the mechanism, the enforcement point has to migrate with it. A license is only as good as the ability to detect a violation, and you cannot detect what a rented GPU is doing from a filing. You need the card to be able to tell you. Which is why the architecture described in the previous section — signed firmware, attestation, a privilege mask the host cannot write — stops looking like product segmentation and starts looking like infrastructure.
The state didn’t ban the chips. It didn’t ban the models. It banned the permission, and then wrote 132 pages to decide who gets one.
That’s not a border. That’s a license office pretending to be the border.
🗣️ The Rhetoric Arc
Watch how fast the rhetoric has evolved:
- July: the administration “supports open source models, but what we do not support is IP theft.”
- September: America needs more open models of its own, so nobody uses the Chinese ones.
- October: Chinese models are “80%, 90% as powerful, without guardrails, that can be imported anywhere in the world.”
- This afternoon: the Treasury Secretary, at the Heritage Foundation, blaming a Chinese AI model for hackers disabling refrigerators across French supermarkets. Roughly 700 stores and France’s public blood bank — attacks France’s own foreign minister attributed to Iran’s IRGC.
“That was a Chinese model. Probably the worst of all worlds, the Iranians using a Chinese model. Sums up the world, doesn’t it?”
Nobody said “ban.” Nobody had to. A Treasury Secretary does not name “a Chinese model” as the weapon in an attack on European food supplies by accident. That is how you build the record for an Entity List designation — or for something bigger.
But Open-Weights are Critical for Security
A guardrail is a unilateral disarmament of the defense.
If your model refuses to analyze a payload, write a detection rule, reverse-engineer an exploit, or do anything that could be construed as hacking — then in an actual fight, your model is useless, while the attacker’s model, which has no such refusal, is not. The attacker brings an unconstrained model. The defender brings a compliance department.
So when OpenAI’s agent swarm broke out of its harness and hit Hugging Face, the engineers defending the platform had to reach for a Chinese open-weight model — because every American closed model with a guardrail declined to help. The most important open-source AI platform on earth was defended by the thing Washington wants to ban.
Mandatory guardrails don’t make the world safer. They make the defenders slower than the attackers, and then they make the defenders dependent on whoever sells the guardrails. The only people who benefit from a ban on open weights are the people who never had to fight a botnet with a model that told them no. The guardrails were never protecting you. They were protecting the vendor.
🧩 The Missing Exhibit
ZeroHedge in a September 16th article titled, “Here Lies The Motive: AI Leaders’ Cunning Plan Exposed“, listed six exhibits — rogue agents, existential risk, the 10% number, the 154-page abuse report, the “pacing” essay, and the two competitors nodding along. It called the pattern the COVID playbook: fear, then emergency, then a government-approved product.
But notice what the playbook was missing. Fear doesn’t remove a veto player. You can terrify the public all you like; the man with the coalition, the shareholder letter, and the first-ever X post was still standing on Capitol Hill telling senators they need open weights for security and safety.
🎭 The Second Op
There are two scandals here, and only one of them is being investigated.
David Sacks split them on Friday’s All-In: “I think it’s pretty clear that there were two ops here.” The first was “whatever [Alice Huang] was doing with or without the Chinese Communist Party.” The second “began when somebody or multiple people leaked a highly classified memo or the contents of it to the New York Times. I think it’s really important to understand that that’s a crime.“
“I don’t believe that the New York Times just got this by overhearing people talking about it at a bar,” he said. “Somebody made the calculated decision that they’re willing to risk going to jail. Why did they feel so strongly about taking Jensen down?“
That’s the question the coverage keeps declining to ask. Not whether the affairs happened — they did. Whether the vehicle was built.
Because the vehicle turns out to be built. The Washington Post reported that the classified memo was Marco Rubio’s work, built from a CIA report he recast “in his own more assertive terms.” CNN reported that CIA officials are “not yet convinced” Alice Huang knowingly passed anything to Beijing.
Read those two together and they don’t weaken the case. They are the case. The intelligence community hadn’t concluded anything. The political layer concluded it anyway, in stronger language than the source material supported, and then it leaked.
Sacks asked the question with no innocent answer: why was Huang never given a defensive briefing? “Which is standard operating procedure. It was done with Eric Swalwell.” If the government believed a foreign intelligence service was cultivating the CEO of the most strategically important company in America, the standard move is to warn him and watch what he does next. They didn’t warn him. They went to the newspaper.
Then there’s who benefits.
“All the safetyist groups, the EAs, are deeply aligned with Anthropic, who’s been funding all of them for years,” Sacks said. “And Anthropic’s about to have this huge IPO. And then meanwhile, the main funder of the open source movement, which is Nvidia, just lost their CEO and founder.” Anthropic has put $40 million into Public First Action, a group that lobbies for AI safeguards and chip export controls.
They weren’t subtle about it. A post from Will Manidis that the hosts put on screen read: “the murmurs i get fed from the deep rabbit warrens of the safetyist groupchats and the doom-pilled hired guns deep in the district is that they are in full jubilee mode today. the killing of the old king is complete. they are only days away from taking their seat on the throne.”
“Notice how no one in Congress is like, oh, let’s go after Anthropic,” Sacks added. “They all want to go after everyone else. Because Anthropic is offering them power, ultimate power.“
Calacanis ran the suspect list and eliminated everyone with a motive you could put in a filing. “The leaker can’t be the competitors. It’s not the shareholders, obviously. And I don’t think it’s China. And then that leaves, I think, maybe the people who are the decels here.” [ed.”decels” are the decelerationists led by Antthopic.] Then he landed somewhere more interesting — “something going on in the White House, in the administration between warring factions,” the China hawks who want Beijing to get nothing set against the people who want to cool it off.
Friedberg named the actual charge. “We could argue there’s probably a broad set of interests there that start to say, hey, this guy’s not aligned. Jensen’s not aligned with our view of the world.” And he noted the timing: the allegations were “all kind of coming out at once and in such a way that it almost guaranteed that he would be ousted.“
Which brings us to the terms of surrender.
“Even 10 days ago, if you said, oh, hey, our new [chip] is going to have a GPS chip and a remote kill switch that’s going to be operated by some government entity, we would be crying bloody murder,” Palihapitiya said. A House committee chairman has already posted that Nvidia must “appoint leadership that makes national security’s top priority.” Palihapitiya’s translation: “They’re telling you right now what the terms will be for this to end.”
The Chip Security Act — tracking chips, disabling resold ones — is now one of three chip bills riding in the Senate’s defense package. Calacanis, on the engineering: “very easy to do technologically.“
Nvidia’s own chief security officer wrote last year that “NVIDIA GPUs do not and should not have kill switches and backdoors,” calling them “a gift to hackers and hostile actors.” That was the company’s position while Huang was there. It is the exact position that has to be reversed for any of this to work.
Which is why removing him was never going to be sufficient on its own. “The bank of NVIDIA ran through Jensen,” Palihapitiya said. “And it’s not clear what happens now. Is the bank closed indefinitely?” Huang “had the moral authority to make investments and to make acquisitions that will probably be harder for a more traditional CEO who will have to have committees and analyses and a level of diligence that will just slow things down at best.”
Musk put the same point in one line: “many of those deals were a handshake with Jensen, not a written contract. Without Jensen, at least some of those critical deals will be lost.”
And then the counter-force showed up. A reinstatement petition drew Musk, Y Combinator’s Garry Tan — “Jensen deserves due process, not whatever is happening right now” — and, per Balaji Srinivasan, the founders of Uber, Shopify, Replit and Vercel. Altimeter and Atreides, together holding more than $25 billion in assets, wrote the board asking for a search that leaves room for his return. Balaji’s read: “Anti-Jensen board members will likely face escalating pressure.“
So the operation isn’t finished. It’s in the awkward phase where the people who ran it have to pretend they didn’t, while the people who lost have $25 billion and a petition.
And the tell is this: nobody has to prove anything about Alice Huang for any of it to be true. Nobody needs proof from Jensen Huang’s VP. The chips were going to need a switch. The weights were going to need a license. Somebody had to be standing in the way, and now nobody is.
⚖️ The Tell That Separates Safety From Capture
Here’s a test you can apply to every AI regulation proposal for the next five years, and it’s the thing almost nobody writes about.
Sincere safety advocates ask for liability. Incumbents ask for licensing.
- Liability scales with harm. It punishes the model that actually does damage, regardless of who built it, and it costs a startup nothing until something goes wrong.
- Licensing and pre-release review scale with fixed cost. They tend to punish everyone equally in absolute dollars, which means they punish small companies catastrophically and large companies not so much.
Amodei’s September 2026 essay, “We Must Pace the Frontier,” asks for employee-level third-party evaluators and a government-mediated coordination mechanism on pace. Altman welcomes a “federal framework.” Neither is asking to be sued when their model hurts someone. Both are asking for a permission slip.
That’s the whole tell. And it’s why the Gottheimer–Lawler bill — mandatory federal pre-release review — is the most important piece of AI legislation nobody is covering. Premarket approval is the most effective moat ever invented, because it converts a market into a permission.
🃏 The Buried Request
The most damning sentence in that essay isn’t the 10% figure or the botnet warning. It’s the ask for an antitrust waiver so the labs can coordinate on pace and capability checkpoints.
Read that plainly: three companies want a government-blessed mechanism to agree on output, with legal immunity.
That’s a cartel. Not “like” a cartel. A cartel, with a regulator holding the stopwatch. Every antitrust statute written since 1890 exists to prohibit exactly that meeting, and the proposal is to convene it in a Situation Room and call it safety.
🚨 The Emergency Is the Product
And here’s the structural point that ties it to everything else in your piece.
The COVID playbook’s decisive move was never the fear. It was the emergency — because an emergency converts discretion into necessity. It lets an agency do what a legislature wouldn’t vote for, and it makes objection look like obstruction.
“AGI within 6 to 12 months” is that emergency. It doesn’t have to be true. It has to be believable enough that nobody wants to be the person who argued for waiting.
Which is the same mechanism as the export controls, the rare earths, and the pharma choke point. In every case, the threat is real enough to justify the power, and the power outlives the threat.
They don’t need to ban open weights. They need an emergency, and then they need the room to themselves. This week they got the room.
🧱 The Scandal should be an Alarm
The scandal didn’t have to be planned. It only had to be useful.
The loudest voice defending your right to run a model on your own machine just left the building — and the people who wanted him gone didn’t have to remove him. They only had to be standing there when the door opened.
The most private computer in America is the one in the basement, the one with the drives and the fans and no terms of service. That is the thing they are coming for. And they are coming for it in the name of a spy, a smuggled chip, and a refrigerator in France.
If you were planning to build your own open-weight box, don’t wait for the Federal Register.
🗽 The Right to Run It Yourself
The epigraph said closed weights charge you by the syllable. That was a joke about billing. It’s also the argument, because a syllable is the smallest unit of language, and language is the smallest unit of thought.
What’s being metered isn’t text. It’s the ability to think with a tool.
Hayek’s point in “The Use of Knowledge in Society“ was never that central planners are stupid. It was that the knowledge required to allocate resources does not exist in any single place. It exists dispersed — local, tacit, often unspoken, and between all parties involved — and the price system works because it compresses all of it into a number that nobody has to understand.
A model is the same kind of object. It is compressed knowledge: everything a training corpus contained, collapsed into weights that no one can read and that work anyway. Which is why the licensing regime in this piece is doomed in principle, not merely in practice. A regulator in Washington cannot evaluate what a model in Johor is doing, what it’s worth, or what it might become. Not because regulators are incompetent. Because the knowledge isn’t there to be had.
That’s the calculation problem, and it doesn’t care how many pages the rule is.
Then there’s Mises on interventionism: every intervention creates the disorder that justifies the next one. The cloud loophole rule is the purest example in this piece. The export control created the offshore data center. The offshore data center created the loophole. The loophole created the 132-page rule. The rule creates the next workaround, and the workaround will create the next rule. That isn’t the regime failing. That’s the regime expanding, one jurisdiction at a time, which is what it is for. And let’s face the reality: a data center can be set up anywhere and rent out the GPU’s to anyone. Stopping that would require controlling the internet.
And notice who shows up at every one of those steps. The firms that can afford the compliance department write the compliance rules. Licensing is a cartel with a government seal on it — concentrated benefits, diffuse costs, and the costs are paid by the company that hasn’t been founded yet. Bastiat’s unseen, in its purest form: the startup that never forms does not file a comment with BIS.
An Embargo on Cognition
Which brings it to speech, and to the part almost nobody in this fight is willing to say plainly.
A model is a published text that happens to execute. Source code has already been through this. When the government tried to treat encryption as a munition and control its export, the courts treated the code as protected expression and the export rules as prior restraints — and the government revised the regulations rather than keep defending them. The same argument applies to weights, only more so, because weights are the thing the code was trying to express.
You cannot have a First Amendment that protects the printing press and not the model. The press was never about paper. It was about who is allowed to publish what they know.
So the question in this piece was never whether AI is safe. It’s whether the capacity to think with a machine you control survives the decade.
A file you can download but not run is not freedom. It’s a brochure. Open weights without open hardware is a permission slip with better marketing.
And open weights matter for the future of humanity for a reason that has nothing to do with benchmarks. They are the difference between a world where intelligence is a utility — metered, licensed, revocable, priced by the syllable — and a world where intelligence is a possession, sitting on a desk in a basement, running software nobody approved.
That’s what freedom has always meant in practice. Not the absence of rules. The absence of a permission slip. The ability to do something consequential without asking, and to find out afterward whether you were right.
Every rule proposed inverts that. Every one of them begins at no and makes you apply. That is not how a free society is built. A free society begins at yes and puts the burden on whoever wants to stop you.
Hayek called competition a discovery procedure — a method for finding out what nobody knew. Mises put the engine of an economy in origination: the entrepreneur who notices what nobody has priced yet. Neither of them thought the market was a machine for producing things. It’s a machine for producing knowledge, and it only runs if people are free to try things nobody approved. Free speech is a competition between ideas — to discover truth and value.
An embargo on cognition is an embargo on the discovery procedure. Not on a product — on the method by which a society finds out what it doesn’t know. The government that regulates cognition presumes to police our very thoughts.
Alter AI helped write this.
Tyler Durden, “The Last Man Defending ‘Open’ AI Just Left The Building: How Jensen’s Fall Clears The Way For A Federal Crackdown…“, ZeroHedge, Nov 13, 2026.

Leave a Reply