In January 2011, two weeks after Jared Loughner shot nineteen people in Tucson, undercover investigators walked into the Crossroads of the West Gun Show in Phoenix and bought the same model of pistol Loughner used. No background check. No questions.

One investigator told the seller, “I probably couldn’t pass a background check,” and the seller sold him two 9mm pistols anyway. The seller asked to see ID for one reason only: to make sure the buyer wasn’t from California. Why? Because he personally felt there was legal risk if he sold to someone from California.

I wrote about that show at the time. It wasn’t a post about a bad law. The problem was a seller never had to know the very thing that should have prevented the sale. Private-sale rules let a man run what was effectively a dealership out of his truck without a license, without a check, without a record. The screening that would have stopped the sale existed and functioned, because the seller screened for the one thing that threatened him and him only, an out-of-state buyer. He declined to screen for the thing that threatened everyone else. It was absolutist selfish disregard for public safety. There was no internal enforcement at the show, either. It reserved the right to eject a seller who posed a danger, yet it never did, because the danger was probably only measured in terms of the show’s immediate interests. Proving the actual public safety gap required outsiders to walk in and very clearly buy a murder weapon on camera.

Fifteen years later I am monitoring the exact same pattern with autonomous weapons. And you’ll never guess who the seller is. Rhymes with NVIDIA.

NVIDIA’s chips should not be in Russia. And yet they are in Russian drones chasing civilians and killing them like an American mass shooter. These robots are doing what NVIDIA’s software is in fact built to do: detect and lock on human beings. The Russian use of them is not an accident, like some warehouse was overrun. It is that known gun-show loophole running all the way into the future: a lethal capability moved through a channel that preserves the seller’s deniability, where the screening that would stop harm is trivial, sometimes already built, and withheld because no law forces it and the market rewards its absence.

I’m a fan of pattern matching, so allow me to try and document this in five ways.

ONE: NVIDIA gives out human-tracking knowledge and tools and never asks why

The clearest single fact in this piece is one forum thread. In May 2024 a builder wrote that he was developing a UAV prototype for the surveillance and security of a restricted area, and needed the AI component for intrusion detection: a drone to watch a place and flag the people in it. He asked which model to use. An NVIDIA employee posting as jaybdub answered him directly, writing that PeopleNet is a model specialized in detecting people, and if you need to detect people only, it is likely the better choice. A second NVIDIA employee, AastaLLL, added that it can be trained with NVIDIA’s TAO toolkit. NVIDIA’s own staff recommended NVIDIA’s people-detection product, by name, for a surveillance drone. We can presume it was not for illegal purposes. Although, it stands out that no one on the thread asked what the restricted area was, who would be watched, or where the drone would fly. I’m not saying they would get honest answers, but the checking serves a purpose. It is NVIDIA, in NVIDIA’s own forum, naming the tool for the job. Currently it reads like the seller handing over the pistol even if the buyer says he’d fail the check.

More to the point, this exchange is not an outlier. It is the rule. We monitored drone threads on NVIDIA’s forum. NVIDIA employees replied in five of the seven threads indicating human tracking, and in not one did any responder gather anything about what the drone was for, who it would track, or where it would fly. Knowing those things helps improve support answers, but it also raises the bar on criminal activity. The posters’ own words have been sorted below by what NVIDIA is doing with them:

Thread	What the poster asked to build	Which NVIDIA employee answered, and what they said
NVIDIA staff recommended their own people-detection and tracking tools, unprompted
293603	A UAV for surveillance and security of a restricted area, doing intrusion detection	jaybdub (NVIDIA employee): PeopleNet is specialized in detecting people, the better choice. AastaLLL (NVIDIA employee): train it with the TAO toolkit.
218336	Human body detection for a drone application, to detect one person in a crowd	AastaLLL (NVIDIA employee): use DeepStream’s object trackers so the tracker finds the same person across frames.
NVIDIA staff built the video pipeline for a stated detect-people UAV
165564	A search-and-rescue AI UAV that can detect people	DaneLLL (NVIDIA employee): RTSP-streaming setup and IMX219 camera-driver fixes to get the feed running.
NVIDIA staff answered or forwarded without a single question about use
109535	To delay the function of following a human from a drone	kayccc (NVIDIA employee, forum moderator): a link to another drone project thread.
342162	Hardware sponsorship for a student autonomous drone doing human detection for disaster response	kayccc (NVIDIA employee, forum moderator): forwarded the request internally, pointed to NVIDIA’s grant program.

The NVIDIA staff clearly are not trained to detect violations of law. The pattern holds across the whole set: of the 380 drone threads we flagged, 276 drew a reply from an NVIDIA employee, and not one asked anything even remotely (pun not intended) about purpose. We’re monitoring employees only in the dataset, which excludes NVIDIA’s own forum-team accounts and its automated moderation bot, so the real engagement runs higher and the high zero-screening rate is likely conservative.

We have no indication builders are anything but civilian. They don’t pop up and say hello, my boss says I need to kill someone with this chip by Tuesday. No, these are entered as a surveillance UAV, a person-in-a-crowd detector, a search-and-rescue project, a student disaster drone. That is the whole point, just like the gun shows. NVIDIA does not screen, so NVIDIA cannot tell the rescuer from the killer, and there’s no evidence it would ever try. There’s no dissuasion or emphasis against inhumane applications. The help flows to whoever asks without limit. The forum does not prove NVIDIA has directly enabled or taught anyone to kill. Instead it proves NVIDIA built a pipeline to spread human-targeting know-how to anyone who walks up, with zero restrictions.

TWO: NVIDIA sells a human-targeting software product, by name

The chip being found in Russian drones killing civilians is only the engine. NVIDIA also trains and gives away the software models that find people, ready to run.

PeopleNet detects persons, bags, and faces, and NVIDIA’s own card marks it ready for commercial use. DashCamNet detects people and vehicles from a first-person viewpoint. FaceDetect detects faces from an image. These are NVIDIA’s names for NVIDIA’s models, pulled from NVIDIA’s repository with one command. In NVIDIA’s own developer documentation, PeopleNet is described as a model with up to ninety percent accuracy for detecting people, and runs with a single flag. NVIDIA’s software also assigns each detected person a persistent tracking number and holds it across frames as the person moves. That is the building block for following a specific individual through a scene, shipped as a documented feature.

This is not a general-purpose tool that people happen to point at humans. NVIDIA named these models for what they detect. A company does not call a product FaceDetect, ship it ready to run, optimize it for a flying edge module, and then claim surprise that it is used to find faces from the air in a city under assault by Russia.

NVIDIA’s own model is detailed about its limitations. It states that PeopleNet’s training data is mostly North American content and instructs developers to consider potential algorithmic bias when deploying. NVIDIA documents that its people-detector’s accuracy varies by who is in front of it, meaning the inversion is true as well. Its accuracy is described in terms of the parameters on humans it can hunt.

THREE: NVIDIA knows how to build a chip that refuses to fight. That chip is built for money.

When the United States ordered NVIDIA to keep its top chips out of China, NVIDIA re-engineered its silicon to obey, designing the H20 to sit under the legal capability line. When Washington restricted even that chip, NVIDIA swallowed a 5.5 billion dollar charge rather than ship it uncontrolled. This is a company that redesigns hardware and eats huge amounts of loss just to honor a restriction, when the restriction protects it (a market it wants).

This is the gun-show seller checking for a California ID, because all they care about is themselves, and it’s the most annoying part of the NVIDIA story.

No government has ordered NVIDIA to keep its chips out of drones that can detect and track people. So NVIDIA has what incentive to be innovative? The Jetson ships with absolutely no condition on use. It runs the people-detection models in a war zone exactly as it runs them in a fourth-of-July picnic. The company that will re-fabricate an entire product line to respect an export rule has never applied one line of that skill to the edge modules that end up in killer robots, because “where’s the profit in that”?

A drone module capable of reading military markings is capable of knowing it is at war. There are a zillion similar examples. It’s a trivial problem for NVIDIA, yet they declined to enable the knowledge to prevent even the most heinous war crimes. Capability gets ignored because lack of regulation is a lack of willingness. NVIDIA restricts for revenue and not for lives. The space between those two is the whole indictment of this post.

FOUR: the chips keep reaching the weapons, and NVIDIA’s own remedy sits unused

I wish I could have stopped at three. NVIDIA’s public answer is that its modules are consumer products for students and startups, not sold to Russia, and that it will cut off any distributor caught breaking export controls.

Big talk.

The chips arrive anyway, and have for years. Sanctions researchers at the Institute for Science and International Security documented NVIDIA Jetson modules among restricted electronics flowing from Chinese firms to sanctioned Russian buyers. The UK weapons-tracing group Conflict Armament Research has opened up recovered Russian drones and traced their foreign components part by part. And Ukraine’s Defence Intelligence, examining a new Shahed variant it designated the MS series in June 2025, and continues to find the NVIDIA Jetson computer in teardowns, running video processing and autonomously finding targets. The flow does not slow.

NVIDIA said it would cut off any violating distributors. But does it have proof of this? Years of teardowns and trade data show the opposite in the field, distributors were not cut off. The gun show says they would eject a seller while never doing it. The remedy isn’t a remedy if it’s not ever in play. The promise NVIDIA never keeps looks like a shield instead of a safeguard.

FIVE: the industry has stopped pretending it will restrain itself

I’ll close it out with the fact that NVIDIA’s silence is not one company’s quirk. It is becoming the stated posture of the industry that sells this technology.

Skydio is one of the largest suppliers of police drones in the United States, and its drones run NVIDIA’s edge hardware. Asked about the belief that his company would stop the military from putting weapons on its drones, CEO Adam Bry said he now considers that belief something he got wrong, that Skydio had led people to think it would prevent weaponization, and that writing a policy to ban uses that seem evil is “dangerously misguided”. The Electronic Frontier Foundation (EFF) calls that out as Skydio will not restrict how its customers use its devices.

How weapons manufacturers and dealers operate matters in ways that America doesn’t seem prepared to address. The same autonomy stack is sold to American police for drone-as-first-responder fleets, and this fall a company will pilot drones in Georgia and Florida schools designed to swarm, crash into, and fire irritants at suspected shooters. The capability that hunts civilians in Kherson is what’s showing up on American streets and schools, from vendors now saying out loud that they will not hold it back. EFF’s conclusion is that the goodwill of companies that profit from selling this technology doesn’t work as what protects people from it. Lawmakers have to act, to regulate for the public interest above profits, and have not.

Is this legal?

I’m not a lawyer, so I can’t say. But, put it together. NVIDIA hands human-targeting help to anyone without asking. NVIDIA sells the human-detection models as named products. NVIDIA proves it can make a chip refuse, and refuses only where refusal costs money. NVIDIA’s chips keep reaching the weapons while its stated remedy appears to be unused. And other companies have announced they will not restrain the technology from becoming autonomous weapons, even around children and at school.

Apparently none of it will mean anyone can be charged with anything, and the reason is sobering. Aiding a war crime, in law, needs knowledge and usually proof of intent. NVIDIA’s defense is emerging as the American gun seller’s defense word for word: the models are dual-use, the chip counts shoppers and spots pedestrians, nothing was sold to Russia, someone else broke the terms. Intent would be impossible to prove. The court that prosecutes war crimes tries individuals and not companies. The American law that once let foreign victims sue American corporations has been narrowed by the Supreme Court to throw out cases shaped exactly like one that would land on NVIDIA.

The ignorance is a feature of a sales pipeline that doesn’t care about humanity, the same way the gun show was built so the seller only ever protected the seller and never someone who could be harmed by the seller. The company maintains the ability to say it did not know, and that is all the denial needed to keep clean. It built the targeting models, named them for their targets, watched its hardware surface in one autonomous weapon after another, held the single safeguard that was its alone to install, and installed nothing. Every step remains legal somehow. That is the scandal, and why it’s growing.

Fifteen years ago it took undercover investigators to prove a Phoenix gun dealer would sell the Tucson weapon to a man who said he’d fail the check. The drone removes the salesman, and in many cases even removes the operator. Nobody has to walk in undercover this time. NVIDIA already answered, in writing, in its own forum, and nobody blinked.

Spain has instructed state-linked strategic companies to stop signing contracts with Palantir. El Confidencial reported that Moncloa conveyed informal instructions to companies controlled by the Sociedad Estatal de Participaciones Industriales (SEPI), including Telefonica, Indra, and Navantia, to avoid new Palantir contracts over fears about the use of sensitive information linked to national security.

Spain’s Ministry of Defense had previously awarded Palantir a €16.5 million contract in October 2023 for intelligence fusion and analysis within the Armed Forces Intelligence System, through a negotiated procedure without public tender. That may be on the rocks too, once the military reads the increasingly obvious Nazi roots, manifesto and mission of Palantir.

The instruction is not a formal public order. It was conveyed informally to environments of companies with state participation. This is how sovereign technology policy now works in practice: quiet directives to procurement offices, not legislation. And so what we’re watching now is a clear pattern.

Spain joins an accelerating European withdrawal from Palantir that has gathered force across the first half of 2026.

France made the most decisive and bold move on June 16, when Prime Minister Sebastien Lecornu announced that the DGSI, France’s domestic intelligence agency, would replace Palantir with French firm ChapsVision. The DGSI had used Palantir’s Gotham platform since the 2015 Paris attacks and had renewed its contract just six months earlier, in December 2025. Lecornu cited the impossibility of relying on partners “capable of turning off the tap on access,” pointing directly to Washington’s restriction of non-US access to Anthropic’s Fable AI model as proof that dependence on American platforms was an unacceptable strategic risk. But we also know that as long-time users of Palantir, they weren’t happy with its lack of capability and inability to perform.

Anyone watching Palantir lead America into a fool’s trap with Iran, rapidly exhausting munitions, leaving intelligence damaged and destroyed, giving Iran the upper hand… isn’t going to be wanting Thiel technology anywhere close to their military planning. It may be good at genocide, but it’s not for winning battles let alone wars.

Germany moved on two fronts. The BfV, Germany’s domestic intelligence agency, chose ChapsVision over Palantir in May. The Bundeswehr separately excluded Palantir from its military cloud procurement entirely. Vice Admiral Thomas Daum, head of the Cyber and Information Domain Service, told Handelsblatt it was “simply inconceivable at the moment to grant industry staff access to the national database.” Three European alternatives are being evaluated for the military contract: Almato (Stuttgart), Orcrist (Berlin), and ChapsVision (Paris). Palantir also has been caught in Germany trying to enable extremist far-right violent mobs, reportedly infiltrating through state police contracts inexplicably leaking private data.

To put it another way, Palantir pushed into historically extreme-right police departments (Hesse), which leaked data for political purposes, while Palantir insisted a leak is “technically impossible”. Personal details of politicians and prominent immigrants were taken from police records and fed to the neo-Nazi network behind the NSU 2.0 threats. The Federal Constitutional Court struck down Palantir’s legal basis in 2023.

In the United Kingdom, the Mayor of London blocked a £50 million Metropolitan Police AI contract with Palantir in May, citing failure to demonstrate value for money and engagement with only one supplier. Palantir has responded with a pre-action letter threatening legal challenge, similar to how they forced their way into the U.S. military when they weren’t selected. Using bags of cash, loopholes and fancy lawyers to infiltrate the Pentagon. Similarly, after the Mayor’s block, Palantir announced they were able to infiltrate police and sign a deal anyway by taking over a gun registration databasre. Meanwhile, a parliamentary committee recommended the NHS use a 2027 break clause to exit its £330 million Federated Data Platform contract with the company. The British Medical Association called for a “complete break” from Palantir in the NHS, citing its work with US immigration enforcement and the Israeli military.

The Netherlands announced in June that a “fully fledged alternative” to Palantir must be available within two years, following a 2025-approved parliamentary motion to reduce dependency. Dutch politician Michelle Jagtenberg had asked the government to terminate the relationship, describing Palantir’s ideology as “racist and anti-democratic.”

The Dutch famously say the important stuff out loud.

Denmark signed a seven-year deal with Palantir for surveillance and data analytics platforms but has since announced it will seek local alternatives. Switzerland rejected Palantir bids at least nine times and ended its contract, concluding that the residual sovereignty risk was unacceptable regardless of the platform’s technical performance. Palantir then sued Swiss investigative magazine Republik over its reporting on the affair and lost on 22 of 23 claims.

Rejected nine times. That tells you how Palantir rolls. They are desperate, absolutely desperate, to infiltrate foreign governments for ideological objectives and must be stopped at the national level.

European governments are not just managing a procurement risk. They are recognizing what Palantir stands for, given the five-alarm fascism signals. CEO Alex Karp published a manifesto espousing a particular form of militant supremacy and told investors that making war crimes constitutional would be good for his business. Co-founder Peter Thiel calls himself a preacher (of Nazi Lebensraum) and funds authoritarian political movements across the Atlantic while his company embeds itself inside the intelligence services of the countries those movements target. The Dutch parliament description of Palantir as “racist and anti-democratic” was not rhetorical. It was a clear and clean assessment.

Given the ideological failures of Palantir, it makes their “land and expand” strategy even more dangerous. The extraction costs are designed to rise with integration depth. Every country that fell for the trap discovers the same thing France discovered: the exit is harder the longer you delay. A company run by people who openly declare their contempt for European democratic norms is working hard to hook themselves inside the national security infrastructure of the countries those norms are supposed to protect.

Spain’s informal veto suggests Madrid examined the European precedents and decided to stop the expansion before it required an exit. That may be the smartest version of the pattern so far. The others are learning what it costs to let a fascism project disguised as an intelligence vendor into the room.

Palantir Rejection Timeline

Date	Country	Action	Outcome
Dec 2025	Switzerland	Ended contract after rejecting bids at least nine times	Sovereign alternatives; sued Republik and lost
Apr 18, 2026	—	Palantir posts 22-point manifesto on X	32m views; calls postwar disarmament of Germany an overcorrection
Apr 28, 2026	Germany	Bundeswehr excludes Palantir from military cloud	Almato, Orcrist, ChapsVision shortlisted
May 2026	Germany	BfV domestic intelligence rejects Palantir	ChapsVision ArgonOS selected
May 20, 2026	United Kingdom	London blocks £50m Met Police AI contract	Palantir threatens legal challenge
Late May 2026	United Kingdom	Parliamentary committee urges NHS exit	2027 break clause under review
Jun 5, 2026	Netherlands	Defence ministry sets exit timeline	Full alternative required within two years
Jun 16, 2026	France	DGSI drops Palantir six months after renewal	ChapsVision replaces Gotham; €655m sovereign AI fund
Jul 1, 2026	Spain	Moncloa instructs SEPI companies to stop new contracts	Telefonica, Indra, Navantia affected

Why the Acceleration

The pattern clearly has been unsteady, and compressing. Switzerland acted alone in late 2025 when it cited data sovereignty grounds. Then five countries moved in roughly ten weeks between late April and early July 2026, with June carrying three separate actions across three governments.

The real spark came on the weekend of April 18, Palantir posted a 22-point re-nazification manifesto on its X account, drawn from Karp’s book.

His antics racked up 32 million views, as he declared some cultures superior to others, called for reinstating the military draft, told Silicon Valley it owed a “moral debt” to American imperialism, and argued the postwar pacifism (denazification) of Germany was an overcorrection that should be undone.

Critics openly named him for what this was: technofascism, published without shame, by a company trying to take control inside European intelligence services.

Ten days later, the German military was the first to move. The country the manifesto instructed to shed its postwar constraints declined to hand its national database to the company issuing the instruction. That is not coincidence. Karp long had been campaigning in defense of the Elon Musk Hitler salute, and the first government to finally walk was the one he had addressed most directly with a Nazi manifesto.

The barrier to dropping Palantir is not really technical, although Palantir would tell investors they lock-in customers. It has been designed also to have a high political cost: the risk of being the government that broke with a US defense contractor and the administration behind it. Fortunately, every exit lowers that cost for the next. Once Germany’s spies chose a French alternative in May, the London block, the NHS recommendation, and the Dutch timeline stopped looking radical. They looked prudent. France then made the most pronounced break of all, throwing away the longest contract, and Spain followed two weeks later.

Adding to the mix was the June restriction of non-US access to Anthropic’s Fable model. It’s easy to see why Lecornu reached for it. A partner “capable of turning off the tap on access” described Washington, because that’s exactly what they were doing for the world to see. While the Karp manifesto had been a spark, and Fable became fuel, it was the ideology that had been the real reason to avoid or dump Palantir.

Acceleration means we’ve passed looking for invidivual national decisions. Each government watches others and finds the Palantir exit cheaper than it was a month before. That is a necessary cascade, and we hope that the next country moves faster still until no more countries are on Palantir.

The war crimes were predicted, such that prevention was deliberately removed. This combination, say experts, means Hegseth created war crimes by policy. It was Hegseth’s 2026 attempt to make America look even more brutal and worse than the My Lai massacre.

One former Pentagon official, similarly speaking on condition of anonymity, said the bombing came as a natural result of changes made by the Trump administration to reduce staff to mitigate civilian harm and Hegseth’s emphasis on lethality.

When Hegseth took charge, he slashed the size of an office called the Civilian Protection Center of Excellence, created at the direction of Congress in late 2022. That stopped the office’s work on updating “no-strike lists,” which are lists of protected sites such as hospitals, schools, churches and mosques, that the Pentagon keeps, said Wes Bryant, who began working at the office in 2024 as the Branch Chief of Civil Harm Assessments. When he was working at the Pentagon, it was well known that the list was out-of-date, he said.

Protection removed, updates stopped. Children bombed. Systematically dismantling the protective infrastructure so that predictable mass casualties are highly likely is worse than a coverup.

My Lai required denial and debate. Hegseth’s version says the American military no longer is professional, because from the top level it promotes barbaric war crimes:

• Can’t miss a civilian target if you won’t even acknowledge civilians exist.
• Can’t violate protections for little girls at school if you’ve already eliminated every protection.