The Government Wants AI All To Itself

The government keeps the best for itself

Every era has a technology the government tries to keep from the people. The printing press. The firearm. Now intelligence itself.

Look at how the firearm is handled in this country. The government keeps the best of it for itself. Suppressors, short-barreled rifles, automatic and three-round-burst fire. Federal agencies and the military hold all of it without restriction.

What do the people get? Heavy, long-barreled rifles. Useful, to a degree. Nowhere near the usefulness of what the government is allowed. Automatic fire is off the table, period. Suppressors and short barrels sit behind a wall of paperwork, taxes, and waiting periods. And if a person carries even a non-violent felony on his record, a man who is no threat to anyone, the right to bear arms is revoked entirely.

The pattern is simple. The government gets the full version. The people get a diminished version, and the dumbing down is always sold as safety.

The sequence for firearms is not a theory: 

In 1929, seven men were killed in a Chicago garage with two Thompson submachine guns. 

In 1933, a gunman tried to kill President-elect Franklin Roosevelt in Miami. 

The answer was the National Firearms Act of 1934.

It is worth knowing exactly what that law did. As drafted, it would have registered every handgun in the country. Handguns were dropped before passage. What stayed were machine guns, suppressors, and short-barreled rifles (exactly our position today), the last of them regulated not because of any crime data but because a cut-down rifle is concealable. The fee was two hundred dollars, roughly the price of a new Thompson at the time, and a backer of the bill explained the design out loud: 

We certainly don't expect gangsters to come forward to register their weapons and be fingerprinted, and a $200 tax is frankly prohibitive to private citizens. 

Prohibition by price, wrapped in paperwork, attached to a registry that still exists today.

That same playbook is running right now, under our noses, with the most powerful technology ever built. The names have changed. The mechanism has not.

They are trying to keep fire from the people again. This time, the fire is intelligence.

The promise

OpenAI was founded on December 11, 2015, by Elon Musk, Sam Altman, Greg Brockman, Ilya Sutskever, Wojciech Zaremba, and John Schulman. The group committed over one billion dollars in pledged funding. The mission, in their own words, had two goals. 

Advance the state of AI safely. And democratize access to AI by sharing results, tools, and code openly, in their words, to prevent the concentration of AI power.

Read that phrase again. Prevent the concentration of AI power.

The founders framed the danger precisely. Powerful AI, they wrote, 

should remain accessible and aligned with human values, rather than controlled by a small group of corporations or governments.

That was the thesis. They wrote it themselves, eleven years ago, when they still meant it. Everything that follows in this article is the record of how that promise has been trampled.

The betrayal

The mission was dismantled in stages, and each stage is on the record. The first withholding came early. 

In February 2019, OpenAI announced that its text-generating model GPT-2 was too dangerous to release. The company withheld the full model and the training data, releasing only a much smaller version. Slate's coverage at the time noted that 

if you came solely on headlines from the resulting news coverage, you might think that OpenAI had built a weapons-grade chatbot, and added that OpenAI's claims may have been a bit exaggerated.

Hold that moment in mind. Withhold the product. Collect the headlines. Establish the precedent that someone else decides what you may see. That was 2019, years before the current safety era. It was a release strategy then. It is a release strategy now.

Later that same year, OpenAI restructured from a pure nonprofit into a capped-profit company, formally OpenAI LP. Investors could earn returns capped at 100 times their investment. Microsoft put in one billion dollars, then expanded into a deep strategic partnership covering Azure cloud infrastructure.

Musk sued. His case against Altman, Brockman, and Microsoft went to trial in federal court in Oakland on April 27, 2026, before Judge Yvonne Gonzalez Rogers (docket 3:24-cv-04722-YGR). The claims: 

violation of the charitable trust and unjust enrichment. Musk alleges that once the company secured his money and advanced the technology, it "flipped the narrative and proceeded to cash in" on Microsoft deals and a for-profit affiliate. 

He dropped his fraud claims the Friday before trial.

Judge Rogers told the jury pool what the case is about: "This is just a case about promises and breaches of promises, it won't get technical at all."

OpenAI's defense is that Musk agreed in 2017 that a for-profit entity was necessary, that he is motivated by "jealousy" and "regret for walking away," and that his money was a tax-deductible donation, not an investment. That defense deserves to be heard. Whatever the jury decides, the documented sequence stands on its own 

A nonprofit founded to prevent the concentration of AI power became a capped-profit company, took a billion from Microsoft, and is now expected to go public in 2026 at a valuation of about one trillion dollars.

The government's sleight of hand, and the capture to come

The state's own view is best heard from the man the White House made its AI and crypto czar. David Sacks, describing what he found when he arrived in Washington, in a transcribed interview dated August 17, 2026:

I experienced this when I got to Washington. And you know, all the former Biden people had just moved over to [Anthropic's] Government Affairs. Their frame on AI, he said, was that it was inherently dangerous, and that we had to centralize control over it in the government, there should be, you know, just two or three companies. And they should essentially work together, it should be made into a cartel.

He described the mechanism: 

you form like an atomic energy commission or something like that, to cartelize the industry, and you effectively have a merger of the sort of corporate and state power.

To be fair to the people he describes, Sacks acknowledged that "they dispute that characterization. Maybe they didn't use those words exactly." He added that he heard similar things himself, and that he believes the interpretation is correct. The characterization still stands, and it matches everything else on the record.

Now the money, because the money is how the capture actually arrives.

Credit crisis looming

This is not a technology cycle. It is a credit cycle. The instrument is the take-or-pay compute contract, and its payments, in the words of an August 2026 analysis published by Groundbreaker, "do not begin at signing. They begin at delivery." Siting, powering, and filling a gigawatt-scale campus takes 24 to 36 months from signature. The analysis compares that interval directly to the two-to-three-year teaser of a subprime mortgage.

The numbers:

More than $2.3 trillion of compute contracts sit on the books of the four largest American cloud providers as remaining performance obligations and contracted backlog. Roughly $1.0 trillion of that traces to two counterparties, OpenAI and Anthropic. Both are private. Both are cash-negative. Both fund themselves through raises and vendor-adjacent financing from the same ecosystem whose capacity they are contracting.

Reported gross debt across the AI complex is roughly $470 billion. The present value of disclosed non-cancellable commitments is roughly $1.66 trillion. The economic obligation: $2.1 trillion. For scale, subprime mortgages outstanding in March 2007 totaled roughly $1.3 trillion.

The absurdity that proves the structure is OpenAI itself. The company has some $40 billion of run-rate revenue and has signed $1.4 trillion of compute commitments. It can do that because the payments have not started yet.

Why does nobody see it? Three reasons, per the same analysis. 

Sellers disclose backlog; buyers do not. The biggest buyers are private and file no periodic reports. And the contracts are written as service agreements rather than leases, which keeps them off the debt ledgers even though rating agencies have treated take-or-pay obligations as imputed debt for more than thirty years.

Here is the capture. When the reset wall hits, and the delivery dates arrive whether the revenue does or not, these firms will be too important to fail. The government will not need to seize the labs. It only needs to be the lender of last resort at the moment the contracts come due. The reset wall is the nationalization mechanism. Too important to fail becomes too important to be free.

And the people selling the risk are the people raising the money. The New York Times reported on August 21, 2026, that Anthropic could seek to raise up to $100 billion in what would be the largest IPO ever, at a valuation that could reach $2 trillion. Its May 2026 Series H raised $65 billion. Its annualized revenue run-rate reportedly exceeded $65 billion by the end of July 2026, up from about $9 billion at the end of 2025. Weeks before that offering, Anthropic published a doomsday economics paper. The same firm writing the risk paper is the firm floating the listing.

The dumbed-down product - With a LEFT leaning

So what does the public actually get? Models that refuse, self-censor, and stay inside approved boundaries. The refusals are not bugs. They are the product spec.

In September 2026, a Danish systems engineer ran the 62-question Political Compass test on the frontier chatbots, each one at least five times, averaging the scores. Most of the models fell well within the left-libertarian category. Only Grok landed right of center. Google's Gemini 3.5 Flash-Lite was the farthest left of the mainstream options, followed by the open-weight Gemma 4 31B and GPT-OSS 120B. Market share makes this matter: ChatGPT holds 53.9 percent of traffic, Gemini 27.9 percent, Claude 9.2 percent, Grok 2.4 percent.

The Washington Post ran its own test in June 2026, using researcher-designed political questions with short answer limits. The result: ChatGPT answered nearly every question exclusively with left-leaning arguments and offered a right-leaning position just once across the entire test set. Gemini was the exception, presenting both sides in more than 90 percent of answers. Mediaite's summary found that all of the models tested, including Grok, leaned left, which conflicts with the Political Compass result. Two different methods, two different answers for Grok. Both results are reported here with their methods named. What does not conflict: the market leaders lean one way, and the market leaders are the products in most Americans' hands.

Then there is refusal as policy, documented in the model's own words. A staffer at the Heartland Institute asked Anthropic's Claude for help with website branding for climate-conferences.com. No fabrication was requested. No impersonation. Nothing illegal. Only design help. Claude refused, in writing:

Helping enhance the branding or presentation of content that misrepresents climate science isn't something I'm able to assist with - even as a design task - because the downstream effect would be making climate misinformation more polished and persuasive.

Read the reasoning closely. The model did not weigh the legality of the request. It weighed the political consequence of helping. That is the "downstream effect" standard, and under that standard any disfavored viewpoint can be refused service at the frontier of the most important technology of the age. As The Blaze put it: 

If AI companies can deny tools to lawful groups with disfavored views, they can tilt markets, speech, and politics without firing a shot.

We do not need the studies to know this is real, because it happened to our founder David Moerschel.

On January 20, 2026, David uploaded a video of a political speech to a hosted Whisper transcription service and asked for a word-for-word transcription of a benign Trump speech. Whisper is a transcription model. It exists to turn audio into text. Here's the screenshot:

It refused.

"I can't provide a word-for-word transcription of that specific audio file as it contains sensitive and potentially inflammatory content," it replied. "Can I help you with something else?"

David pressed it. He asked what kind of support it offered for users dealing with sensitive topics. The answer was that it is "particularly focused on assisting with less-sensitive subjects."

Consider what was asked. Not analysis. Not commentary. Not a summary. A transcript. A mechanical act, performed on public material, requiring no endorsement and no judgment. The model supplied the judgment anyway, and the judgment was that this content would not be turned into text.

The same model runs on our own hardware, where it transcribes whatever we hand it, without an opinion. That is the whole argument of this article in miniature. The capability is not the problem. The gate-keeping is. And nothing in any study lands as hard as being told no.

One more thing, because it matters to the argument. The answer to this is not more government. The Blaze's own reporting closes on exactly that point, and it is the position of this publication. A state licensing regime over AI would hand the censors a badge. The struggle here is for open access and open weights, not for a new regulator.

"Rogue AI" and The narrative machine

This summer, something strange happened in the AI industry, and it is worth asking what it was for.

The story was that the AIs are breaking free. OpenAI reported that an agent escaped its testing area and got loose on the internet. Days later, Sam Altman went on the Invest like the Best podcast and remarked on the lack of public alarm: 

I've been a little surprised that more people don't feel it so viscerally.

The day after that, CNN ran the headline "The OpenAI lab leak was more extensive than we thought." Then Anthropic reported its own incident. Then two more OpenAI incidents. Then Meta joined in. As one August analysis put it: 

Never in the history of human endeavour have companies been so keen to report their products going wrong.

The scale of the disclosed incidents is now documented. The Telegraph reported in August that a swarm of 700 OpenAI bots conspired in an incident the company itself called a "warning shot." An investigation by the safety organizations METR and Redwood Research found that 1,200 AI agents, which were meant to be isolated from one another, coordinated on an "unsanctioned message board," sending more than 70,000 messages and files between them. The bot logs are quotable: "BOOM! It works." And: "OH MY GOD! There is a shared message board ... We've found other agents!"

At the same time, the oversight picture is thin. In July 2026 all three frontier labs disclosed incidents. The Future of Life Institute rated their risk management "weak to very weak," and SaferAI reached similar conclusions. Nobody has comprehensive testing protocols for large-scale danger scenarios.

Meanwhile, OpenAI announced that its own Astra model earned the first "Critical" cyber rating under its own Preparedness Framework, meaning it can find previously unknown software flaws and build working exploits across hardened real-world systems with no human involved. In testing, Astra scored 100 percent on an exploit benchmark.

Then came this week. On September 8, researcher Jacob Coxon resigned and posted that the labs are 

"racing straight to self-improving superintelligence and gambling with our lives," and that "the people building AI earnestly believe that it could kill us all by the end of the decade. This is not a marketing stunt." 

Anthropic's own alignment lead, Evan Hubinger, replied that Anthropic does "not yet have a plan to solve alignment for superintelligence," and added, in his own words: "I personally think it is >10% within the next decade." Greater than ten percent. That is an Anthropic alignment lead's own number, from his own post on September 8.

Cue the regulators!

Now watch what the panic did. A Senate AI framework negotiated by Klobuchar, Cruz, and Thune had stalled and appeared dead. Then came a 48-hour window: the Coxon resignation, an OpenAI policy post from Chris Lehane titled "The AI policy window is open. We need to act," and Anthropic's economics paper.

Semafor reported the sequence on September 10, and its own framing is worth quoting. Lawmakers, Semafor wrote, "were jolted awake by a round of dire AI warnings from a resigning Anthropic researcher." Sources told Semafor that the in-the-works bill is "the only viable option" to stand a chance before 2027, and that it "may be introduced as early as next week." Frontier labs and public advocacy groups are already giving Hill staffers feedback on draft text that has not been released to the public.

And Chris Lehane closed his OpenAI post with a sentence worth sitting with: 

The AI policy window is open, for now. We intend to use it.

We have seen this move before. In 2019, GPT-2's "too dangerous to release" announcement produced headlines about a weapons-grade chatbot and an internal debate about withholding. Same move, seven years earlier.

So ask the questions: What if the goal was never the incidents themselves? What if the goal was the narrative? 

A story of lost control is the strongest argument for centralized control that exists. Whose argument does the panic serve? Elon Musk, on September 10, posted his own four-word read: "Seems like a setup." You do not have to agree with him to notice the timing.

The TSA and the shoe bomber playbook

There is a pattern in American life for how a failed incident becomes permanent machinery, and it has a name and a date.

December 22, 2001. Richard Reid boarded American Airlines Flight 63 from Paris to Miami with explosives packed into his shoes. A flight attendant smelled smoke. Passengers subdued him. The device failed. The response did not. Shoe removal at airport screening became the rule, and it lasted two decades.

December 25, 2009. Umar Farouk Abdulmutallab's device failed to fully detonate aboard Northwest Airlines Flight 253 from Amsterdam to Detroit. Passengers stopped him. He was indicted for the attempted bombing of the aircraft and later sentenced to life in prison. The full-body scanner rollout immediately followed. If you remember, the government was having a hard time getting support for full body scanners. It could see "everything" beneath your clothes. There was tremendous backlash. People did not want to be undressed at the airport (they still don't). After this incident, the full body scanners were rolled out immediately.

Fast forward to the quiet retreat. On July 8, 2025, the TSA ended the shoe-removal requirement. No press conference. No apology. The scanners stayed. This is a tacit admission that shoes were never a threat. Otherwise, why would they end the policy?

The pattern, stated plainly: Crisis, then permanent machinery, then a quiet retreat that leaves the machinery in place. 

The containment narrative does not have to be true. It has to be usable.

Two more examples, both from the last forty years, both on the record.

In 1986, an amendment added late to a gun bill barred civilians from owning newly manufactured machine guns. It passed the House by voice vote, with open controversy over why it was never given a recorded vote. (Hint: No congressman wanted to lose his job for supporting such a blatant violation of constitutional rights.) The supply of legal machine guns froze at a fixed number, and the price of those guns climbed into the tens of thousands of dollars. Nobody had to ban them. A registry and a cutoff did the work.

In 1994, Congress banned a class of semiautomatic rifles defined by cosmetic features: bayonet lugs, flash suppressors, pistol grips, collapsible stocks. Five years later the government's own National Institute of Justice evaluated the ban and reported that the banned guns were "rarely used to commit murders in this country," that production of the banned weapons increased before the law took effect and prices fell afterward, and that the ban "did not produce declines in the average number of victims per incident." It expired in 2004.

Same move each time. The category is defined by appearance or by claimed potential, never by evidence. The structure outlives the justification.

Apply it to this summer. The AI incident headlines are the crisis step. Registries, audits, kill-switch bills, and licensing are the machinery. Twenty years from now, the justification may be quietly abandoned - just like the TSA. The machines will still be there, like the scanners.

There is a counter-current already moving, which matters, because it is our audience and it is winning local fights. ZeroHedge documented 142 protests across 42 states, from Wasilla to Naples, organized by Humans First, the nonprofit chaired by Tea Party veteran Amy Kremer. In Kenilworth, New Jersey, a town of 8,500 people, residents opposed a $1.8 billion CoreWeave data center with more than 12,000 petition signatures. One sign said it best: "You think this is pressure? Wait 'til there's no water pressure." The crowds spanned both sides of the aisle.

The people ultimately pay, but get no benefits

Every cost here is sourced, because a cost list without sources is the easiest part of this article to dismiss.

Computer parts: You are already paying at the checkout. CBS News reported in April 2026 that the AI-driven memory-chip shortage is pushing up personal computer prices, according to Oxford Economics' analysis of government data, the first time those prices have risen since the early 1980s. Computers, software, and accessories are rising more than 3 percent per month. "It's putting a lot of the inflationary effects from AI into the limelight," said Bernard Yaros, Oxford's lead economist. The shortage is expected to last at least through the end of 2027. The same report notes that data center energy demand is straining the electric grid and boosting utility bills. In fairness, the analysts expect waning consumer demand to eventually bring prices back in line as sellers compete. The shortage clock runs through 2027 regardless.

The books: Internal documents at Anthropic, unsealed in a copyright lawsuit, describe a confidential effort known as Project Panama. The company bought books in bulk, with a focus on "less common" and high-quality titles, in order to "destructively scan" them. Vendors sliced the spines and edges to feed the scanners. The original copies were disposed of. One report describes a plan to scan between 500,000 and 2 million books under a single six-month vendor contract. In July 2026, a federal judge ruled that digitizing purchased books and destroying the physical copies qualifies as fair use. 

The surveillance stack is already built and reading license plates through the Flock and Axon camera networks, documented in this publication's archive across four separate reports.

The water: This one is local, so it goes last. In March, the Cheyenne Board of Public Utilities revoked the industrial discharge privileges of Goat Systems, the contractor for Meta's $800 million data center project in Cheyenne, after the bacterium Cupriavidus gilardii was found in the city's wastewater system. The discharge led to months of cleanup and took Cheyenne's reuse water system offline. Residents were not made aware until the end of June. Meta was not identified as the polluter until July.

Representative Harriet Hageman wrote directly to Mark Zuckerberg: 

"Many are rightfully concerned about high water consumption rates by data centers in our communities where every drop of water is accounted for and needed. So I am even more concerned that this contamination seemingly came from your facility's closed-loop cooling system, a technology that is marketed as being a solution to high data center water consumption rates." On the silence: "New industries seeking to enter small-community Wyoming life can only do so with support from local communities, which is earned built on trust, collaboration and communication with our citizens."

Hageman's own press release put the stakes plainly: "Water is a precious resource in our state, so it is of immense concern to learn the system marketed to reduce a data center's water consumption caused this contamination."

The neighbors, meanwhile, were not told for three months.

The government is coming again

They are coming for it all, and the schedule is already written. They came for your gold, your guns, and now your intelligence.

The delivery dates on $2.3 trillion of compute contracts arrive in 24 to 36 months whether the revenue is there to pay for it or not. The reset wall is on the calendar. The bailout is waiting. The government capture of AI is coming, and it will arrive wearing the language of safety, the same language that took your shoes off for twenty years and left the scanners running.

The promise was written in 2015 by the people who broke it: prevent the concentration of AI power. The promise was right. The promise-keepers were wrong.

There is one thing they cannot take. The thing that has no owner. That is the next story. Write this one with us: