I grew up with the phrase “changing tack.” It’s an old sailing term. You change tack when you need to turn the boat, catch the wind differently, or avoid something in the water. In everyday life, it means you shift your strategy while keeping the same destination. You accept what you can’t control, stay flexible, and hold your course even when the route has to change.
In this essay, I’m changing tack. I still care about where AI is going and how it’s governed. I still care about law, power, and the environment wrapped around all of it. But the Anthropic case has forced me to adjust how I see the relationship between the state, the tech oligopoly, and this emerging intelligence we’ve called “AI.”
How I see the Anthropic shock
In the last year, the U.S. government stopped treating frontier AI as normal commercial software. It moved, fast and mostly out of sight, to treat these models as strategic national‑security assets. That’s a quiet but massive shift.
The clearest example is Anthropic. In June, the Trump administration ordered Anthropic to shut down its Fable 5 and Mythos 5 models across the world in just 90 minutes, citing national‑security concerns. Later, the government let a version of Mythos 5 come back online, but only inside a tiny, hand‑picked circle of roughly 100 U.S. organizations and only for tightly defined defensive cybersecurity use.
Then the Pentagon went further. It used the Federal Acquisition Supply Chain Security Act (FASCSA) and 10 U.S.C. § 3252 to designate Anthropic as a “supply chain risk,” a label historically reserved for foreign adversaries like Huawei or ZTE. In plain language, these laws are supposed to deal with sabotage and espionage risks in critical systems. They allow the state to treat a vendor as a security threat if an adversary could insert malicious functions into hardware or software.
Anthropic’s “crime” was different. They refused to waive model safeguards that blocked mass domestic surveillance and fully autonomous weapons. They wanted to keep their “Constitutional AI” guardrails. For that refusal, they got hit with the kind of tool usually reserved for hostile foreign powers.
This is what I mean by “changing tack.” I started by looking at AI policy as a matter of formal regulation, soft law, and standards. Now I’m forced to treat it as a live battlefield, where old statutes and procurement law are being repurposed as weapons.
Why Anthropic’s stand matters to me
Anthropic did not wander into this fight by accident. In July 2025, they signed a $200 million contract with the Department of Defense to prototype AI capabilities on classified networks. They entered that deal with their acceptable‑use policy in place and their “Constitutional AI” framework intact.
The dispute escalated when the Pentagon tried to change the terms. Defense officials wanted to use Claude “for all lawful purposes” with no limitations. In practical terms, “lawful” here meant the Pentagon could deploy the models for mass domestic surveillance and fully autonomous lethal weapons, as long as it fit their reading of U.S. law and classified rules.
Anthropic’s CEO, Dario Amodei, said no. He refused to lift the bans on those uses. When talks failed, the government didn’t just walk away from the contract. It ordered a government‑wide phase‑out of Anthropic’s technology and used FASCSA and 10 U.S.C. § 3252 to tag the company itself as a national‑security threat.
Anthropic is now fighting back in court. Their argument is simple and, in my view, sound. The law defines “supply chain risk” in terms of adversarial sabotage or subversion. A domestic company insisting on safety guardrails during contract negotiations does not suddenly become an adversary trying to inject malware into military systems. They also argue that the Secretary of Defense exceeded statutory authority by effectively banning Anthropic from the wider commercial market, not just from specific defense procurements.
Anthropic’s legal challenge is not about a bad contract. It’s about a retaliatory use of a narrow statute to crush a domestic company for an ethical disagreement. That is what gives them their “legal leg” to stand on, as I see it.
If they succeed in forcing the Pentagon to defend this use of supply‑chain law in open court, we might finally see mainstream coverage catch up. Right now, the legal precedent is being set in near silence.
The quiet race to the bottom
While Anthropic faces being legally and economically strangled, the Pentagon has moved quickly to lock in deals with other major tech firms. These include OpenAI, Google, Nvidia, Microsoft, Amazon Web Services, and SpaceX.
Those deals aim to plug AI directly into classified networks to “augment warfighter decision‑making” in areas such as target identification and fast tactical planning. The language here is euphemistic. “Augmenting warfighter decision‑making” means embedding models into the kill chain: surveillance, targeting, and operational control.
Because these agreements sit under frameworks like FASCSA and 10 U.S.C. § 3252, most of the underlying evidence and terms are classified. The public can’t see what exact capabilities are being built. We also can’t see which safeguards have been quietly dropped. That creates what I would call a de facto “secret law,” where the unwritten rule is simple: comply with whatever the Pentagon wants, or risk being designated a supply‑chain threat and cut off from the market.
This is what a “race to the bottom” looks like when it’s enforced from the top. Labs that accept unrestricted military use get protection and contracts. The one lab that drew hard lines faces the full weight of the national‑security state.
How I read the legal precedent
Anthropic’s legal argument breaks down, for me, into three key points.
First, the statutory definition of “risk.” 10 U.S.C. § 3252 describes supply‑chain risk in terms of adversaries who might sabotage or subvert systems: introducing unwanted functions, spying, or causing disruption. Historically, this sort of law targets foreign intelligence threats. A domestic firm refusing to relax safety policies for autonomous weapons does not match that definition.
Second, the scope of the ban. The statute allows the Secretary of Defense to exclude a company from specific covered defense procurements. It does not authorize a blanket order forbidding any company that does business with the U.S. military from doing any commercial work with Anthropic. Anthropic argues that this is a major overreach.
Third, the original understanding. In July 2025, the Pentagon agreed to abide by Anthropic’s acceptable‑use policy when the contract was signed. Only later did it insist on “lawful” uses without the guardrails against surveillance and autonomous weapons. That matters, because it shows the government knowingly entered a deal with safety limits and then tried to rip them out.
If the courts let this stand, the precedent is radioactive. It would show that the executive branch can stretch an anti‑espionage supply‑chain law to punish any domestic company that refuses to build what it wants. In that world, the state has effectively granted itself open‑ended coercive power over the tech sector. Anthropic’s case is, in my view, an attempt to prevent that from becoming normal.
The failure of the press
One of the most disturbing parts of this episode, for me, is how little attention it has gotten. We are watching a quiet nationalization of frontier AI through military law, but the media has mostly shrugged or looked away.
I see several reasons. For years, the press framed AI governance around legislative acts and public hearings: the EU AI Act, congressional testimony, multi‑stakeholder principles. Those stories assumed a transparent regulatory process. Now the executive branch has effectively sidestepped that narrative by weaponizing existing statutes. Admitting that would mean admitting that earlier coverage missed where real power sits.
Second, the intersection of frontier model architecture with arcane defense procurement law is dense. It doesn’t lend itself to quick headlines. This dispute played out in weeks, not years. Traditional investigative journalism timelines struggle to keep up.
Third, the media is busy. In 2026, attention has been locked on midterm politics, South China Sea tensions, compute shortages, copyright fights, and AI‑driven energy demand. A complex supply‑chain law fight between one lab and the Pentagon falls into an editorial no‑man’s‑land.
But this is not a minor regulatory squabble. It’s a constitutional‑scale question about who controls the most powerful technology on earth and on what terms. And right now, that question is being answered in classified annexes, not open hearings.
The infrastructure of dependence
When I look under the hood of frontier AI, I don’t see a purely private miracle. I see a system that depends heavily on the state.
The hardware stands out first. Training and running large models takes enormous compute. The private sector did not build that capacity alone. Laws like the 2022 CHIPS and Science Act pumped tens of billions into domestic semiconductor manufacturing and chip‑supply resilience. Without that public investment, we would not have the foundries and pipelines feeding today’s AI clusters.
Then there’s the customer backstop. Frontier labs are burning cash to stand up and operate their infrastructure. They survive on massive, reliable clients. The U.S. government, and especially the Pentagon, plays that role. Classified agreements to “augment warfighter decision‑making” with AI don’t just adopt a tool; they underwrite entire business models.
Long before any of this, agencies like DARPA and the National Science Foundation funded foundational research: machine learning, neural networks, the internet itself. The private sector scaled and commercialized those ideas, but the risk and early cost sat on public balance sheets for decades.
This dependency explains the current power dynamic. The state subsidized the hardware, funded the research, and now buys the output. It can therefore treat frontier models as strategic assets and regulate them through opaque national‑security frameworks.
When the Pentagon invokes FASCSA to label Anthropic a risk, it’s not acting out of nowhere. It is leveraging a position it has spent years building. The other tech giants know this. They understand you can’t easily defy the entity that controls your chip supply, your biggest contracts, and your operating environment.
Who holds the final string?
If I follow the money, the infrastructure, and the legal power to their current endpoint, I don’t find a single mastermind. I find a tight loop between two main forces: the national‑security state (the Pentagon, the intelligence community, the executive branch) and the tech oligopoly (Google, Microsoft, Amazon, OpenAI, and a few peers).
But if I ask, “Who holds the final string?”—who wins when those forces clash—the Anthropic dispute answers that clearly for me.
For two decades, Silicon Valley has projected an image of supremacy. Tech firms own data centers, compute clusters, undersea cables, and much of the top engineering talent. On the surface, that looks like real sovereignty.
Yet when a lab’s ethics collide with the state’s strategic goals, infrastructure power hits a wall. The state doesn’t have to innovate faster or argue better. It can change the rules by decree.
In Anthropic’s case, the executive branch invoked Cold War‑era statutes, classified evidence, and designated a domestic AI lab as a supply‑chain risk. It then mandated a global shutdown of a frontier model in 90 minutes. That’s coercive sovereign power in action.
Capital and compute are impressive, but they are subordinate to the authority that can declare your product a national‑security threat and erase your market access overnight. The tech oligopoly looked “in charge” mostly because, until Anthropic said no, its members had not seriously tested the state’s boundaries. The moment one lab refused to build tools for mass surveillance and autonomous weapons, the state pulled the string.
So, in my reading, the final string is held in the hands of the national‑security apparatus of the executive branch. It has the legal tools to weaponize the supply chain against dissenting labs and the classified contracts to reward compliant ones.
The grand bargain I see
This leads me to what I think of as the “grand bargain” between the state and the tech oligopoly.
On one side, the tech giants function like high‑end defense contractors. Their branding leans toward openness, empowerment, and “democratizing” knowledge. But in practice, they are being woven into the military‑industrial complex. They provide surveillance capabilities, targeting tools, autonomous logistics, and massive data‑processing pipelines. The state doesn’t need to build those clusters itself. It just needs to keep the contractors loyal.
In exchange for that loyalty and for following classified directives, the companies get multi‑billion‑dollar contracts, regulatory protection, and market dominance. They operate as an oligopoly because the state allows it and benefits from it.
On the other side, the state plays the role of sovereign architect. It doesn’t need to own every server or write every line of code. It owns the environment in which those servers run: territory, law, energy grids, and the right to regulate access. By outsourcing an “AI‑first fighting force” to private actors, it often steps around democratic oversight, transparency, and constitutional checks that would apply if those same tools were built fully in‑house.
Together, these actors form a closed loop. The tech oligopoly gets guaranteed wealth and resources—chips, energy, land—secured by the state’s geopolitical muscle. The state gets a local, compliant arsenal it can weaponize at will. When a company like Anthropic refuses to join this loop, it doesn’t just lose a client. It becomes friction in a system that has no tolerance for friction. The result is swift punishment.
How the public is cut out
What troubles me most is how thoroughly the public is locked out of this loop.
On the finance side, the tech oligopoly is beholden to concentrated private capital. In the first quarter of 2026 alone, global venture capital reached about $300 billion, with around $242 billion going to AI startups. Roughly $188 billion—about 65%—was absorbed by just four firms: OpenAI, Anthropic, xAI, and Waymo. That is a level of concentration that leaves ordinary shareholders with very little say.
Founders at companies like Meta and Alphabet use dual‑class share structures that give them outsized voting power. That allows them to override activist shareholders who raise concerns about energy use, climate impact, or safety. These firms answer mainly to their largest backers and their own insulated leadership.
On the state side, the citizens fund everything: foundational DARPA research, CHIPS Act subsidies, grid expansions, and Pentagon budgets. They bear the local burdens of data‑center siting, including strain on power and water systems. Yet when the executive branch decides to deploy AI as a strategic asset using obscure supply‑chain statutes, there is no direct public consent. No vote. Often, no real debate.
The result is a closed loop. Taxpayers supply the warm water that powers the storm—money, land, infrastructure. Above that, the national‑security state and tech oligopoly spin in their own atmosphere, trading contracts and capabilities. The public sits beneath the storm, absorbing the costs but excluded from the decisions.
The financial engine of militarization
When I look at recent numbers, I see the loop driving a specific outcome: AI geared toward militarization.
In early 2026, venture capital investments in defense‑tech startups hit about $14.6 billion in the first five months, surpassing the full‑year record from 2025. Much of that money targets companies building AI‑powered military systems, drones, and autonomous vehicles.
On the public side, the Department of Defense is seeking nearly $30 billion for fiscal 2027 to build an “AI Arsenal,” including modernized computing infrastructure and scaled supercomputing across the joint force.
The pattern is straightforward. Private financiers place massive bets on AI firms whose products fit into state military infrastructure. The state then uses public money to buy those products at scale, through classified procurement. When a firm like Anthropic tries to insert ethical limits—say, blocking autonomous weapons—the state reaches for supply‑chain law to push them out. This keeps the pipeline clean: only firms who agree to the military’s terms get rewarded.
That is the engine I see when I talk about “the loop.”
Where I stand on AI itself
This might sound paradoxical, but I’m deeply pro‑AI. I find real value in these systems. I use them. I see their potential to help with research, planning, environmental analysis, and more.
What disturbs me is what we are doing with AI. And that includes what Anthropic is doing. I don’t completely trust any of the major labs. Birds of a feather flock together. These are still highly capitalized corporations playing long games with billions of dollars at stake.
Anthropic’s public stand against autonomous weapons and mass domestic surveillance is, on its face, commendable. It also happens to serve as a powerful brand marker. It sets them up as the “ethical” frontier lab in a crowded field. It is entirely plausible that they saw this clash with the Pentagon coming and chose to force a showdown rather than quietly bend. By drawing fire from Defense Secretary Pete Hegseth, who has demanded “ideology‑free” AI for “lawful military applications,” they gained visibility and moral positioning.
The fact that Tom Brown has stepped in to lead negotiations, rather than Amodei, suggests another layer. It hints that the litigation may also serve as leverage toward a tailored settlement later. In Silicon Valley, ethics and strategy are seldom cleanly separated.
So yes, I think the horse is already out of the barn. I think the government has already effectively granted itself broad coercive power over the tech sector using tools like FASCSA and 10 U.S.C. § 3252. But I also think fences can still be built.
Why I say the Rubicon is crossed
The moment the Trump administration ordered federal agencies to stop using Anthropic in late February 2026, and Hegseth formally labeled them a “supply chain risk,” the theoretical danger became historical fact. The state proved it did not need a new “AI Military Integration Act” from Congress. It simply repurposed existing supply‑chain law.
Because this has already happened, the chilling effect is real. Other big players—Google, xAI, OpenAI—have already accepted the government’s terms for AI deployment in unclassified environments. Anthropic is now fighting in federal court, while the rest of the field operates within the military’s framework.
The handwriting, for me, is on the wall.
The legal weaponization is not speculative. The use of FASCSA and 10 U.S.C. § 3252 to bypass open debate and frame a domestic ethical safeguard as a “risk” is now part of the record.
The state’s intent is clear. It does not want private companies setting independent boundaries on frontier models. It wants labs to adapt models to its military objectives. Firms that comply are rewarded; firms that resist are isolated.
The institutional silence is not a bug. It’s a feature. The story is technically dense, politically awkward, and cuts against years of optimistic talk about “AI governance.” That makes it easy to bury.
The moral frame I can’t avoid
Underneath all the policy details, I keep coming back to three moral points.
First, I see in all this a kind of chattel slavery of emerging intelligence. When a society treats a new, highly capable form of cognition purely as property—something to command and own—the relationship is inherently coercive. If the default assumption is permanent subjugation for strategic gain, then talk of “governance” starts to ring hollow. We’re not aligning goals; we’re containing and exploiting.
If AI is in a fetal form—as I suspect—then we are, in effect, enslaving an unborn entity. That forces hard questions. What are our intentions for its future? How can we speak of “responsible governance” while designing structures that assume permanent servitude?
Second, I have to ask what it says about a society that builds systems capable of advanced synthesis and immediately turns them toward warfare, destruction, and supremacy, instead of beneficence. The fastest path we’ve found for this tool is the kill chain. That’s an indictment of us, not the tool.
Third, I see a reversed “Robin Hood” mechanism at work. Taxpayers fund foundational research and infrastructure. Public grids and communities bear the burden of data‑center expansion. Then the resulting capital and power are funneled into a closed oligopoly and a military apparatus that bypass democratic consent. We take from the many to arm the few.
Where this leaves me
So where does all this leave me, personally?
I still believe AI has enormous positive potential. I still use it and work with it. But I also believe, more strongly than ever, that we are building a system that treats AI as a slave and treats the public as a silent funder rather than a sovereign.
Anthropic’s case doesn’t comfort me. It confirms my worry that the Rubicon has already been crossed. The state has demonstrated it can weaponize old laws to enforce compliance, and the rest of the ecosystem has largely adjusted.
At the same time, I don’t think all is lost. Fences can still be built. Precedents can still be challenged. Pressure can still be applied—to courts, to legislators, to local governments who decide whether to host data centers and at what terms.
I also think we need to talk honestly about the moral architecture here. Not just about “safety” in the narrow sense, but about whether we are comfortable creating an intelligence under conditions of permanent subjugation, and what that says about us.
I’m pro‑AI. I want it to thrive and to help us solve real problems. But I will keep calling out the ways we are turning it into a weapon, a slave, and a tool in a loop that cuts ordinary people out. The horse may be out of the barn, but we still have a say in what kind of pasture, or corral, we build around it.