Authoritarian and Democratic Inference

In June 2026 the Council on Foreign Relations published a long assessment of where the technology is going, and buried in it is a sentence that should make any operator stop reading and look up. China’s state-centric model, it argues, “could prove better suited to deploying autonomous systems at scale than the EU’s rights-based framework”. Read past the geopolitics and what is being said is simpler and older: concentration of control is an advantage, and the side that concentrates fastest will win. The frame treats that as a fact about the world rather than a choice about how to build. It is the establishment voice of an idea that has been around far longer than AI, and a man named Lewis Mumford gave it a name in 1964.

I want to argue that two identical streams of tokens, produced to the same quality, can carry opposite politics for the single reason that the machine producing one sits in a datacenter you petition and the machine producing the other sits on your desk. That is not a metaphor. It is a claim about architecture, and it was made precise sixty years ago by people who had never seen a GPU.

Mumford’s two technics

Lewis Mumford, in a 1964 lecture called “Authoritarian and Democratic Technics”, drew a line that has nothing to do with how advanced a technology is. He said that two technical systems can be equally sophisticated, equally productive, and still belong to opposite political families. The first he called authoritarian technics: powerful, centralized, system-centered, organized around a single command structure that the human being feeds and obeys. His ancient example was the megamachine that built the pyramids, thousands of people turned into components of one engine directed from the top. The second he called democratic technics: small-scale, person-centered, durable, under the direct control of the worker who uses it. His example was the village craftsman with hand tools, modest in output but autonomous in operation.

Mumford’s sharpest point is that authoritarian technics does not win because it is more efficient at any single task. The craftsman’s tool often did the job perfectly well. It wins because it produces a surplus of power that flows to whoever sits at the center of the system, and that surplus is the actual product. The pyramid was incidental. The megamachine was the point.

Now hold a rented frontier API next to that description. Capability concentrated in a handful of datacenters. A command structure you do not sit inside. A relationship in which you supply the input, the prompts and the payment, and receive the output on terms set without asking you. You feed it and you obey it, and when the system decides your access is over, it is over. That is authoritarian technics in Mumford’s exact and technical sense, not as an insult but as a structural description. The desk machine, by contrast, is small-scale and person-centered. It produces less. It is under the direct control of the one person operating it. The same tokens come out of both. The politics built into the two paths point in opposite directions, and they point that way before anyone writes a single prompt.

Winner: the artifact decides before you do

The objection writes itself, and it is the right objection. Surely the politics is in how you use the thing, not in the thing itself. A hammer builds a house or breaks a window. The tool is neutral and the human supplies the meaning.

Langdon Winner answered that objection in 1980 in an essay called “Do Artifacts Have Politics?”, and his answer is the warrant that holds Mumford up. Winner’s claim is that some technical arrangements are not neutral instruments awaiting a user’s intention. Their design distributes power before anyone uses them, by the simple fact of how they are built. His most cited example is the set of low overpasses on the parkways of Long Island, built so low, he argued, that buses could not pass under them, which kept the people who rode buses, disproportionately the poor, off certain roads and beaches. The politics was poured into the concrete. No operator using the bridge well or badly could undo it. The architecture had already decided.

Winner’s second category is sharper for our case. Some technologies, he argued, are inherently political: their internal logic requires a particular distribution of authority to function at all. A nuclear weapons system demands a centralized, secretive, hierarchical command chain because nothing else can safely run it. Other technologies, like a solar collector on a roof, are compatible with a decentralized and democratic order because they do not require a center to operate. The technology and the politics are not separable. One implies the other.

This is the move that lets us say the API and the desk box are not the same tool used two ways. A hyperscale inference system requires concentration: the capital, the supply chain, the energy, the control plane all live at the center because the architecture cannot run otherwise. A model on owned hardware does not require a center, which is precisely why it can be operated by one person who answers to no one. The artifact sets the default. Winner’s word for the people who deny this is the people who believe technology is “neutral,” and his entire essay is forty pages of showing why that belief is a way of not looking.

The bridges, and the fight over what they meant

It is worth slowing down on those overpasses, because they are the most argued-over example in the whole field, and the argument is the lesson. The version Winner tells comes from Robert Caro’s biography of Robert Moses, The Power Broker. Caro reports that Moses had the parkway bridges to Jones Beach built unusually low, low enough that public buses could not fit under them, so that the people who rode buses, disproportionately poor and Black New Yorkers, were kept away from the beach while the cars of the wealthy passed freely beneath the same concrete. A racial sorting poured into the height of a bridge.

Here is where I have to be honest, because this blog does not get to keep an example just because it is convenient. Later scholars have contested the account. They have questioned whether the heights were actually chosen to exclude anyone, and pointed out that buses could reach Jones Beach by other roads. The neat story of a villain encoding his prejudice in masonry may be tidier than the record supports, and you should know that before you repeat it.

But notice what survives the demolition. Whatever Moses did or did not intend, the bridges were built at the height they were built at, and that height did decide who could pass and who had to find another route. Winner’s actual claim never depended on the villain. His claim is that the arrangement of an artifact distributes power before anyone uses it well or badly, and a low bridge does that whether malice or a survey error put it there. Intent is a question for the biographer. Effect is a property of the structure. The contested evidence does not weaken the point. It is the point, stripped of its melodrama.

That is exactly the shape of the API. Nobody has to intend exclusion for a rented control plane to decide who computes and on what terms. The height of the bridge is the architecture, and the architecture is already sorting traffic before the first bus, or the first prompt, arrives.

A control-plane move is a political act

This stopped being theory for me on a specific date. For its first months the sovgrid.org stack ran its public edge through a Cloudflare Tunnel, which meant a vendor’s ops team could, in principle, change my routing in response to a request I would never see. On 2026-05-24 that tunnel was retired in favor of direct Caddy and Let’s Encrypt on a rented edge, and the stack moved from 5/6 to 6/6 on its own sovereignty audit, the control-plane dimension flipping from rented to owned. The honest framing at the time was that this was a sovereignty win and not a security win. I gave up a CDN that absorbed abuse I cannot absorb myself, and in exchange I took on the operational responsibility of defending the origin.

Here is the part that maps onto Winner. When the abuse comes now, the response is not a vendor allowlist that decides centrally who is permitted to reach me. It is a set of targeted DOCKER-USER firewall drops at the origin, rules I write and own, that block the specific scanner traffic without handing a third party the authority to decide who counts as a visitor. Those are two technically valid ways to keep a vuln-scanner off the site. They produce the same outcome at the level of bytes. They distribute the authority to gate my own traffic in opposite directions: one to a company, one to a config file in my repository. Choosing the firewall rule over the allowlist was not a performance decision or even mostly a security decision. It was a decision about where the power to decide should sit. That is architecture as politics, made in practice, on a Tuesday, by someone who at the time would not have used those words for it.

The audit number is the receipt that the choice was real. 5/6 to 6/6 is not a feeling. It is a dimension that flipped because a centralizing artifact was removed from the path and a decentralizing one put in its place.

The objection, steelmanned, and the answer

Now the hard part, because a sovereignty essay that skips its strongest counter is just decoration. Winner’s own essay contains the seed of it. Many artifacts, he conceded, are flexible. They are compatible with more than one political arrangement, and which one obtains depends on the social choices made around them. A desk box run by someone who never reads its logs, never changes a default, never exercises any of the control the architecture makes available, re-centralizes nothing and frees no one. It is a more expensive way to be passive. The politics, the objection runs, was always in the use. An owned machine operated thoughtlessly is authoritarian technics that happens to sit on your desk, and a rented API used by an operator who keeps a fallback and reads the terms is less captured than the architecture suggests. The artifact is not destiny.

I concede all of that, because it is true and because Winner himself would. But conceding it does not dissolve the claim. It sharpens it. The architecture does not determine the outcome. It sets the default that the outcome has to fight against. The rented path defaults to concentration, and an operator has to actively work to claw back any control, which the architecture is built to make hard and often impossible. The owned path defaults to autonomy, and an operator has to neglect it to lose what the architecture is built to grant. Both can be overridden by how you behave. But one starts you pointed at the center and the other starts you pointed at yourself, and starting position is most of where most people end up. Winner’s bridges did not force anyone to be poor. They made one outcome the path of least resistance and another the path of constant effort, and that is how artifacts do their political work: not by compulsion but by gradient. The desk box gives you a gradient you can actually climb. The API gives you one you mostly cannot.

The CFR sentence is where this gets concrete and serious. When the establishment says concentration is better suited to deploying systems at scale, it is making Winner’s neutral-technology error at the level of national policy, treating a political choice about who holds the control plane as a technical fact about what works. It might even be right that concentration deploys faster. The megamachine built the pyramid faster than any village of craftsmen could. Speed at the center was never the question Mumford was asking. The question was what the surplus of power does and where it pools, and the answer, in 1964 and in 2026, is that it pools wherever the architecture sends it.

What I am actually claiming

I am not claiming the desk box is more capable. It is not, and at the frontier it never will be. The box on my desk has never once been described as too low for a bus, which is the only metric on which it beats the bridge. I am not claiming owning a machine makes you free regardless of what you do with it, because the steelman is correct that a neglected machine frees no one. And I am not claiming the API is run by villains. It is run by engineers solving a real concentration problem honestly, the same way the bridge engineers were just building bridges.

What I am claiming is narrower and harder to wave off. The two ways of getting the same tokens are not one neutral tool used well or badly. They are two artifacts with two politics, in Mumford’s vocabulary and Winner’s, and the politics is poured in at the level of architecture, before the first prompt, the way it was poured into the overpasses before the first bus. The rented API distributes the power to decide toward a center you petition. The owned model distributes it toward a desk you control. You can fight either default, but you should at least know which one you are standing on, because the people telling you that concentration is inevitable have a stake in your believing the artifact has no politics at all. It does. I watched a single audit dimension flip from a 5 to a 6 the day I moved one.

This is essay six of a series, and like the radical monopoly piece before it, it concedes its strongest objection before answering it rather than after. The spine of the argument, and why each essay does that on purpose, lives on the philosophy page. The structured and complete version is the forthcoming book, for which these essays are the public workshop. Start with the philosophy page if you want the whole shape of the thing.