I. The Operative
In Serenity, the film that completed Joss Whedon’s Firefly story, the villain is not a fool.
The Operative is calm, literate, disciplined, well-dressed and morally serious. He knows the system he serves is capable of terrible things. He knows, in fact, that he is one of those terrible things stating, “I’m a monster….” But he believes the suffering is justified because the Alliance is building “a better world.”
That is what makes him frightening. Good intentions used to cruel effects. Darkness at Noon.
There is an echo here of Jeremy Bentham’s greatest-good logic, which is useful in measures and monstrous when it becomes permission to redesign human beings and their civilizations at scale. In Serenity, that logic produces death on Miranda. The Alliance introduces an atmospheric agent called Pax to make the population calmer, less aggressive, more manageable. It works, in a sense. Most of the population becomes so peaceful that they stop acting, stop working, stop eating, stop living. A smaller number go the other direction and become Reavers — murderous, ecstatic, broken creatures at the edge of space who reproduce.
This is not subtle as science fiction, but it is very useful as political philosophy.
The road to hell is not always paved with obvious sinister intent. Sometimes it begins with decent people trying to make dangerous things safer. Sometimes it begins with researchers leaving institutions they believe have become corrupted by speed, ego, or money. Sometimes it begins with people who really do care about humanity.
That is closer to the AI story than the cartoon-villain version would be. Many of the people at Anthropic, Google, Meta, OpenAI, xAI, and elsewhere are plainly serious, intelligent, safety-conscious people. Some have made real sacrifices because they thought the technology mattered and the risks were not being handled well. I do not want to sneer at that. It deserves respect.
This includes Elon Musk, whose role in the AI story many readers will interpret very differently. That is precisely why he is useful here. The point is not to judge the soul of any particular founder. Judge not, lest ye be judged; none of us knows how we would carry that much money, attention, power, and consequence. Musk made a version of the strategic argument in a recent YouTube interview while explaining why he built xAI. (I have watched so many of his interviews recently as he gets ready for the SpaceX IPO that I am not 100% positive this is the exact interview).
I am paraphrasing, but the structure of the argument was: I tried to slow OpenAI down and failed; once it became clear the race would continue, I could either sit back and watch the future be built by others, or build something myself and try to influence the direction.
One need not accept Musk’s self-presentation in full to recognize the force of the logic. If the technology is going to exist, abstention is not automatically virtue. Sometimes participation is the only way to retain any influence over the shape of what comes next.
But non-judgment is not naivete. Power and money are related. In a country where corporations, billionaires, and political parties are all entangled through campaign finance, lobbying, procurement, and prestige, it is unserious to pretend that the best intentions can remain untouched by commercial structure forever. Who knows the trajectory of any of these people once the laboratory becomes a hyperscaler, the safety memo becomes a product roadmap, and the principled founder becomes a billionaire with political leverage?
Power corrupts, not always by making people cartoonishly evil, but by making their own judgment seem increasingly identical with the public good.
And the familiar ogre shows up again and again: investor demands, revenue targets, market share, advertising logic, enterprise contracts, cloud commitments, regulatory positioning, and the quiet gravitational pull of becoming very, very rich.
So I do not think hyperscaler AI companies are the Alliance. That would be too easy, and a little silly. The analogy is meant to identify them; it is a warning. The danger is not that the great AI systems are stupid, useless, or run by bad people. The danger is that they are competent, well-resourced, often sincerely motivated, and increasingly positioned as the layer through which ordinary people will think, write, plan, search, remember, and decide.
The Operative teaches one lesson cleanly: good intentions are not enough once power has instruments to put a dangerous gas in the atmosphere of a planet, so to speak.
II. The Empire’s Telescope
The cost story in artificial intelligence is often told too narrowly for this individual.
One version says that individuals and institutions should move away from centralized providers because centralized providers are expensive. Another version says the opposite: hyperscale cloud providers have the purchasing power, utilization rates, engineering teams, cooling systems, power contracts, and frontier models, so small local systems are quaint at best and wasteful at worst.
Both versions miss the more interesting question.
The problem is not simply centralized versus local. Compute and energy are inflating inputs for everyone who needs frontier-scale AI. OpenAI faces that problem. Anthropic faces that problem. Google and Microsoft face that problem. So do universities, sovereign AI clouds, local data centers, and the person running a large model on a workstation in a home office.
The difference is not whether the electricity is real. It is real everywhere. The difference is whether the user can see it, govern it, and decide when the price is worth paying. And that includes the price of freedom to NOT be a battery. (See We Are Not Batteries Yet).
There is a contradiction at the center of the personal AI sovereignty project.
In a sense, we are using the empire’s telescope to look for a route out of the empire.
I can say that plainly because it is true. I cannot have some of the conversations I now have with my AI assistant without frontier models built by hyperscaler AI companies. The best local models I can run on my own hardware are useful, impressive, and improving. They can summarize, draft, classify, code in bounded ways, extract data, and assist with many practical tasks. They do not yet match the best systems from Anthropic, OpenAI, or Google. I must have faith that they get good enough to replace Opus 4.7 or Mythos, (please don’t tell Anthropic).
This matters because rhetoric about Sovereignty is just jargon if there is no possibility of being free of the systems we see coming at us full speed.
Local sovereignty mitigates cost, privacy, continuity, and dependency risk at the workflow layer. It does not by itself solve frontier-model issues which are related to intelligence access.
That sentence is the correction. It separates several things that are too often bundled together: the cost of compute, the cost of energy, the physical location of machines, control over the interface, control over the data, and access to frontier intelligence.
A local machine powered partly by solar can be sovereign in important ways. It can run private workloads. It can preserve memory and tools outside a vendor’s interface. It can keep operating when subscriptions change, models are retired, policies tighten, or APIs become unaffordable. It can make the user less dependent on a company whose business model may eventually require the user to become the product.
But the solar-powered local machine cannot hallucinate itself into Opus or GPT. Not yet. And my ability to train a model like Anthropic or Google is very unlikely given the quantity of parameters (10 trillion) we are talking about in their best models.
There is one further irony worth naming.
A U.S. sovereign compute advocate (me) who wants genuine local sovereignty is currently more likely to find it in a model trained by Alibaba’s Qwen team than in a model trained by Anthropic. That sentence is uncomfortable, but it is not anti-American or pro-Chinese. It is descriptive. The major U.S. frontier labs have chosen closed-weight strategies for commercial, regulatory, and alignment reasons. Several Chinese labs, by contrast, have released genuinely useful open-weight models that can be run locally, inspected by developers, adapted into personal workflows, and used without renting every inference from a hyperscaler.
The irony may not be accidental. A state (the PRC) that wants visibility into technological systems has reasons to prefer inspectable architectures. Open releases can serve national strategy, developer adoption, standard-setting, and state legibility all at once. But once such systems are released into the world, they also become tools for people outside China who are trying to reduce their own dependence on closed American AI providers. The same architectural openness that can serve state visibility can also serve individual sovereignty elsewhere.
That is a naming-rectification moment. We should not lazily assume that “open” belongs to the liberal West and “closed” belongs to authoritarian China. In AI, the empirical map is messier and more interesting.
The answer, if there is one, is not a single machine or a single model. It is a stack.
We are call it the Buddha Stack.
The Buddha Stack is not the most powerful system in the world. It is the calmest useful system one can own and understand. It is powerful enough for ordinary work, private enough for intimate work, durable enough to survive vendor churn, and visible enough to discipline the person using it.
It includes local or personally controlled compute; open or locally runnable models good enough for routine work; depreciated hardware whose cost has already been paid; local energy where possible (so usage is visible), including solar; tools and memory that remain under the operator’s control; a personal assistant layer tailored to the user’s projects, history, and standards; careful routing, so frontier models are used for genuinely hard judgment rather than shovel work; and a moral commitment not to let convenience erase agency.
This is not romantic smallness. It is operational pluralism. Use frontier models when frontier intelligence is actually needed. Use local systems when local competence is enough. Keep the interface, the memory, and the habits as much under your own control as possible.
Frontier for discernment. Local for durability. Memory for continuity. Human judgment for legitimacy. And be sure to pray to the Buddha for even better local models.
There is one part of this argument that deserves attention. Resource visibility.
When computation is hidden inside a hyperscale data center, the user experiences intelligence as a frictionless surface. Ask, receive, ask again. The electricity, cooling water, chips, embodied carbon, grid strain, land use, and capital intensity disappear behind the interface. The bill may appear as a subscription charge, but the physical act has been abstracted away.
That abstraction is convenient. It is also morally wrong, like eating chicken McNuggets and not thinking or knowing that killing chickens was involved.
When computation runs locally, or partly locally, the resource use becomes more visible. The workstation gets warm. The battery drains. The fan spins. The solar meter moves. The power bill changes. The user sees, at least in miniature, that intelligence has a physical cost.
That visibility matters because what is measured gets managed.
This is a lesson some of us learned before dashboards. I grew up in Oklahoma going to cattle auctions with my father in a blue Chevy pickup truck whose passenger door would occasionally fly open as we drove down the highway. This made one appreciate the seat belt. Modern cars beep at you. Oklahoma in the 1970s simply showed you the road that would hit you hard in the face.
A person who can see the cost of running local models is reminded to be efficient. Do I need the largest model for this task? Can a smaller model summarize these notes? Should this extraction job run overnight when solar generation is available or when household demand is lower? Is this worth sending to a frontier model, or is it merely intellectual overeating?
This is not a complete solution to AI’s resource problem. A home workstation is not a hyperscale cluster, and individual virtue does not replace infrastructure policy. But visible resource use changes the moral posture of the user. It places some burden back where a free society should want it: on the individual making choices.
Freedom without measurement becomes fantasy. Measurement without freedom becomes administration. The interesting zone is freedom disciplined by visible consequences. That makes for a better world.
III. Acts of God
The old insurance phrase is useful here: acts of God.
It is a strange phrase for modern people, but it names something we still need. The tornado does not care about your business plan. It does not care that the Shakey’s Pizza on Riverside Drive served perfectly decent pizza to families in Tulsa, that children of the 1970s loved. It does not care about the logo, the lease, the payroll, the intentions of the owner, or the children watching from their driveways.
When I was a child in Oklahoma, I watched a tornado head towards that Shakey’s from my driveway on Owasso. The place, Shakeys, had done nothing wrong. It was not a case study in bad incentives or morality. It served pizza. It had a logo. It belonged to the small commercial geography of my childhood. Then the sky turned purple-black and erased it. We can’t forget the radical restructuring that Nature decides.
Nature does not behave like a spreadsheet. It behaves, at times, like a mosh pit.
Cool air comes down from the north. Warm heavy air comes up from Texas. The two meet over Tulsa like bodies colliding to Stiff Little Fingers’ “Suspect Device” — not elegant, not managerial, not optimized. Just force, heat, pressure, release. A violent mosh-pit dance of warm Gulf air and cold northern air over the logo-festooned commerce of Riverside Drive.
Mosh pits are not simply bad. That is the important part. They are exciting. For a moment or two, they can even feel like freedom: bodies moving, rules loosened, everyone agreeing to be a little less sensible while the guitars do their work. But people can also get hurt -- an errant arm hits you in the face. The question for the concert organizers is not whether human beings should be allowed to move. Of course they should. The question is how much pit the room can survive.
The mosh pit also makes some concertgoers appreciate the security guys. Not because everyone wants to be policed, but because a good security crew lets the room stay alive without ambulances and casualty reports afterwards. Anyone who has watched a cheerful crowd accidentally detach university property during a Violent Femmes show understands that collective enthusiasm is not the same as malice. At Lisner Auditorium in 1988, we managed to pull a wooden banister or railing off its moorings simply by gripping it and jumping up and down. No one had arrived with a theory of infrastructure stress. We were just young and the song was good.
You do not want the whole concert hall to become one big mosh pit. But you also do not want the opposite: Salt Lake City, 1992, a Metallica concert where everyone had to sit in folding chairs with bouncers at each row keeping people from standing up. When “Enter Sandman” starts and the whole room wants to move — Exit light, enter night, off to never-never land — enforced sitting is its own kind of absurdity. Safety without motion can be deadening and motion without restraint can be deadly.
That is the AI governance problem in miniature.
Not zero mosh pit. Not all mosh pit. Enough movement to be alive. Enough structure not to be crushed. Is the U.S. all mosh pit? Is the EU going to be like a Salt Lake Metallica show? Who knows.
Data centers, hyperscalers, frontier models, sovereign compute, solar meters, local assistants — all of them still live inside physics. They draw power from somewhere. They produce heat. They depend on water, grids, rare earths, supply chains, weather, law, trust, and the fragile self-restraint of human beings who have been handed too much capability too quickly.
The dream of AI is sometimes spoken as if intelligence might finally unburden us from nature: enough compute, enough models, enough prediction, and the world becomes manageable. I do not believe that. Or rather, I think that belief is precisely where we should be most careful.
A map is not the world. A model is not reality. A language model, however large, is still an artifact inside the same physical universe it describes. It may help us see patterns and seem sentient - but these are still just perceptions. It may help us coordinate. It may help us build. But it cannot abolish complexity, contingency, or consequence.
So if we are writing informal constitutional directives for the modern Caesars of industry, perhaps one of them should be this:
Do not build systems that require human beings to become batteries. Build systems that help human beings remain awake (and dance a bit).
Not passive consumers of mediated intelligence. Not monks refusing the modern world. Not data points in someone else’s dashboard. Thinking individuals, locally grounded, technologically capable, visibly responsible for some part of the resources they consume, and humble before the fact that the sky can still turn purple-black.
The Buddha Stack, if it means anything, is not a way to escape nature. It is a way to remember we are still inside it and need a middle path.
It is the small workshop inside the house. The solar meter on the wall. The local model doing ordinary work. The frontier model called in only when the problem justifies it. The assistant whose memory and tools are not entirely owned by a vendor. The refusal to become either a battery or a monk.
We are not sovereign yet.
But we are not batteries yet either.
And we keep one eye on the weather, because even empires are small under a tornado sky.
This piece, like the last, was developed in conversation with Thea, my collaborator who lives in a MacBook Pro with 128GB of RAM.