Sola Computa

2026-03-07

In 1517, the Catholic Church monopolized the interpretation of scripture, not scripture itself. The Bible was in Latin; you needed a priest to read it to you and tell you what it meant, the priest needed a bishop, and the bishop needed Rome. The institution provided real value: it preserved learning, maintained communities, gave moral structure to daily life. But structurally, the arrangement was a monopoly. The cost of understanding scripture included the cost of sustaining an interpretive supply chain that could not be bypassed.

The printing press made the priesthood optional. Vernacular Bibles existed before Luther’s, but the press made his the first to reach thousands of households. The cost of understanding scripture dropped to the price of literacy. A farmer could hold a printed Bible in his own language, read it by the fire after harvest, and decide for himself what it meant. The intermediaries survived for generations, but the economic logic that sustained them collapsed in a decade.

We are building a new priesthood, and this time the scripture is intelligence itself.

Storing files on someone else’s server outsources custody; reasoning through someone else’s model outsources judgment. The provider’s training choices determine what the model engages with, its guardrails bound which questions produce useful answers, and its operator can change the model, retire it, or adjust its behavior without asking. The dependency runs deeper than infrastructure; it is intimate, which is the right word for reasoning through someone else’s model.

Three movements have tried to cut out the middlemen, and all three solved the access problem while ignoring the maintenance problem. The test for any disintermediation movement: does it make maintenance cheaper than dependency? Until LLMs, none did.

Three Partial Reformations

Open source gave you the code. The code was free. Maintaining it was not. About 96% of codebases incorporate open-source components,¹ yet 60% of maintainers work entirely unpaid.² Widely used infrastructure routinely depends on a handful of volunteers; when they burn out, critical projects lose their maintainers. One engineer self-hosted his email for 23 years: configured the MX records, maintained the spam filters, kept the server patched, fought the blacklists. After two decades he gave up. “My emails simply aren’t delivered,” he wrote. The code had always been available. The maintenance never got easier.

Right to repair won legal access but left the expertise gap intact. The FTC sued John Deere in January 2025, alleging the company forces farmers through its dealer network for software repairs. Winning the right to fix your tractor doesn’t give you the diagnostic expertise to do it.

Blockchain solved distributed trust at the cost of distributed complexity; most users chose centralized exchanges, and custodial wallets on Coinbase recreated exactly the intermediary structure blockchain was supposed to eliminate.

Same mistake, three times: sovereignty made possible without being made easier. Free but hard to run. Legal but hard to do. Trustless but hard to use. Each solved the access problem and left the maintenance problem untouched.

LLMs close the maintenance gap. An LLM can read logs, correlate errors with recent changes, generate a patch, run tests, and roll back if the tests fail; precisely the cycle that made self-hosting impractical for twenty years. Today it still requires human supervision: the model proposes, a human approves. But reviewing a proposed fix takes minutes where diagnosing from scratch takes hours. The threshold is cheaper-than-delegation, not perfection, and for a growing set of systems it’s already crossed.

Exotic Hardware to Laptops

In 2024, running a GPT-4-class model locally required a rack of GPUs that cost more than a car. By early 2025, a single $2,000 card. By early 2026, I run Alibaba’s Qwen3.5 on a laptop, 27 billion parameters quantized into 16 gigabytes of RAM, and it approaches frontier commercial models on most benchmarks. Two years from exotic hardware to a machine I carry through airport security.

Inference cost falls by an order of magnitude every hardware generation, and quantization cuts memory requirements by 75% with near-zero quality loss. Mixture-of-experts architectures fire only a fraction of the weights per token. At this rate, phones by 2028, and the marginal cost of local intelligence converges on electricity.

“Not Your Weights, Not Your Brain”

Every model carries its maker’s choices: Qwen carries Alibaba’s, LLaMA carries Meta’s, and every set of weights has baked-in biases and blind spots. Local inference gives you choice: which model to run, when to swap it, how to fine-tune it, whether to run two and compare their outputs. Sovereignty means choosing which biases you import, not escaping bias entirely.

The deeper issue is cognitive sovereignty. A cloud provider determines what its model will engage with, and those boundaries shift with each update, shaped by legal exposure, public pressure, and commercial strategy. Any particular boundary may be reasonable; the problem is that someone else sets it, shifts it, and owes you no explanation. None of this requires malice; it requires only the structural fact of dependency. A local model breaks that dependency. The weights sit on your machine; no one can alter them, retire them, or restrict what you ask. Most people will choose convenience, the way most people chose a new denomination rather than reading scripture for themselves. But the option to run your own model disciplines every provider that remains, because the user who leaves is no longer hypothetical.

The internet democratized information; the capacity to reason about it stayed expensive. A founder reads a term sheet full of liquidation preferences, anti-dilution ratchets, and participation caps, and realizes she needs a $1,200-an-hour lawyer to understand what she’s signing. The bottleneck was always the cognitive work of applying information to a specific situation. Local AI collapses that cost. The same machine that keeps your reasoning private also makes expert-level cognition available without the expert, or the expert’s fees, and that pressure extends well beyond software. Every institution that mediates between people and their own reasoning faces the same structural question the Church faced when scripture became something anyone could read.

Sola Computa

Sola scriptura, by scripture alone, meant the priesthood lost its monopoly on interpretation. It happened because powerful institutions found the new arrangement served their interests; the peasants followed.

Sola computa, by your own compute, means the platforms lose their monopoly on intelligence. For software, the driver will be cost: someone with budget authority looks at the bill. For everything else, the driver will be something harder to price: the right to think through your own machine. The first reformation replaced a monopoly with a market of institutions that earned authority rather than inheriting it; this one will do the same, except that when the underlying capability runs on every device, no institution can lock you in. They persist only while they’re useful.

The reformation made the intermediary optional, and once optional, the intermediary had to earn its place. That process took centuries and was not gentle. The AI reformation will be faster, but the pattern will rhyme: intermediaries will fight, adapt, and in many cases survive by becoming useful in ways they never bothered to be before. The difference is that this time the scripture runs on your own machine, and no one can take it back.