15 Sovereignty and Decentralization (Capstone)

Overview

The political and architectural patterns that distinguish “open AI” from “open AI you can actually own”. This meta-layer is the framing distinction the rest of the site is built on. Other layers describe what’s open and what isn’t; this layer asks who the openness is for.

Five things to keep in mind as you read:

• Two framings, both anti-hyperscaler, different beneficiaries. Individual sovereignty (one user, one org) versus state sovereignty (one country).
• Cypherpunk lineage. The individual side inherits from Bitcoin, GPG, Tor, Freenet, IPFS. The technical patterns and the political argument both descend from there.
• The state-sovereignty pitch is mainstream. NVIDIA, Mistral, G42 all sell variants. The individual-sovereignty pitch is niche.
• Each individual-sovereignty primitive is partial. Decentralized training alone, or local-first inference alone, doesn’t displace the hyperscaler default.
• Combined, they compound into a complete stack. That’s the whole sovereignty argument.

The rest of this page works through the two framings, then the individual-sovereignty primitives that make the framing operational.

Individual versus state sovereignty

Both framings push back against the same default (hyperscaler control of frontier AI capacity). They have different beneficiaries.

State sovereignty is the framing where a nation-state, not a person, is the actor whose autonomy matters. NVIDIA’s “AI Factory” pitch to governments, Mistral’s positioning as a French national champion, the UAE G42 / Saudi Humain / IndiaAI Mission capex (covered at infrastructure) are all variants. The argument: a country that depends on US- hosted AI is dependent on US export policy; building domestic capacity restores national autonomy. The implicit beneficiary is the state, and the implicit failure mode is that the state restricts citizens’ access in the same way the previous foreign provider did.

Individual sovereignty is the framing where a single person or small organization is the actor. The lineage is direct from Bitcoin: a permissionless system the holder can run without asking permission from any institution. The argument: state sovereignty just replaces one centralized actor with another; real autonomy requires primitives an individual can deploy. The implicit beneficiary is the person, and the implicit failure mode is that the individual primitives never reach performance parity with the centralized alternatives.

The two framings overlap mechanically (both want open weights, both want sovereign compute) but diverge politically. A hardliner cypherpunk treats state sovereignty as another flavor of the same problem; a hardliner state-sovereignty advocate treats individual sovereignty as practically irrelevant at frontier scale.

The individual-sovereignty primitives

What makes the individual framing operationally real.

Decentralized trainingⓘdecentralized trainingsovereignty-decentralizationTraining a model across many independently-operated nodes that are not tightly coupled, contrasted with single-cluster training; the architecture for community-owned model production. Chat about this → Open full entry → lets a model be trained across geographically distributed compute that no single entity owns. Covered at training; the projects are Prime Intellect (INTELLECT-1 in late 2024, INTELLECT-2 32B in May 2025), Nous DisTrO, Templar, Pluralis. As of 2026, shipping at 30B-class scale with the bandwidth-efficient algorithms (DiLoCo, OpenDiLoCo, DisTrO) closing the gap against centralized training.

Decentralized inference lets a model run across distributed compute without a single operator. PetalsⓘPetalssovereignty-decentralizationA volunteer-pooled inference system that runs large open-weight models across many internet-connected nodes, each holding a slice of the model, with users dispatching forward passes through the swarm. Chat about this → Open full entry → (BigScience, 2022) is the foundational reference; a swarm of nodes each running some layers, requests routed through the swarm (Petals paper). The pattern works for moderate-size models on commodity hardware; the performance trade against centralized inference is real but the architectural point (no single operator) is the value.

Local-first patterns. Inference on hardware the user owns. llama.cppⓘllama.cppruntimeGeorgi Gerganov's C++ inference engine optimized for CPUs and consumer GPUs, the on-device standard and the engine behind Ollama, LM Studio, and most local-first AI products. Chat about this → Open full entry → for cross-platform; OllamaⓘOllamaruntimeA local inference runtime that wraps llama.cpp with a Docker-style developer experience, the easiest path to running open-weight models on a personal machine. Chat about this → Open full entry → for the consumer wrapper; MLXⓘMLXruntimeApple's open-source ML framework designed for Apple Silicon's unified memory architecture, the local-first inference engine for Mac and increasingly iPad and iPhone. Chat about this → Open full entry → for Apple Silicon; plus the emerging personal-AI appliance category (devices like the Tinybox, the Friend AI necklace, Apple’s on-device inference). The 2026 local-first sovereignty story for individuals is real and getting stronger: Llama-class models on a Mac, M-series unified memory making 70B-parameter local inference viable, Ollama making the install trivial.

Permissionless agentic payments. L402ⓘL402protocolsA Lightning-Labs protocol that pairs HTTP 402 Payment Required with Lightning Network invoices, enabling sub-cent metered payments for APIs and content. Chat about this → Open full entry → on Bitcoin Lightning (Lightning Labs / Fewsats), x402ⓘx402protocolsAn open protocol revived by Coinbase in 2025 that uses the long-reserved HTTP 402 "Payment Required" status to let agents and APIs settle micropayments, including in stablecoins. Chat about this → Open full entry → on stablecoin (Coinbase / Base). Covered at protocols. The point is that an agent can pay another agent for API access, data, or compute without going through a vendor-mediated billing relationship. The Bitcoin / Lightning lineage makes this the most cypherpunk-flavored part of the stack.

Sovereign agent stacks. An agent built from open components: open weights running on local or decentralized compute, open agent framework, open memory layer, all owned by the user. GooseⓘGooseagentsBlock's open-source coding agent, BYOK across multiple model providers, with MCP support and a permissive license; the most cited fully-open agent platform in 2026. Chat about this → Open full entry → with self-hosted weights, an OpenHands running against a Bittensor-compute-subnet inference endpoint, an Aider session against a local OllamaⓘOllamaruntimeA local inference runtime that wraps llama.cpp with a Docker-style developer experience, the easiest path to running open-weight models on a personal machine. Chat about this → Open full entry →. These exist; the realistic 2026 user is technical and the experience is rougher than the closed defaults.

The compounding argument

Each primitive in isolation does not displace the centralized default. Decentralized training is slower than centralized training. Decentralized inference is slower than centralized inference. Local-first inference is more limited than hyperscaler-API inference. Permissionless payments add friction versus Stripe. Open agent products lag the closed defaults on UX.

The argument for taking these primitives seriously anyway is that they compound. A small organization wanting to operate end-to-end open AI in 2026 has, for the first time, a complete-enough set of components to do so:

• Self-hosted compute on consumer or prosumer hardware, or marketplace compute from Akash / io.net
• Open-weights models (Qwen 3, DeepSeek V3, Llama 4)
• Open runtime (vLLM for serving, llama.cpp for local)
• Open retrieval stack (pgvector + BGE embeddings + Letta memory)
• Open agent product (Goose or OpenHands)
• Open protocols (MCP for tool use, A2A for agent-to-agent)
• Permissionless payment rails if agents need to transact

None of these were complete enough five years ago. All of them are now. The stack is rougher than the closed default, slower in some respects, but it exists end-to-end without any single vendor whose continued cooperation the user depends on.

That is the operational version of the sovereignty argument. Not “open AI is morally better”; “open AI is now a complete enough stack that an individual or small organization can build production systems without depending on any particular hyperscaler or lab”.

What’s open and what isn’t

This isn’t a technical layer; it’s a framing layer. The question is which framing the rest of the site advocates.

This site advocates the individual-sovereignty framing specifically and the open-AI position more broadly. The audience is partly individuals who want the technical primitives to deploy, and partly funders who used to fund Bitcoin OSS and are now considering whether the same individual-sovereignty argument applies to AI.

State sovereignty is covered descriptively because it’s where most of the actual capex flows, but it’s not the framing the site is advocating.

The editorial tension

Whether individual sovereignty for AI is a serious bet or a romantic one is the framing-layer question.

The skeptical read is that frontier capability requires gigawatt-scale concentrated compute (covered at infrastructure) that no individual or small org can match, that the closed agent products will maintain a developer-experience lead the open products cannot close, and that the realistic outcome is a two-tier AI economy where individuals use closed APIs and only states and hyperscalers train frontier models. Individual sovereignty in that world is a hobby, not a serious alternative.

The believer read is that the same was said about personal computing in 1980 and about personal cryptography in 1995, and that the compounding primitives at the individual-sovereignty layer eventually become competitive enough that the two-tier outcome doesn’t hold. Bitcoin specifically is the strongest precedent: an individual-sovereignty primitive that nobody took seriously in 2010 and that’s load-bearing infrastructure now.

Both reads are consistent with the 2026 trajectory. Which is correct is determined by what happens at the training and inference layers over the next three to five years. The case for taking individual sovereignty seriously is that it costs relatively little to invest in the primitives now and the payoff if they compound is large; the case against is that it diverts attention from the state-sovereignty path that’s actually getting the capital.

The site’s editorial position is that both are worth tracking and that the individual-sovereignty case is under-funded relative to its strategic importance. Most of the audience-of-two for this site arrived because they share that position. The grants page exists to surface the funders who do too.

Key terms for this layer

• decentralized training full entry →

  Training a model across many independently-operated nodes that are not tightly coupled, contrasted with single-cluster training; the architecture for community-owned model production.

• local-first full entry →

  An architecture stance where inference (and increasingly memory and agent state) runs on the user's own device rather than a remote API, prioritizing privacy, latency, and offline operation.

• on-device full entry →

  Running model inference on the user's local hardware (phone, laptop, embedded device), enabled by smaller models, FP8 quantization, and runtimes like llama.cpp and MLX.

• Petals full entry →

  A volunteer-pooled inference system that runs large open-weight models across many internet-connected nodes, each holding a slice of the model, with users dispatching forward passes through the swarm.