** WARNING: connection is not using a post-quantum key exchange algorithm.
** This session may be vulnerable to « store now, decrypt later » attacks.
** The server may need to be upgraded. See https://openssh.com/pq.html
The second cypherpunk battle
Seventeen years ago a pseudonymous cryptographer published nine pages of PDF and a C++ tarball, and a handful of strangers on a mailing list decided they would rather run the money themselves than ask a bank for permission. That decision compounds. It is why your rack pulls forty amps. It is why your watt-hour meter matters to you in a way it does not matter to your neighbour. It is why you know what a PSU’s 12V rail droop sounds like on a bad day. You are fluent in a kind of sovereignty that most people have never had to practice.
Now look at the most consequential new technology of the decade. Frontier artificial intelligence — ChatGPT, Claude, Gemini — is centralized intelligence. Every prompt is metered. Every response is logged. Every model is hosted inside a jurisdiction that can, and eventually will, be leaned on. Every API key is a kill switch. Every terms-of-service update is a unilateral rewrite of what you are allowed to think out loud about. This is the correspondent-banking system of cognition, and it was built in eighteen months while nobody who matters to us was watching.
The cypherpunk instinct that built Bitcoin has a second battle. The first battle was money. The second battle is intelligence. Same adversary class, same playbook, same answer: run it yourself, on hardware you own, with code you can read, heated by electricity you paid for. We are calling this sovereign AI, and we are publishing /ai/ because the plebs are already halfway there and don’t know it yet.
What « open-source AI » actually means
If you have never touched a large language model, the vocabulary sounds like marketing. It is not. It maps cleanly onto hardware you already understand.
Model weights are the numbers. Billions of floating-point parameters, trained by someone with a very large GPU cluster, released as a file you can download. Think of weights the way you think of a miner’s firmware image: compiled artifact, signed or unsigned, flashable to hardware you control. Meta’s Llama family, Mistral’s models, Google’s Gemma, Alibaba’s Qwen, DeepSeek’s R1 and V3 — these are open-weight releases. They exist on your filesystem. They do not phone home.
Runners are the execution layer. Georgi Gerganov’s llama.cpp is the bare-metal reference implementation — a single C++ binary that loads a quantized model and serves inference across CPU, CUDA, Metal, ROCm, Vulkan. The Ollama team wrapped llama.cpp in a daemon with a sane CLI and a model registry; it is to llama.cpp roughly what bitcoind is to the reference client for someone who just wants a node running. LMStudio is the GUI-first option for operators who want a desktop app. Pick one. They are not competitors; they are different entry points to the same stack.
Quantization is compression. A 70-billion-parameter model at full precision wants 140 GB of VRAM. Quantized to 4-bit it fits in 40 GB. Quantized to 2-bit it fits on a single consumer GPU, with a quality cost you can measure. This is the knob that makes pleb-scale inference possible.
Context window is how much text the model can see at once. Tokens are the chunks — roughly three-quarters of a word each — that the model thinks in. Inference is what happens when you send it tokens and it sends tokens back. Frontier model is the industry’s term for the largest, most capable models at the research ceiling — historically closed, increasingly matched by open-weight releases.
Interfaces sit on top. Open WebUI is the self-hosted chat interface most operators land on — it talks to Ollama or any OpenAI-compatible endpoint and looks like ChatGPT, except the traffic never leaves your LAN.
Beyond text: Black Forest Labs released FLUX.1 as an open-weight image model of genuinely frontier quality. OpenAI’s Whisper v3 — one of the few things the lab actually released openly — does transcription at human parity. Stability AI’s Stable Diffusion line remains the workhorse for image generation at the lower end. Hugging Face is where most of this lives; it is the GitHub of model weights, hosting hundreds of thousands of open releases.
None of these are D-Central’s work. We did not train a model. We did not write a runner. We stand on the shoulders of Meta, Mistral, Google, Alibaba, DeepSeek, Black Forest Labs, Stability, the Ollama team, the LMStudio team, Gerganov, and the thousands of unnamed contributors at Hugging Face. Our job is to make their work usable inside a Hashcenter that was originally built for a different sovereign workload.
Why the Bitcoiner has an unfair advantage
Look at what sovereign AI actually requires, as a stack of prerequisites, and then look at what the average pleb already has in the garage.
It requires power. Not laptop power — real power. A single 3090 pulls 350 watts under inference load. A four-GPU rig pulls 1.6 kilowatts nominal, 2+ under spikes, and needs a dedicated circuit. Most households do not have this. You do. You have a 200-amp service and you have already balanced it across three subpanels. You know which breaker feeds which outlet and you labelled them yourself.
It requires PSU fluency. You know what a redundant platinum PSU costs, why the 12V rail matters, why a cheap unit sags under transient load and crashes your workload. You have replaced PSUs in the field, at 2 a.m., with the lights off. A normie spinning up inference is going to buy a gaming PSU and complain when the kernel panics.
It requires thermal management. Inference is a heat machine, same as hashing. The power you put in comes out as heat, every joule of it. You already know how to move that heat. You have ducted miners into a furnace return. You have sized extraction fans. You understand static pressure. You have argued with a partner about whether the basement is a workshop or a sauna. Normies do not have this reflex.
It requires rack space and Linux. You have both. You are not afraid of a shell. You have compiled something from source this year. You run a node. You have probably already tried curl | bash more times than is wise, and you know how to read a systemd unit. The median AI-curious consumer in 2026 does not know what a tty is.
It requires kWh tolerance. Your relationship with your power bill is different from most people’s. You treat electricity as a production input, not a household expense. You think in dollars per kWh and hashrate per joule. You will look at an inference workload and immediately ask the right question — tokens per second per watt — before anyone tells you that’s the right question.
And, crucially, it requires an allergy to SaaS. You already refused to hold your coins on an exchange. You already refused to trust a hosted node. You already refused to let Cloudflare terminate your TLS for your private services. You are not going to be thrilled about feeding every thought you have to a logging endpoint owned by a company that answers to a subpoena.
The plebs are sitting on a decade of accumulated sovereign-infrastructure muscle memory, and the only thing missing is the AI-specific piece on top. That piece is smaller than you think.
Hashcenter, not datacenter
At D-Central we call any compute facility optimized around a sovereign workload a Hashcenter. The word is deliberate, and it is worth defining because the rest of this argument depends on it.
A Hashcenter is not a datacenter. A datacenter is a leased, SLA-wrapped, compliance-gated facility selling compute as a service to someone else. Hut 8 runs datacenters. Core Scientific runs datacenters. IREN runs datacenters. TeraWulf runs datacenters. The workload inside is hashing, today, but the configuration is a hyperscaler configuration: raised floors, redundant feeds, enterprise networking, operators in polos, tier certifications, and — most importantly — the intelligence and the revenue flow to a counterparty you do not control.
A Hashcenter is a compute facility optimized for a sovereign workload under individual or small-group ownership. High power density. Commodity hardware you can buy on the open market and service yourself. Thermal design that optimizes for workload output, not aesthetic uptime theater. No SLA, because the operator and the beneficiary are the same person. A basement with four S19j Pros and a ducted return is a Hashcenter. A shed with a Canadian-winter intake and twenty decommissioned Whatsminers is a Hashcenter. A garage with a rack of retired 3090s running Llama 3.3 is a Hashcenter.
Hashcenters predate Bitcoin, technically — the word is ours but the pattern is old. The Bitcoin pleb reinvented the pattern at scale during the 2017–2021 era, when home miners realized they did not need to colocate. The ETH miners did the same thing with GPUs until merge-day. What we are arguing is that the same pattern — sovereign workload, commodity hardware, owner-operated, heat-positive — extends directly to inference, and that the plebs already own most of the infrastructure required.
The distinction between running your own Hashcenter and renting inference from OpenAI is the same distinction as running your own node versus using Coinbase. Everyone reading this knows which side of that line they belong on. The only question is whether the line has been drawn yet in their head for the AI case. We are drawing it.
The three sovereignty layers
Bitcoin sovereignty is two-layered and you already practice both. AI adds a third.
Hardware sovereignty — you own the silicon. You bought it, you racked it, you can unplug it. Nobody can revoke your miner. Nobody can push a firmware update that bricks it without your consent, assuming you control the flash path. The entire reason DCENT_OS exists as a GPL-3.0 open-source firmware project is to harden this layer further. Plebs have this. It is the ground floor.
Heat sovereignty — the joules heat what you want them to heat. This is the layer most miners discover by accident and then get religious about. Your hashing is not just making money; it is displacing your furnace. The kWh you would have paid for hashing and the kWh you would have paid for heating are the same kWh, counted once. D-Central has been building around this thesis for years — we already make heaters that hash, and the Bitcoin-space mining-heater segment has become its own category. Plebs have this too, whether they think of it that way or not.
Weights sovereignty — this is the new layer, and it is the one that AI adds. It means: the model that generates your tokens runs on hardware you control, against weights you possess, without network egress to a third party. No API call. No telemetry. No « we have updated our terms of service. » The model on your disk today will still work the day the company that trained it is acquired, shut down, or subpoenaed. Weights, once released, cannot be unreleased. This is the cypherpunk property — it is the same property that makes a signed Bitcoin transaction final. Once the numbers are out, they are out.
Hardware sovereignty says nobody can take your silicon. Heat sovereignty says nobody can take your joules. Weights sovereignty says nobody can take your intelligence. Stack all three and you have a Hashcenter that cannot be turned off from outside, cannot be repurposed from outside, and cannot be censored from outside. That is what we mean when we say sovereign AI.
What the plebs should actually run — first pass
This section is deliberately short because each item below deserves its own post, and those posts are coming. Treat this as a pointer, not a guide.
For chat and general reasoning: start with Meta’s Llama 3.3 70B or Google’s Gemma 3. Either runs respectably on two 3090s at 4-bit quantization. If you only have one GPU, run the smaller Gemma 3 12B or Llama 3.2 variants — they are good enough to replace most ChatGPT usage for most plebs. DeepSeek R1, in its distilled forms, brings frontier-tier reasoning into the same hardware envelope.
For code: Qwen 2.5 Coder 32B is the current pleb-scale answer. It lives on a single 3090 at reasonable quantization and writes code well enough that D-Central uses it internally for firmware drafting work.
For image generation: Black Forest Labs’ FLUX.1 is the current frontier-grade open model. Stable Diffusion XL is the lighter, faster fallback. Both run on consumer hardware; FLUX wants more VRAM to be comfortable.
For transcription: OpenAI’s Whisper v3 runs on CPU if you are patient and on GPU if you are not. It is the reason you will never pay a transcription SaaS again.
For the runner: Ollama for most people. It is the closest thing to « it just works » in this ecosystem. Drop to llama.cpp directly if you want to tune quantization and batching by hand, or if you are deploying to a server without a package manager you trust. LMStudio if you are GUI-first and want a Mac-native experience.
For the interface: Open WebUI, self-hosted, pointed at your Ollama instance, accessible over your LAN or over your Tailscale mesh.
A starter rig for a pleb already running miners: one or two used 3090s, a 1200W platinum PSU, a 64 GB system with a modern Ryzen, an NVMe large enough for a few models, and a spare 20-amp circuit you are not using for hashing. The total bill-of-materials is less than the cost of a single S21, and the electrical and thermal headroom exists in your Hashcenter already.
Future posts in /ai/ will walk each of these in operator-grade detail: Install Ollama in 10 Minutes, The Pleb’s Guide to Self-Hosted AI, and Heating Your Home With Inference are the first three in the pipeline.
A note on DCENT’s own stack
We are a hardware and firmware company, not an AI lab. Our existing products — DCENT_OS, DCENT_axe, and the DCENT Toolbox — are all closed beta, GPL-3.0, public beta summer 2026. The planned sibling project for GPU inference rigs, DCENT_Inference OS, is in the same bucket: closed beta, GPL-3.0, public beta summer 2026. None of it is a secret. All of it is built on the shoulders of llama.cpp, Ollama, the Linux kernel, and the open-weight releases from the labs above.
The reason we are publishing /ai/ now, before DCENT_Inference OS is public, is that the open-source AI stack that already exists is sufficient for almost every pleb use case today. You do not need us. You need Ollama, a GPU, and an evening. We will publish our own tooling when it is ready, and we will credit everything it is built on. In the meantime, the stack is the stack, and the stack works.
The hardware lottery has already been won
Here is the part that should feel like a homecoming.
Between 2017 and 2022, plebs bought GPUs by the crate. The ETH-mining era left a sedimentary layer of used Nvidia silicon — 3080s, 3090s, A4000s, P40s — sitting in bins, in closets, in sheds, on shelves labelled « for parts. » That hardware was, for a while, a depreciating asset. Ethereum went proof-of-stake in September 2022. GPU hashing became uneconomic on the main chains. A lot of that silicon was sold at a loss, or mothballed, or repurposed into gaming rigs that didn’t need that much VRAM.
That silicon is now an appreciating intelligence asset. A used 3090 with 24 GB of VRAM, bought for retail in 2020 and worth $400 on the second-hand market in 2023, is today the single most cost-effective piece of inference hardware a pleb can own. The VRAM is the constraint; the VRAM is what you already have. Quad-3090 rigs that were built for ETH are Llama 3.3 70B rigs without a single hardware change. The risers, the PSUs, the open-frame chassis, the 20-amp circuits, the PDUs, the extraction fans — all of it transfers. The only thing that changes is the workload on top.
Bitcoin miners sitting in sheds represent a different version of the same argument. The ASICs will not run Llama — inference is a GPU-shaped workload, not an ASIC-shaped one — but the facility will. The 200-amp service, the thermal design, the cable runs, the racks, the ambient-temperature sensors, the HVAC that moves real air: that is a Hashcenter, and a Hashcenter with a single GPU rack bolted on becomes a dual-workload Hashcenter, hashing and inferencing from the same envelope. On cold nights you tilt toward inference. On warm ones you tilt toward hashing. The joules all end up as heat. The heat all ends up in your house.
The plebs already won the hardware lottery. They won it twice — once with ASICs, once with GPUs. They already own the power, the thermal envelope, the sovereignty reflex, the Linux fluency, and the refusal to rent. What is missing is the software layer on top, and that layer is open, free, and installable tonight.
Why we are publishing /ai/
D-Central is a Bitcoin mining company. We sell miners, we make firmware, we build heaters that hash. We are not pivoting to AI and we are not rebranding. What we are doing is extending — additively — into the adjacent sovereign workload, because the audience is the same and the infrastructure is the same and the fight is the same.
The fight is this: intelligence, like money, is being rebuilt in the shape of a permissioned service. A small number of well-capitalized labs, sitting inside a small number of jurisdictions, are becoming the sole providers of the cognitive substrate that every other knowledge worker, every other business, every other citizen is increasingly dependent on. This is exactly the configuration Bitcoin was built to refuse. We refused it for money. We will refuse it for intelligence.
The case for self-sovereign LLMs is being argued from the cryptographic-protocol side already. We are arguing it from the hardware-and-heat side. They are the same argument. Open weights, local inference, owner-operated hardware, encrypted transport, no third-party kill switch. The full stack of sovereignty.
The plebs are our audience because the plebs are who this actually works for. You already have the 200-amp service. You already have the PSU fluency. You already have the Linux comfort. You already have the refusal-to-SaaS reflex. You already have the spare GPU in the closet from the ETH days. You already have a Hashcenter. What you do not yet have is the weights on the disk and the runner on the boot, and that is a one-evening project.
One more layer decentralized. Bitcoin decentralized money. Open-source AI decentralizes intelligence. Your Hashcenter decentralizes both. Welcome to /ai/.
Bitaxe Heatsink — High-Performance Aluminum Cooler for Bitaxe & Nerdaxe Gamma / Supra / Ultra — Silent Operation & Stable Overclocking" width="80" height="80" loading="lazy" style="width:80px;height:80px;object-fit:contain;border-radius:6px;background:#1A1A1A;flex-shrink:0;">
Acheter les dissipateurs
