The Math Says We Cannot Read Your Prompt

2026-05-08T00:00:00+00:00

Most cloud AI privacy guarantees are contractual. The provider says: we won’t read your prompts.</em> You believe them, or you don’t. There’s nothing about the system that prevents them from reading. The OS user on the inference box can read process memory. The deploy account can pull database rows. Anyone who compels the provider can compel the data.</p>

This is also true of Archipelag.io for Tier-0 commodity workloads, and we want to be honest about that. A workload running in a Docker container on a contributor’s GPU is visible to that contributor’s docker exec</code>, their tcpdump localhost</code>, their memory snapshots. A prompt that flows through the coordinator is visible in NATS and in PostgreSQL. We promise we don’t look. We can’t prove we didn’t.</p>

For most workloads — public chat, image generation, code formatting — that’s fine. The same trust model is what every standard LLM API runs on, and it’s been good enough for the last three years. But there’s a class of workloads where contractual privacy isn’t enough. Medical notes. Legal drafts. Internal company data that nobody outside the company is supposed to see. For those, we want to ship something stronger.</p>

Today we’re shipping the first piece: Tier-1 confidential inference on Apple Silicon.</strong></p>

What “Tier 1” actually means</h2> We split workloads into three privacy tiers:</p> Tier</th> Name</th> Coordinator can read?</th> Island operator can read?</th></tr></thead> 0</strong></td> Commodity</td> Yes</td> Yes</td></tr> 1</strong></td> Hardened</td> Yes</td> No</strong></td></tr> 2</strong></td> Confidential</td> No</strong></td> No</strong></td></tr> </tbody></table> Tier 0 is what we’ve always had. Tier 1 closes the Island side: the contributor running the hardware can no longer read the prompt or the response. Tier 2 closes the coordinator side too — the operator (us) can’t read it either. Tier 1 is shipped today on Apple Silicon. Tier 2 is on the roadmap (P2 in PRIVACY_ROADMAP.md</code></a>).</p> This work is heavily inspired by Eigen Labs’ April 2026 paper on private decentralized inference</a>. They sketched the access-path-elimination approach for Apple Silicon; we took that approach, applied it to a consumer compute network, and shipped working code.</p> Anatomy of a confidential job</h2> Here’s what happens when you submit a job with privacy_tier: 1</code> to a macOS Island today.</p> 1. The Island attests itself</h3> When a Mac mini or MacBook joins the network as a Verified Island, it doesn’t just register. It runs MacOSIsland.bootstrap()</code>:</p> Hardware identity.</strong> It generates a P-256 keypair inside the Secure Enclave</em>. The private half never leaves the SE — there’s no API to extract it, even with root. The public half goes into the Island’s persistent identity.</li> Hardening checks.</strong> It calls PT_DENY_ATTACH</code> to refuse debugger attachment, queries csr_check()</code> to confirm SIP is enabled, and verifies that no kSecCSDynamicInformation</code> flags are set on its own running binary. If any of these fail, bootstrap aborts.</li> Binary self-hash.</strong> It hashes its own running binary (runningBinarySHA256</code>) so the coordinator can compare against an allowlist of known-good Island builds.</li> Fetch a nonce.</strong> It asks the coordinator for a fresh challenge nonce — a one-time random value, expires in 60 seconds.</li> Sign.</strong> It builds a canonical-JSON attestation blob — {hardware_id, security_posture, binary_hash, nonce, public_keys}</code> — and signs it with the SE-bound key.</li> Submit.</strong> POST to /api/v1/island/apple-silicon/attest</code>. The coordinator verifies the signature against the public key, checks the nonce is fresh, validates the hardening claims, and either issues a 15-minute blessing or rejects the Island outright.</li> </ul> A Verified Island has privacy_tier ≥ 1</code> in the database, a binary_hash</code> we recognize, and a blessed_until</code> timestamp. Re-attestation runs every 10 minutes; a single transient failure doesn’t drop you, but anything sustained does.</p> 2. The consumer encrypts to the Island’s hardware-bound key</h3> When a consumer submits a Tier-1 job, the SDK first hits /api/v1/island/encryption-target</code> to get the picked Island’s current</em> encryption pubkey (the SE-bound one from attestation). The consumer:</p> Generates an ephemeral P-256 keypair in their own RAM. Single use.</li> Computes ECDH(consumer_priv, island_pub)</code> → 32-byte shared secret.</li> Derives a session key with HKDF-SHA256 over a fresh 16-byte session_id</code>.</li> Seals the prompt with ChaCha20-Poly1305, fresh 12-byte nonce.</li> Sends the envelope (consumer_pub_xy</code>, session_id</code>, nonce</code>, ciphertext</code>, tag</code>) to the coordinator.</li> </ul> The coordinator persists the envelope verbatim. It cannot decrypt — it has no private key on either side of this exchange.</p> 3. The Island decrypts inside the Secure Enclave’s domain</h3> The coordinator forwards the envelope to the assigned Island over a Phoenix Channel. Inside the Island’s hardened process:</p> The SE-bound private key + the consumer’s pubkey produces the same shared secret. Same HKDF over the same session_id, same key.</li> ChaCha20-Poly1305 verifies the Poly1305 tag (constant-time) and decrypts the ciphertext.</li> The plaintext exists only</em> inside this process’s address space, on this hardware, for the duration of the job.</li> llama.cpp runs the inference (TinyLlama 1.1B Q4_K_M for our smoke test, larger models for production).</li> The response is re-sealed with the same shared secret, a fresh</em> session_id and nonce — never reuse a (key, nonce) pair — and published as a single sealed chunk.</li> </ul> 4. The consumer decrypts locally</h3> The coordinator routes the sealed response back. The consumer derives the same shared secret (their saved private key + the published Island pubkey), unseals, recovers the plaintext.</p> End to end, the prompt was readable in exactly two places: the consumer’s RAM (where it was typed), and the inside of the hardened Island binary on attested hardware (where inference happened). Nowhere else. The coordinator saw ciphertext only.</p> We verified this end-to-end on a real Apple M4 yesterday — full round-trip, encrypted prompt → SE-attested island → TinyLlama inference → encrypted response → consumer-side decrypt.</p> What this actually protects against</h2> A non-exhaustive enumeration:</p> A curious or compromised Island operator.</strong> They can’t docker exec</code> into the inference process — there’s no Docker container. They can’t snapshot memory without disabling SIP, which the attestation refuses. They can’t extract the private key — it’s in the Secure Enclave. They can’t substitute their own binary — the binary hash is in the attestation blob.</li> A coordinator operator who tries to read prompts.</strong> They have the ciphertext and nothing else. ChaCha20-Poly1305 with a 256-bit key derived from a fresh ECDH is not breakable in practice.</li> Network observers.</strong> Same answer; everything sensitive is encrypted under hardware-bound keys.</li> </ul> What this does not</em> protect against</h2> We promised honesty about scope. Here’s what Tier 1 doesn’t do:</p> A compromised coordinator operator (i.e. us) who acts maliciously.</strong> The coordinator decides which Island handles your job. A malicious coordinator could route every job to a single attacker-controlled Island that’s been reverse-engineered. Tier 1 closes the Island side of the trust model; the coordinator side stays open until Tier 2 ships. Tier 2 puts the coordinator itself inside an Intel TDX or AMD SEV-SNP TEE so even we</em> can’t see which job went where.</li> A determined attacker with physical hardware access.</strong> PT_DENY_ATTACH</code> and csr_check</code> are speed bumps, not bunkers. Someone with the actual Mac mini, willing to disable SIP, attach JTAG, or run an unsigned kernel extension, can defeat both. The Secure Enclave still protects the private key — but they could read live memory before encryption. We mitigate by routing only to Islands whose hardening posture is currently</em> attested; we do not claim full physical resistance.</li> Side-channel attacks on the SE itself.</strong> Apple has had a good track record here, but no chip is perfectly side-channel-resistant. If you’re modeling a nation-state attacker, this isn’t your threat model.</li> Linux Islands.</strong> Today only Apple Silicon Islands can serve Tier-1 jobs. Linux is on the roadmap (TPM 2.0 + IMA) but not shipped.</li> Android Islands.</strong> Same — Play Integrity API is on the roadmap, not shipped.</li> The consumer’s own machine.</strong> If your laptop has malware, no amount of SE-bound encryption on the inference end helps. We can secure the wire and the compute; you have to secure the keyboard.</li> </ul> We’d rather be conservative about what Tier 1 means than ship a marketing claim that crumbles when an actual security researcher pokes at it.</p> Why we built this on consumer hardware</h2> Most confidential-compute work targets datacenter silicon — Intel SGX, Intel TDX, AMD SEV-SNP. Those are useful, and Tier 2 will use them. But there’s an interesting property of consumer Apple hardware that makes it a good Tier-1 substrate:</p> The Secure Enclave is in every chip Apple ships.</strong> Every M-series Mac and every recent iPhone has one. There are hundreds of millions of these in the wild, sitting idle most of the time.</li> CryptoKit exposes SE-bound P-256 ECDH directly.</strong> No vendor-specific SDK, no platform attestation token brokered through a third party. Just SecureEnclave.P256.KeyAgreement.PrivateKey</code>.</li> Hardened Runtime + code signing give us a binary-identity primitive.</strong> The same machinery Apple uses to make sure App Store apps haven’t been tampered with, we use to make sure the Island binary running on someone’s Mac mini is the one we built.</li> </ul> So the math worked out: a privacy substrate built on hardware most cloud providers don’t have, on devices that already exist, mostly idle, in people’s homes and offices. The same observation that drove “Your Phone Is an Island”</a> — the chip is everywhere, it’s idle, let’s use it — applies to its security features just as much as to its inference throughput.</p> What’s next</h2> The privacy roadmap covers six phases:</p> P0 — Threat Model & Honest Statement.</strong> Done. See the updated SECURITY.md</a>.</li> P1 — Apple Silicon Tier 1.</strong> Done. Live in production today.</li> P2 — Coordinator TEE.</strong> Run the coordinator itself inside Intel TDX or AMD SEV-SNP so the operator can’t read prompts either. This is the bridge from Tier 1 to Tier 2.</li> P3 — End-to-end encryption.</strong> Done at the protocol layer (this post). Awaits P2 to close the trust loop.</li> P4 — Linux Tier 1/2.</strong> TPM 2.0 + IMA-measured boot for Linux Islands; SGX/TDX/SEV-SNP for Linux Tier 2.</li> P5 — Continuous Verification.</strong> Periodic re-attestation with audit logs, transparent reporting of what each Island did and when.</li> </ul> The full document is in our engineering context repo</a>.</p> If you have a workload where contractual privacy isn’t enough — set privacy_tier: 1</code> on your Cargo. The job will only route to a Verified Island within its blessing TTL, the prompt will be encrypted to that Island’s SE-bound key, and the response comes back the same way. The coordinator carries ciphertext; the contributor’s machine sees the work; nobody else sees either.</p> That’s the difference between “we promise we don’t read your prompts” and “the math says we cannot.” For Tier 1, we shipped the math.</p> Europe Without Datacenters 2026-05-01T00:00:00+00:00 It is the first of May. A day about human labour, which makes it a good day to be honest about two things at once: that AI is going to take a meaningful share of the work humans used to have to do, and that the infrastructure deciding whose</em> AI does it is being poured in concrete right now, mostly outside Europe.</p> Archipelag.io is a bet on both of those things. We think AI inference is permanent — not a hype cycle, not a winter waiting to happen, but a new utility on the level of electricity and bandwidth. And we think Europe is currently weak on the layer that delivers that utility, in a way that will define the next twenty years of European innovation if nothing changes.</p> There is a separate, longer argument for that claim — about why “most innovative country” rankings measure yesterday’s patents rather than today’s substrate, and why Switzerland sitting first on a fifteen-year-old composite says less than people think. Our founder, Raffael Schneider, makes that case in full on his personal blog</a>. The short version is the one that matters for what we are building: the leaderboards lag the substrate by a decade, and Europe’s substrate is thin.</strong></p> The infrastructure gap is real</h2> Europe has world-class researchers. It has serious universities. It has, in patches, a manufacturing base that other regions cannot conjure on demand. What it does not have, in 2026, is sovereign training and inference capacity at the scale the next decade will demand. The hyperscale buildout is happening in Virginia, in Texas, in the Gulf states, in Seoul, in Hefei. The European share of frontier compute is small, and most of it sits inside foreign clouds whose access can be revoked, repriced, or repurposed by a jurisdiction that is not ours.</p> The reflex answer is “build more datacenters.” It is not wrong, but it is slow. A new gigawatt-class facility takes years of permits, grid headroom Europe largely does not have spare, and capital that flows more easily to other geographies. By the time the concrete cures, the workload has moved on.</p> There is a faster answer hiding in plain sight.</p> Datacenter-less inference</h2> Every modern phone has a neural accelerator. Every modern laptop has a GPU that idles 95% of the time. Every workstation under a desk in a Basel lab, a Stockholm studio, a Lisbon co-working space is, for most of the day, a compute resource doing nothing.</p> Archipelag.io turns that latent capacity into a network. An iPhone in Zürich runs a 1.1B-parameter model for someone in Lyon. A Mac Studio in Berlin handles a diffusion job submitted from Madrid. The coordinator places work on the nearest capable Island; the host earns credit for the cycles they contribute; the consumer gets inference that is fast, cheap, and — this is the part that matters for Europe — running on hardware whose physical location and legal jurisdiction are knowable.</p> This is not a substitute for hyperscale where hyperscale is genuinely needed. Frontier training will keep happening in big buildings. But a very large fraction of the inference economy — chat, summarization, OCR, transcription, image edits, code assistance, the daily volume that adds up to most of the bill — does not need an H100 cluster. It needs enough</em> compute, near enough</em> to the user, under the right jurisdiction</em>. That is a substrate problem, not a datacenter problem. And it is one Europe can build now, with hardware that already exists, in homes and offices across the continent.</p> Sovereignty as a side effect, not a slogan</h2> “Digital sovereignty” gets used as a marketing word. We try to be precise about what we mean. When inference runs on an Island whose operator is a person in a known country under a known legal regime, you get answers to questions that hyperscale clouds cannot answer cleanly: where the bytes physically were when the model touched them, whose law governs disclosure, who can compel a shutdown, whose export controls bind the silicon. None of that requires a new treaty. It requires the network to be built out of hardware that is already inside the jurisdiction.</p> The corollary is that the network is hard to lock in. There is no single vendor whose terms-of-service govern the substrate. Islands can leave; consumers can leave; cargos are signed and portable. Anti-lock-in is a leading indicator of innovation: the cheaper a system is to exit, the more the ecosystem iterates around it.</p> The labour question</h2> Labour Day is not a technology holiday. It is about people, and about which work humans should and should not have to do. We are clear-eyed about what AI is going to change. A meaningful share of the work that has historically been arduous, repetitive, low-agency, or quietly dangerous is going to be done by machines in the next decade. That is the fourth industrial revolution as it is actually arriving — not chrome humanoids, but inference cycles quietly absorbing the parts of human labour that humans were never especially glad to do.</p> That shift is good if the substrate underneath it is plural, jurisdictionally clear, and within reach of the people whose lives it will reshape. It is bad if it is owned by three companies on another continent.</p> Archipelag.io is our piece of making it the first one. A network where the compute is everywhere, where every contributor is a stakeholder, and where the most important question about an AI workload — whose hardware, under whose law</em> — has an answer that does not require trusting a logo.</p> The leaderboards will catch up in ten years. The substrate is built now.</p> Your Phone Is an Island 2026-04-16T00:00:00+00:00 There’s a chip inside your phone that most software never touches.</p> Apple calls it the Neural Engine. It’s a dedicated matrix-multiplication accelerator — sixteen cores on the A17 Pro, capable of 35 trillion operations per second. It was designed for machine learning: face recognition, photo enhancement, Siri’s on-device processing, that sort of thing. But most of the time, across most of the billion-plus iPhones in the world, it does absolutely nothing.</p> The same is true for the GPU. The M-series chips in MacBooks and iPads have GPUs that rival discrete desktop cards from a few years ago. The A-series chips in iPhones have Metal-capable GPUs that can run real inference workloads. Not toy demos — actual language models, generating coherent text at usable speeds. Apple spent billions designing this silicon. It sits in people’s pockets, idle.</p> We looked at that and thought: what if it didn’t have to be?</p> A prompt walks into a phone</h2> Here’s what we built. It’s easier to show than to explain.</p> Person A opens an app on their iPhone. The app connects to the Archipelag.io coordinator — our control plane — and registers the phone as an Island on the network. The phone reports what it can do: which chip it has, how much memory is available, what model formats it supports. From this moment, the phone is a compute node, ready to accept work.</p> Person B opens a browser on their laptop, navigates to Archipelag.io, and types a message into a chat interface. Something like: “Explain distributed computing in one sentence.”</p> What happens next takes about two seconds:</p> The coordinator receives the prompt and looks for an available Island that can handle it.</li> It finds Person A’s iPhone — online, charged, on WiFi, thermally stable — and dispatches the job.</li> The iPhone receives the job over a WebSocket connection, loads a language model (TinyLlama 1.1B, quantized to 4-bit, about 650 megabytes), and begins inference.</li> Tokens stream back in real time — word by word, through the coordinator, through a WebSocket, into Person B’s browser.</li> Person B sees the response appear progressively, just like any chat interface. Except the model that generated it is running on a phone across town.</li> </ol> No cloud GPU was involved. No data center. No NVIDIA. Just a phone, a browser, and a coordination layer in between.</p> Why this is not a gimmick</h2> The natural reaction is: okay, but a phone is slow compared to an A100. Why would anyone do this?</p> Fair question. Here’s why we think this matters.</p> The numbers are better than you’d expect.</strong> An iPhone 15 Pro (A17 chip) runs TinyLlama at roughly 25 tokens per second. That’s fast enough for conversational use — most people read at about 4-5 words per second, and each token is roughly three-quarters of a word. The model is small (1.1 billion parameters), but it’s coherent, and it’s enough for a wide range of tasks: summarization, Q&A, classification, simple creative writing. Larger phones with more RAM can run bigger models. The M-series iPads and MacBooks can run 7B and 13B parameter models at speeds that rival some cloud endpoints.</p> The hardware already exists at enormous scale.</strong> Apple has shipped over 2.3 billion iPhones since 2007. Even if you only count devices from the last three years — phones with Neural Engines capable of running quantized LLMs — that’s hundreds of millions of devices. Each one is a potential compute node. No one needs to buy new hardware. No one needs to provision a cloud instance. The compute is already distributed across every city, every neighborhood, every pocket.</p> Latency can be better, not worse.</strong> When a GPU is physically close to the person making the request, the network round-trip is measured in single-digit milliseconds instead of the 50-200ms typical of cloud endpoints in a different region. For streaming applications — where tokens arrive one at a time and perceived speed matters as much as throughput — local inference on modest hardware can feel faster than remote inference on expensive hardware.</p> It’s resilient in ways centralized infrastructure isn’t.</strong> A network of ten thousand phones spread across a city doesn’t have a single point of failure. There’s no region to go down, no availability zone to lose, no cloud provider to have an outage. Individual phones go offline — they run out of battery, lose WiFi, get put in a pocket. The coordinator routes around them. The aggregate network stays up.</p> Under the hood</h2> For the engineers in the room, here’s how the pieces fit together.</p> The Island software</h3> The Archipelag.io Island app is a native Swift application built on Apple’s platform stack. At its core is AgentCore</code>, which manages the lifecycle of the phone as a compute node: registration, heartbeating, job acceptance, execution, and result streaming.</p> When the app starts, it generates a persistent host ID (a UUID stored locally), detects the device’s capability class based on the chip generation, and opens a WebSocket connection to the coordinator using the Phoenix Channel protocol. Registration includes the device’s supported runtimes (llmcpp</code> for quantized language models, coreml</code> for Apple’s ML framework, wasm</code> for WebAssembly modules), available memory, and performance estimates.</p> The phone sends a heartbeat every ten seconds: current status, battery level, charging state, thermal condition, active jobs. The coordinator uses this to decide whether the phone is a viable target for new work.</p> Inference</h3> Language model inference runs through LlamaExecutor</code>, a Swift wrapper around llama.cpp</a> via the LLM.swift</a> library. Models are GGUF-format files — quantized (4-bit, Q4_K_M) to fit in mobile memory while preserving output quality.</p> The first time a phone receives a job for a given model, it downloads the GGUF file (about 650MB for TinyLlama), verifies its SHA-256 hash, and caches it locally. Subsequent jobs skip the download entirely. The coordinator can also push preload recommendations — telling the phone to fetch a model in advance based on anticipated demand, so it’s ready before the first job arrives.</p> During inference, tokens are generated sequentially and streamed back to the coordinator in real time. Each token is sent as a discrete message with a sequence number, allowing the coordinator to relay them to the consumer’s browser in order. The consumer sees text appear word by word, with the same progressive feel as any cloud-hosted chat interface.</p> The coordinator</h3> The coordinator is an Elixir/Phoenix application that serves as the control plane for the entire network. When a consumer submits a prompt, the coordinator selects the best available Island using a multi-dimensional scoring system:</p> Speed score</strong>: estimated throughput for this model on this hardware.</li> Fit score</strong>: how well the workload matches the Island’s resources — not too tight (risk of failure), not too loose (wasted capacity).</li> Headroom score</strong>: remaining capacity for concurrent work.</li> </ul> For mobile Islands running LLM inference, speed is weighted heaviest (50%), because on-device inference time dominates the user experience. The coordinator also prefers “warm” Islands — those that have recently run the same model and still have it loaded in memory — to avoid the latency of cold-starting a model load.</p> Once a host is selected, the coordinator dispatches the job over the WebSocket channel, including the model URL, expected hash, context window size, and temperature setting. The Island software picks it up, runs inference, and streams tokens back through the same channel.</p> Pairing</h3> How does a phone become an Island? Through a pairing flow designed to be as simple as scanning a QR code.</p> A user logs into Archipelag.io on their laptop and navigates to the pairing page. The system generates a six-character code (alphanumeric, no ambiguous characters like I/O/1/0) that’s valid for ten minutes. The user opens the Archipelag.io Island app on their phone, enters the code (or scans the QR), and the phone connects to the coordinator with that code as a credential. The coordinator verifies it, links the phone to the user’s account, issues a 30-day auth token, and the phone is now a registered Island.</p> From that point forward, the phone reconnects automatically whenever it’s on WiFi and its preconditions are met.</p> Preconditions</h3> A phone is not a server. It overheats, runs out of battery, loses connectivity, and gets shoved into a back pocket mid-inference. The Island software accounts for all of this.</p> Before accepting any job, the phone checks: Is it connected to WiFi? Is the battery above 20%? Is it charging (waived in demo mode)? Is the thermal state nominal or fair (not throttling)? Is it within the owner’s configured availability schedule?</p> If any of these conditions fail, the phone reports itself as unavailable and the coordinator routes around it. If conditions change mid-job (the phone overheats, WiFi drops), the coordinator’s lease system detects the stall — leases renew every 30 seconds, and if a renewal is missed, the job is requeued to another Island.</p> Beyond phones</h2> The mobile Island is a proof of concept for a broader thesis: compute should come from everywhere, not just data centers.</p> Today, Archipelag.io routes AI inference jobs to desktop GPUs, Mac Studios, Linux workstations, and now phones. The architecture doesn’t privilege any particular hardware. The same coordinator, the same job dispatch protocol, the same streaming pipeline works whether the Island is an RTX 4090 in a gaming rig or an A17 chip in an iPhone. The scoring system adapts — it knows a phone is slower than a desktop GPU and routes accordingly, sending lightweight jobs to phones and heavy inference to big GPUs.</p> But more importantly, the architecture isn’t limited to AI inference at all. The coordinator dispatches jobs to containers. Those containers can run anything: OCR, image processing, document conversion, video transcoding, code formatting, data transformation. The Cargo Registry</a> already lists dozens of non-AI workloads. A phone can’t run all of them — it doesn’t have Docker — but it can run WASM modules and native inference, and as Apple’s hardware continues to improve, the range of feasible workloads grows with every chip generation.</p> This is what makes Archipelag.io fundamentally different from inference-only networks that are locked to a single hardware vendor. We’re not building a relay for LLM API calls. We’re building a compute fabric — one where any device, running any supported runtime, can contribute capacity to the network and earn credits for the work it does.</p> The phone in your pocket is not just a client. It’s a node. It’s an Island.</p> The economics</h2> During our open beta</a>, all credits are virtual — nobody pays, nobody gets paid. But the economic signals are real.</p> An Island earns credits for every job it completes. The Compute Exchange</a> sets prices through supply and demand: Islands post asking prices, consumers place bids, and the market clears. A phone that reliably completes lightweight inference jobs earns credits at a rate proportional to the work it does. Over time, Islands build up karma — a reputation score based on completion rate, speed, and reliability — which gives them access to higher-value jobs.</p> The marginal cost of running inference on a phone is effectively zero. The device is already purchased, already charged (usually), already on WiFi. The electricity consumed during inference is negligible — pennies per day. There’s no rack to rent, no cooling bill, no bandwidth fee. This means even modest earnings represent near-pure margin for the phone’s owner.</p> At scale, the aggregate capacity is striking. If one percent of iPhones sold in the last three years — roughly two million devices — contributed one hour of compute per day, the network’s total throughput would rival a mid-sized cloud region. Not in raw FLOPS on paper, but in actual inference requests served, close to the people who need them.</p> What comes next</h2> The mobile Island is live in our beta today for iOS. Android support (using ONNX Runtime for on-device inference) is in active development. We’re working on expanding the set of models that run efficiently on mobile hardware — Phi-3 Mini (3.8B parameters) runs well on recent iPhones with 8GB+ of RAM, and the M-series iPads and Macs can handle much larger models.</p> We’re also exploring distributed inference — splitting a single large model across multiple nearby devices, so a cluster of phones could collectively run a model that none of them could run alone. This is early-stage research, not a shipped feature, but the architecture was designed with it in mind.</p> If you want to turn your phone into an Island: become an Island</a>, pair your device with your account, and start contributing compute. If you want to use AI served from the network: sign up</a> and open the chat. Your prompt might be answered by a GPU in someone’s basement, or by a phone on someone’s desk. Either way, it’ll be closer to you than any data center.</p> The hardware is already everywhere. We just connected it.</p> Archipelag.io Enters Open Beta 2026-03-13T00:00:00+00:00 There’s a GPU in your computer that spends most of its life doing nothing.</p> Maybe it’s a gaming rig that sits idle while you’re at work, a machine that cost you the better part of a paycheck and now runs Steam updates and screensavers for twenty hours a day. Maybe it’s a workstation that renders video for two hours and sleeps the other twenty-two. Maybe it’s a machine you built during the mining craze that’s been collecting dust since ETH went proof-of-stake. Maybe it’s the M3 MacBook on your kitchen table. Or the iPhone in your pocket, with a neural engine that Apple spent billions designing, currently dedicated to scrolling Instagram.</p> All of these devices can do meaningful compute work. Not just the beefy rigs, even a phone can run a small language model, handle a classification task, or contribute to a distributed inference job. The compute is already there, paid for, powered on, doing nothing useful.</p> Meanwhile, someone across town is paying a cloud provider to rent a GPU in a data center three time zones away, waiting 200ms for each token of a chatbot response, and wondering why AI feels so sluggish.</p> We looked at this situation and thought: this is fixable.</p> The compute is already here</h2> The way AI infrastructure works today is that a handful of companies own the GPUs, and everyone else rents time on them. Prices go up, capacity gets scarce, and if you’re not a Fortune 500 company with a cloud contract, you’re at the back of the queue. Meanwhile, NVIDIA can’t manufacture chips fast enough, export controls are tightening, and every quarter the gap between who needs compute and who can get it grows wider.</p> But here’s the thing that gets lost in all the hand-wringing about chip shortages: there is an enormous amount of compute sitting in people’s homes right now, doing absolutely nothing. The gaming GPU market alone ships over 30 million discrete cards a year. Add laptops with capable integrated graphics, workstations, retired mining hardware, phones with neural engines, tablets with ML accelerators. The aggregate compute power in ordinary people’s hands dwarfs what any single cloud provider operates. It’s just fragmented, unconnected, and idle.</p> We thought there might be a different way. What if the people who need compute could get it from the people who already have it? What if your neighbor’s RTX 4070 could serve your LLM requests with lower latency than any data center, because it’s literally down the street? What if a network of regular people’s machines — from a teenager’s gaming PC in Munich to a photographer’s Mac Studio in Lisbon — could collectively form something more resilient and more accessible than any centralized cloud?</p> That’s Archipelag.io. A distributed compute network where you contribute what you have and use what you need. The hardware exists. The bandwidth exists. The people exist. We built the coordination layer to connect it all.</p> Why this matters beyond convenience</h2> We’re building this because we think the current trajectory of AI infrastructure is fragile and exclusionary, and we’d rather not wait to find out how that plays out.</p> When three companies control the majority of GPU cloud capacity, that’s a single point of failure for an increasingly critical technology. When a startup in São Paulo pays the same rate as a hedge fund in Manhattan but gets worse latency because the nearest data center is in Virginia, that’s not a market working well. When a solo developer can’t afford to experiment with AI because inference costs eat their entire budget, that’s talent and ideas we never get to see.</p> We believe AI compute should work more like the internet itself: decentralized, regional, and owned by the people who participate in it. Not because centralization is evil, but because resilient systems need diverse infrastructure. A network of ten thousand independent Islands spread across real neighborhoods is harder to take down, harder to price-gouge, and closer to the people who actually use it than any single data center campus.</p> This is partially idealistic, yes. But it’s also practical. Hardware scarcity is not a hypothetical scenario for next decade, it’s already here. Lead times for high-end GPUs stretch months. Cloud spot instance prices spike unpredictably. Every new foundation model release triggers another wave of demand that supply can’t absorb. The people who already own capable hardware are sitting on an asset that could generate passive income as an Island while making AI more accessible for everyone around them. That’s not utopian thinking, that’s just a better allocation of resources that already exist.</p> So, open beta</h2> Starting today, Archipelag.io is open to everyone for three months, through June 13, 2026</strong>.</p> Here’s what that means in practice:</p> All the money is fake.</strong> Credits, earnings, payouts, everything financial on the platform during the beta is virtual. Nobody gets charged, nobody gets paid. We need to test the billing system, the karma scores, the payout logic, and the best way to do that is to run it all for real, just without real money attached. When we flip the switch to real billing after the beta, everyone who participated will get bonus credits as a thank-you.</p> Things will break.</strong> We’ve been building this for over a year, and it works on our machines, in our test environments, with our carefully crafted scenarios. But we haven’t seen what happens when a few hundred people with different hardware, different network conditions, and different ideas about what “use AI” means all show up at once. That’s the whole point. We need you to find the bugs we can’t.</p> We might reset things.</strong> If we need to change how the database works or restructure accounts, we might have to wipe data. We’ll warn you first, but don’t store anything important here yet.</p> What you can actually do</h2> Use AI.</strong> Chat with LLMs (Mistral 7B, Llama), generate images with Stable Diffusion XL, or hit our OpenAI-compatible API with your existing code. There are Python and JavaScript SDKs if you want to build on top of it. The web UI has a playground for trying models and a registry for browsing what’s available.</p> Run an Island with a GPU.</strong> If you’ve got a decent graphics card (RTX 3060 or better), install the Island software on Windows, macOS, or Linux. Your machine picks up inference jobs from people nearby, runs them, streams the results back. You set your own availability, your machine works when you’re not using it, and you earn credits. Think of it as your gaming rig picking up a side job while you sleep.</p> Run an Island on your phone.</strong> If you have a recent iPhone or an M-series Mac, you can run smaller models directly on-device. It won’t replace a 4090 for heavy inference, but it can handle lightweight tasks, and it lets you see how the network works without dedicating a full machine. We’re building this to scale from phones to server racks, because a truly distributed network needs every tier of hardware.</p> The compute exchange sets prices through supply and demand, so even modest hardware can find work at a price point that makes sense. During the beta, all credits are virtual, but the economics are real: you’ll see what your Island’s hardware is worth on the network.</p> What we’re hoping to learn</h2> Honestly, we don’t know exactly what will happen, and that’s why we’re doing this.</p> We want to find out if the routing works well enough, if requests actually land on nearby Islands or if they bounce around. We want to see where the onboarding loses people, which error messages are confusing, which features nobody uses. We want to know if the karma system (where Islands build reputation through reliable service) actually incentivizes good behavior or just creates busywork.</p> We want to hear from you. If something is broken, open an issue</a>. If something is confusing, email us</a>. If you have an idea for something we should build, tell us.</p> Getting started</h2> Want to use AI?</strong> Sign up</a>, open the chat, and start talking to a model. It’s running on someone’s Island, probably not very far from where you’re sitting.</p> Want to contribute compute?</strong> Grab the Island software</a>, point it at your API key, and set when you’re available. Jobs will start arriving.</p> Want to build?</strong> The docs</a> have everything, including the API reference</a> and SDK guides.</p> What happens next</h2> After the beta, we plan to turn on real payments, add more models, and grow the network into more regions. The specifics depend on what we learn over the next three months.</p> The longer arc is this: we want to build a compute network that’s as decentralized and resilient as the internet was supposed to be. One where a freelance developer in Kraków can spin up an AI feature without a cloud contract. Where a student in Lagos can access the same models as a researcher at Stanford, served from hardware in their own city. Where the person running an Island earns enough to cover their electricity bill and then some, turning idle hardware into a genuine income stream.</p> We’re not pretending this is easy. Distributed systems are hard, incentive design is hard, getting strangers to trust a network with their hardware and their Cargos is hard. But the alternative, a future where all AI compute funnels through three or four corporate clouds, is not a future we want to build toward.</p> We’ll keep posting updates here and on GitHub</a>. This is a small team building something that only works if people show up and use it. So thank you for showing up.</p> Let’s see what we can build together.</p> Privacy Policy 2026-01-26T00:00:00+00:00 Last updated: January 26, 2026</em></p> This Privacy Policy describes how Archipelag.io (“we”, “us”, or “our”) collects, uses, and shares information about you when you use our website, platform, and services (collectively, the “Services”).</p> Information We Collect</h2> Information You Provide</h3> Account Information</strong>: When you create an account, we collect your email address, username, and password. If you sign up using GitHub, we receive your GitHub username and email.</li> Payment Information</strong>: When you purchase credits, our payment processor (Stripe) collects your payment card details. We do not store full card numbers on our servers.</li> Island Information</strong>: If you register as an Island (compute contributor), we collect information about your hardware specifications, IP address, and geographic region.</li> Communications</strong>: When you contact us, we collect the information you provide in your messages.</li> </ul> Information Collected Automatically</h3> Usage Data</strong>: We collect information about how you use our Services, including API calls, Cargo types, and timestamps.</li> Device Information</strong>: We collect information about the device you use to access our Services, including browser type, operating system, and device identifiers.</li> Log Data</strong>: Our servers automatically record information including your IP address, access times, and pages viewed.</li> Cookies</strong>: We use essential cookies to maintain your session and preferences. We do not use third-party tracking cookies.</li> </ul> Information from Islands</h3> When Cargos run on Island machines:</p> Islands receive only the minimum data necessary to execute Cargos</li> All data is encrypted in transit using TLS 1.3</li> Cargos run in isolated containers with no persistent storage</li> Islands cannot access or retain consumer data after Cargo completion</li> </ul> How We Use Your Information</h2> We use the information we collect to:</p> Provide Services</strong>: Process your requests, route Cargos, and deliver results</li> Manage Accounts</strong>: Create and manage your account, process payments, and handle payouts</li> Improve Services</strong>: Analyze usage patterns to improve performance and develop new features</li> Ensure Security</strong>: Detect and prevent fraud, abuse, and security threats</li> Communicate</strong>: Send service announcements, security alerts, and support messages</li> Comply with Law</strong>: Meet legal obligations and respond to lawful requests</li> </ul> How We Share Your Information</h2> We do not sell your personal information. We share information only in these circumstances:</p> With Service Providers</h3> Stripe</strong>: Payment processing</li> Cloud infrastructure</strong>: Hosting and data storage (EU-based)</li> Email services</strong>: Transactional email delivery</li> </ul> With Islands</h3> Islands receive anonymized Cargo data necessary for execution</li> Islands do not receive your identity, account details, or payment information</li> Geographic routing data is approximate (city-level) only</li> </ul> For Legal Reasons</h3> We may disclose information if required by law, court order, or government request, or to protect rights, safety, or property.</p> Business Transfers</h3> If we are involved in a merger, acquisition, or sale of assets, your information may be transferred as part of that transaction.</p> Data Security</h2> We implement appropriate technical and organizational measures to protect your information:</p> Encryption</strong>: All data encrypted in transit (TLS 1.3) and at rest (AES-256)</li> Access Controls</strong>: Strict access controls and authentication for internal systems</li> Monitoring</strong>: Continuous security monitoring and intrusion detection</li> Audits</strong>: Regular security assessments and penetration testing</li> Incident Response</strong>: Documented procedures for security incident handling</li> </ul> Data Retention</h2> Account Data</strong>: Retained while your account is active, then deleted within 30 days of account closure</li> Usage Logs</strong>: Retained for 90 days for operational purposes</li> Payment Records</strong>: Retained for 7 years as required by financial regulations</li> Support Communications</strong>: Retained for 2 years after resolution</li> </ul> Your Rights</h2> Depending on your location, you may have the following rights:</p> Access</strong>: Request a copy of the personal information we hold about you</li> Correction</strong>: Request correction of inaccurate information</li> Deletion</strong>: Request deletion of your personal information</li> Portability</strong>: Request your data in a machine-readable format</li> Objection</strong>: Object to certain processing of your information</li> Restriction</strong>: Request restriction of processing in certain circumstances</li> </ul> To exercise these rights, contact us at hey@archipelag.io</a>.</p> International Data Transfers</h2> Our primary infrastructure is located in the European Union. If you access our Services from outside the EU, your information may be transferred to, stored, and processed in the EU. We ensure appropriate safeguards are in place for any international transfers.</p> Children’s Privacy</h2> Our Services are not directed to children under 16. We do not knowingly collect personal information from children. If you believe we have collected information from a child, please contact us immediately.</p> Changes to This Policy</h2> We may update this Privacy Policy from time to time. We will notify you of material changes by posting the new policy on this page and updating the “Last updated” date. Your continued use of the Services after changes constitutes acceptance of the updated policy.</p> Contact Us</h2> If you have questions about this Privacy Policy or our privacy practices, contact us at:</p> Email</strong>: hey@archipelag.io</a></li> Address</strong>: Archipelag.io, Amsterdam, Netherlands</li> </ul> For data protection inquiries in the EU, you may also contact your local data protection authority.</p> This policy is provided for informational purposes. For the legally binding terms, please refer to the Terms of Service</a>.</em></p> Security 2026-01-26T00:00:00+00:00 Security is fundamental to Archipelag.io. We’ve designed our platform with security at every layer, ensuring both consumers and Islands can participate safely in our distributed compute network.</p> Security Architecture</h2> Zero-Trust Model</h3> We operate on a zero-trust security model where:</p> Consumers don’t trust Islands</strong>: All consumer data is encrypted and Islands never see plaintext content</li> Islands don’t trust Cargos</strong>: All Cargos run in isolated containers with strict resource limits</li> The coordinator is the authority</strong>: Cryptographically signed Cargos and authenticated communications</li> </ul> Encryption</h3> In Transit</strong>: All communications use TLS 1.3 with modern cipher suites</li> At Rest</strong>: Sensitive data encrypted using AES-256-GCM</li> Cargo Data</strong>: End-to-end encryption for sensitive Cargo payloads</li> API Keys</strong>: Hashed using Argon2id before storage</li> </ul> Workload Isolation</h3> Every Cargo on our network runs in complete isolation:</p> Container Sandboxing</strong>: Docker containers with seccomp profiles and AppArmor</li> Resource Limits</strong>: Strict CPU, memory, and network quotas</li> No Persistent Storage</strong>: Cargos cannot write to Island filesystems</li> Network Restrictions</strong>: Outbound-only connections, no Island network access</li> Signed Images</strong>: Only cryptographically signed container images can execute</li> </ul> Island Security</h2> Agent Security</h3> The Archipelag.io Island software is designed with security as a priority:</p> Written in Rust</strong>: Memory-safe language eliminates entire classes of vulnerabilities</li> Minimal Privileges</strong>: Runs with least-privilege access</li> Automatic Updates</strong>: Security patches delivered automatically</li> Open Source</strong>: Agent code is available for security review</li> </ul> Network Security</h3> Outbound-Only</strong>: Islands never accept inbound connections</li> WireGuard VPN</strong>: Encrypted tunnels for all coordinator communication</li> No Port Forwarding</strong>: No router configuration or firewall changes required</li> IP Anonymization</strong>: Consumer IPs are not shared with Islands</li> </ul> Consumer Security</h2> Account Protection</h3> Strong Password Requirements</strong>: Minimum 12 characters with complexity requirements</li> OAuth Integration</strong>: Secure authentication via GitHub</li> Session Management</strong>: Automatic session expiration and secure token handling</li> API Key Scoping</strong>: Fine-grained permissions for API keys</li> </ul> Data Protection</h3> Regional Processing</strong>: Data processed in your geographic region by default</li> No Data Retention</strong>: Cargo inputs and outputs not stored after completion</li> Audit Logging</strong>: Complete audit trail of account and API activity</li> Right to Deletion</strong>: Request complete data deletion at any time</li> </ul> Audit Logging</h3> Hash-Chained Logs</strong>: Immutable audit trail with tamper detection via hash chaining</li> Trust Events</strong>: All approvals, rejections, and suspensions are logged with full context</li> Security Incidents</strong>: Automated incident recording with forensic detail</li> Chain Verification</strong>: Integrity of the audit log can be verified at any time</li> </ul> Rate Limiting & Abuse Prevention</h3> API Rate Limiting</strong>: 100 requests per minute per API key</li> Authentication Rate Limiting</strong>: 5 requests per minute for magic link login</li> Input Validation</strong>: Message content capped at 32KB, job input at 256KB, max 128 messages per request</li> Body Size Limits</strong>: 1MB maximum request body</li> </ul> Infrastructure Security</h2> Platform Security</h3> EU Data Centers</strong>: Primary infrastructure hosted in EU data centers</li> DDoS Protection</strong>: Multi-layer DDoS mitigation</li> Web Application Firewall</strong>: Protection against common web attacks</li> Regular Penetration Testing</strong>: Third-party security assessments</li> </ul> Monitoring & Response</h3> 24/7 Monitoring</strong>: Automated security monitoring and alerting</li> Incident Response</strong>: Documented procedures for security incidents</li> Vulnerability Management</strong>: Regular scanning and patching</li> Security Logging</strong>: Comprehensive logging for forensic analysis</li> </ul> Compliance</h2> We maintain security practices aligned with:</p> GDPR</strong>: EU data protection requirements</li> SOC 2 Type II</strong>: (In progress) Security, availability, and confidentiality controls</li> ISO 27001</strong>: (Planned) Information security management</li> </ul> Vulnerability Disclosure</h2> We welcome responsible disclosure of security vulnerabilities.</p> Reporting a Vulnerability</h3> If you discover a security issue, please report it to:</p> Email</strong>: hey@archipelag.io</a></p> PGP Key</strong>: A security.txt</code> file with our PGP key is planned and will be available at /.well-known/security.txt</code></p> What to Include</h3> Description of the vulnerability</li> Steps to reproduce</li> Potential impact</li> Your contact information (optional, for follow-up)</li> </ul> Our Commitment</h3> Acknowledgment</strong>: We’ll acknowledge receipt within 24 hours</li> Assessment</strong>: Initial assessment within 72 hours</li> Resolution</strong>: Critical issues addressed within 7 days</li> Credit</strong>: Public acknowledgment for responsible disclosures (if desired)</li> No Legal Action</strong>: We will not pursue legal action against good-faith researchers</li> </ul> Scope</h3> In-scope systems include:</p> archipelag.io and all subdomains</li> api.archipelag.io</li> The Island software</li> Mobile applications</li> </ul> Out of scope:</p> Third-party services (Stripe, GitHub)</li> Social engineering attacks</li> Physical attacks</li> Denial of service testing</li> </ul> Security Best Practices</h2> For Users</h3> Use Strong Passwords</strong>: Or sign in with GitHub for OAuth security</li> Protect API Keys</strong>: Never commit keys to version control</li> Monitor Usage</strong>: Review your account activity regularly</li> Enable Notifications</strong>: Get alerts for account changes</li> Report Suspicious Activity</strong>: Contact us if you notice anything unusual</li> </ol> For Islands</h3> Keep Systems Updated</strong>: Enable automatic updates for the Island software</li> Secure Your Network</strong>: Use a firewall and secure your home network</li> Monitor Resources</strong>: Watch for unusual CPU or network activity</li> Dedicated Hardware</strong>: Consider using dedicated hardware for contributing compute</li> Review Logs</strong>: Periodically review Island logs for anomalies</li> </ol> Security Updates</h2> We publish security advisories for significant issues at:</p> Status Page</strong>: status.archipelag.io (coming soon)</li> Email Notifications</strong>: Security alerts sent to registered users</li> GitHub</strong>: Security advisories in our public repositories</li> </ul> Contact</h2> For security questions or concerns:</p> Security Team</strong>: hey@archipelag.io</a></li> General Support</strong>: hey@archipelag.io</a></li> </ul> Security is an ongoing process. We continuously improve our security posture and welcome feedback from our community.</em></p> Terms of Service 2026-01-26T00:00:00+00:00 Last updated: January 26, 2026</em></p> These Terms of Service (“Terms”) govern your access to and use of Archipelag.io’s website, platform, APIs, and services (collectively, the “Services”). By accessing or using our Services, you agree to be bound by these Terms.</p> 1. Acceptance of Terms</h2> By creating an account or using our Services, you agree to these Terms and our Privacy Policy</a>. If you do not agree, do not use our Services. If you are using the Services on behalf of an organization, you represent that you have authority to bind that organization to these Terms.</p> 2. Description of Services</h2> Archipelag.io provides a distributed compute platform that connects consumers who need AI inference capabilities with Islands who contribute computing resources. Our Services include:</p> For Consumers</strong>: Access to AI models (LLM chat, image generation) via web interface and API</li> For Islands</strong>: Software to contribute compute resources and earn credits</li> Credit System</strong>: Digital credits used to pay for compute services</li> </ul> 3. Account Registration</h2> 3.1 Account Creation</h3> To use most features, you must create an account. You agree to:</p> Provide accurate and complete information</li> Maintain the security of your account credentials</li> Promptly update any changes to your information</li> Accept responsibility for all activities under your account</li> </ul> 3.2 Account Security</h3> You are responsible for maintaining the confidentiality of your account credentials. Notify us immediately at hey@archipelag.io</a> if you suspect unauthorized access.</p> 3.3 Account Termination</h3> We may suspend or terminate your account if you violate these Terms, engage in fraudulent activity, or for any other reason at our discretion with reasonable notice.</p> 4. User Conduct</h2> 4.1 Acceptable Use</h3> You agree to use the Services only for lawful purposes. You will not:</p> Violate any applicable laws or regulations</li> Infringe intellectual property rights of others</li> Submit malicious code, viruses, or harmful content</li> Attempt to gain unauthorized access to systems or data</li> Interfere with or disrupt the Services or servers</li> Use the Services to generate illegal, harmful, or abusive content</li> Circumvent usage limits or access controls</li> Resell or redistribute the Services without authorization</li> </ul> 4.2 Content Guidelines</h3> When using AI models through our Services, you agree not to generate content that:</p> Is illegal in your jurisdiction</li> Depicts child sexual abuse material (CSAM)</li> Promotes violence, terrorism, or self-harm</li> Constitutes harassment, hate speech, or discrimination</li> Infringes on others’ privacy or intellectual property</li> Is deceptive, fraudulent, or misleading</li> </ul> 4.3 Enforcement</h3> We reserve the right to review content and suspend accounts that violate these guidelines. Serious violations may be reported to law enforcement.</p> 5. Island Terms</h2> 5.1 Island Requirements</h3> To participate as an Island (compute contributor), you must:</p> Have legal authority to operate computing equipment at your location</li> Ensure your hardware meets minimum specifications</li> Maintain adequate internet connectivity</li> Comply with local laws regarding electricity usage and business operations</li> </ul> 5.2 Cargo Execution</h3> As an Island, you agree to:</p> Run only approved, signed Cargos from our platform</li> Not inspect, copy, or retain consumer data processed through Cargos</li> Maintain the security and isolation of Cargo containers</li> Report any security incidents immediately</li> </ul> 5.3 Earnings and Payouts</h3> Earnings are calculated based on compute time and hardware tier</li> Minimum payout threshold is $25 USD</li> Payouts are processed monthly via Stripe</li> You are responsible for any applicable taxes on earnings</li> </ul> 6. Credits and Payments</h2> 6.1 Credit Purchases</h3> Credits are purchased in packages as displayed on our pricing page</li> All purchases are final and non-refundable except as required by law</li> Unused credits do not expire</li> Credits have no cash value and cannot be transferred</li> </ul> 6.2 Credit Usage</h3> Credits are deducted based on actual compute usage</li> Usage rates vary by Cargo type and are displayed before submission</li> We reserve the right to adjust pricing with 30 days notice</li> </ul> 6.3 Refunds</h3> Refunds for unused credits may be requested within 30 days of purchase</li> Partially used credit packages are not eligible for refunds</li> Contact hey@archipelag.io</a> for refund requests</li> </ul> 7. Intellectual Property</h2> 7.1 Our Property</h3> The Services, including software, APIs, documentation, and branding, are owned by Archipelag.io and protected by intellectual property laws. You receive a limited, non-exclusive license to use the Services in accordance with these Terms.</p> 7.2 Your Content</h3> You retain ownership of content you submit to the Services. By submitting content, you grant us a license to process that content as necessary to provide the Services.</p> 7.3 Output Content</h3> Content generated by AI models through our Services is provided as-is. You are responsible for ensuring your use of generated content complies with applicable laws.</p> 8. Privacy and Data</h2> Your use of the Services is subject to our Privacy Policy</a>, which describes how we collect, use, and protect your information.</p> 9. Disclaimers</h2> 9.1 Service Availability</h3> The Services are provided “as is” and “as available.” We do not guarantee uninterrupted or error-free operation. We may modify, suspend, or discontinue features at any time.</p> 9.2 AI Output</h3> AI-generated content may contain errors, biases, or inappropriate material. You are responsible for reviewing and validating any output before use. We do not guarantee the accuracy, completeness, or suitability of AI-generated content.</p> 9.3 Third-Party Services</h3> The Services may integrate with third-party services (Stripe, GitHub, etc.). Your use of those services is subject to their respective terms and privacy policies.</p> 10. Limitation of Liability</h2> TO THE MAXIMUM EXTENT PERMITTED BY LAW:</p> We are not liable for any indirect, incidental, special, consequential, or punitive damages</li> Our total liability for any claims arising from these Terms or the Services shall not exceed the amount you paid us in the 12 months preceding the claim</li> We are not liable for any loss of data, profits, or business opportunities</li> </ul> These limitations apply regardless of the theory of liability (contract, tort, strict liability, or otherwise).</p> 11. Indemnification</h2> You agree to indemnify and hold harmless Archipelag.io, its officers, directors, employees, and agents from any claims, damages, losses, or expenses (including reasonable attorneys’ fees) arising from:</p> Your use of the Services</li> Your violation of these Terms</li> Your violation of any third-party rights</li> Content you submit through the Services</li> </ul> 12. Dispute Resolution</h2> 12.1 Informal Resolution</h3> Before filing any formal dispute, you agree to contact us at hey@archipelag.io</a> and attempt to resolve the dispute informally for at least 30 days.</p> 12.2 Governing Law</h3> These Terms are governed by the laws of the Netherlands, without regard to conflict of law principles.</p> 12.3 Jurisdiction</h3> Any disputes not resolved informally shall be resolved in the courts of Amsterdam, the Netherlands.</p> 13. Changes to Terms</h2> We may modify these Terms at any time. Material changes will be notified via email or prominent notice on our website at least 30 days before taking effect. Your continued use of the Services after changes constitutes acceptance of the modified Terms.</p> 14. General Provisions</h2> 14.1 Entire Agreement</h3> These Terms, together with the Privacy Policy, constitute the entire agreement between you and Archipelag.io regarding the Services.</p> 14.2 Severability</h3> If any provision of these Terms is found unenforceable, the remaining provisions will continue in effect.</p> 14.3 Waiver</h3> Our failure to enforce any right or provision does not constitute a waiver of that right or provision.</p> 14.4 Assignment</h3> You may not assign your rights under these Terms without our consent. We may assign our rights at any time.</p> 14.5 Force Majeure</h3> We are not liable for delays or failures caused by circumstances beyond our reasonable control.</p> 15. Contact</h2> For questions about these Terms, contact us at:</p> Email</strong>: hey@archipelag.io</a></li> General inquiries</strong>: hey@archipelag.io</a></li> </ul> By using Archipelag.io, you acknowledge that you have read, understood, and agree to be bound by these Terms of Service.</em></p> Introducing Archipelag.io 2025-01-26T00:00:00+00:00 This started with a question that kept nagging at us: why are we sending AI requests halfway across a continent when there’s a perfectly good GPU two blocks away?</p> Not a hypothetical GPU. A real one. Sitting inside a gaming PC in someone’s apartment, drawing 15 watts at idle, waiting for its owner to get home from work so it can render some frames in a first-person shooter. Or inside a video editor’s workstation that finished a render at 2pm and won’t touch the GPU again until tomorrow morning. Or inside one of the thousands of machines people built during the Ethereum mining era, now gathering dust in closets and garages because proof-of-stake made them obsolete overnight.</p> We kept coming back to this picture: enormous amounts of capable hardware, already purchased, already powered on, already connected to fast internet, doing nothing. And on the other side, people paying cloud providers to rent equivalent hardware in a data center that might be three time zones away, accepting the latency and the markup because that’s how things work.</p> We decided to see if we could connect those two sides.</p> What Archipelag.io actually is</h2> The short version: you send us an AI Cargo (a chat message, an image prompt, an API call), and we route it to a nearby Island that can handle it. That machine runs the job, streams the result back to you, and the person who owns it earns money for the work.</p> The longer version involves a coordinator service that tracks which Islands are online, what hardware they have, how reliable they’ve been, and where they are relative to you. When a job comes in, the coordinator picks the best available Island based on all of that, dispatches the job over NATS (a messaging system built for exactly this kind of thing), and the Island software picks it up, runs it in a sandboxed container, and streams output back through the coordinator to your browser or API client.</p> Islands build up a reputation score we call karma. Complete jobs reliably, respond quickly, stay online when you say you will, and your karma goes up. Higher karma means you get offered better jobs and can charge higher rates. Flake out or produce bad results, and the system routes around you. It’s a simple mechanism, but it aligns incentives in a way that matters when you’re trusting strangers’ hardware with real Cargos.</p> What you can do with it today</h2> We’re launching with AI inference, because that’s where the mismatch between supply and demand is most obvious.</p> You can chat with Mistral 7B and Llama models, with responses streaming token by token from an Island near you. You can generate images with Stable Diffusion XL. If you’re a developer, there’s an OpenAI-compatible API, so if your code already talks to OpenAI, pointing it at Archipelag.io is a configuration change, not a rewrite.</p> On the Island side, we have the Island software that runs on Windows, macOS, and Linux. Install it, point it at your API key, tell it when you’re available, and it handles the rest: pulling the right container images, managing GPU memory, streaming results, reporting health back to the coordinator. You don’t need to understand how inference works. You just need a decent GPU (RTX 3060 or better) and an internet connection.</p> We chose to start with inference because it’s a contained problem with clear value on both sides. But the architecture underneath doesn’t know or care that it’s running language models. It dispatches jobs to containers. Those containers could do rendering, transcoding, scientific simulation, whatever. We’ll get there.</p> Why we think this matters</h2> There’s a practical argument and a structural one.</p> The practical argument: latency. When someone in Berlin asks a chatbot a question and the response comes from a GPU in Frankfurt instead of Virginia, the difference is noticeable. Not just in raw milliseconds, but in the feel of the interaction. Streaming tokens from 30ms away feels like a conversation. Streaming them from 200ms away feels like waiting.</p> The structural argument: concentration. Right now, a small number of companies control most of the world’s GPU cloud capacity. That’s fine when things are working, but it means pricing is opaque, capacity crunches hit everyone at once, and if you’re not a large enterprise with a negotiated contract, you’re paying retail for a commodity. Meanwhile, NVIDIA ships over 30 million discrete GPUs a year into the consumer market alone. The compute exists. It’s just not connected to the people who need it.</p> We’re not trying to replace cloud providers. Data centers are good at sustained, high-throughput workloads, and that’s not going away. What we’re building is the other layer: the local, distributed, community-operated compute that handles the long tail of requests where proximity matters more than raw scale. A freelancer in Lisbon shouldn’t have to send their inference requests to us-east-1. A student in Nairobi shouldn’t be priced out of experimenting with AI because the nearest cloud region is in South Africa.</p> What comes next</h2> Over the coming months, we’re adding more models, expanding into more regions, and shipping SDKs for Python and JavaScript so developers can build on top of this without thinking about the infrastructure underneath. We’re also working on a mobile Island, because modern phones have serious compute capabilities and there’s no reason they shouldn’t participate in the network too.</p> The harder work is in the economics and the trust model. How do you price compute fairly when every Island has different hardware, different electricity costs, different availability patterns? How do you build enough trust that people are willing to send real Cargos to strangers’ machines? We have initial answers to both of these, but they’re hypotheses, not conclusions. We need real usage to test them.</p> Try it</h2> If you want to use AI on the network: sign up</a>, get 10 free credits, and send a message. It’ll be served from someone’s GPU, probably closer to you than you’d expect.</p> If you want to contribute compute: grab the Island software</a>, set your schedule, and see what happens. Your machine works while you don’t.</p> If you want to follow along or contribute: everything is on GitHub</a>. If something breaks or confuses you, tell us</a>.</p> This is a small team at the start of something that only works if people show up and participate. We’re grateful you’re here.</p>

Archipelag.io

Eight Billion Parameters in One Gigabyte

The Math Says We Cannot Read Your Prompt

Europe Without Datacenters

Your Phone Is an Island

Archipelag.io Enters Open Beta

Privacy Policy

How We Share Your Information</h2> We do not sell your personal information. We share information only in these circumstances:</p>

For Legal Reasons</h3> We may disclose information if required by law, court order, or government request, or to protect rights, safety, or property.</p>

Business Transfers</h3> If we are involved in a merger, acquisition, or sale of assets, your information may be transferred as part of that transaction.</p>

Children’s Privacy</h2> Our Services are not directed to children under 16. We do not knowingly collect personal information from children. If you believe we have collected information from a child, please contact us immediately.</p>

Changes to This Policy</h2> We may update this Privacy Policy from time to time. We will notify you of material changes by posting the new policy on this page and updating the “Last updated” date. Your continued use of the Services after changes constitutes acceptance of the updated policy.</p>

Contact Us</h2> If you have questions about this Privacy Policy or our privacy practices, contact us at:</p>

Security

Security Architecture</h2>

Zero-Trust Model</h3> We operate on a zero-trust security model where:</p>

Encryption</h3> In Transit</strong>: All communications use TLS 1.3 with modern cipher suites</li> At Rest</strong>: Sensitive data encrypted using AES-256-GCM</li> Cargo Data</strong>: End-to-end encryption for sensitive Cargo payloads</li>

Island Security</h2>

Consumer Security</h2>

Infrastructure Security</h2>

Vulnerability Disclosure</h2> We welcome responsible disclosure of security vulnerabilities.</p>

Reporting a Vulnerability</h3> If you discover a security issue, please report it to:</p> Email</strong>: hey@archipelag.io</a></p> PGP Key</strong>: A security.txt</code> file with our PGP key is planned and will be available at /.well-known/security.txt</code></p>

What to Include</h3> Description of the vulnerability</li> Steps to reproduce</li> Potential impact</li>

Scope</h3> In-scope systems include:</p>

Security Best Practices</h2>

Contact</h2> For security questions or concerns:</p> Security Team</strong>: hey@archipelag.io</a></li> General Support</strong>: hey@archipelag.io</a></li> </ul> Security is an ongoing process. We continuously improve our security posture and welcome feedback from our community.</em></p>

Terms of Service

3. Account Registration</h2>

3.1 Account Creation</h3> To use most features, you must create an account. You agree to:</p>

4. User Conduct</h2>

4.1 Acceptable Use</h3> You agree to use the Services only for lawful purposes. You will not:</p>

4.2 Content Guidelines</h3> When using AI models through our Services, you agree not to generate content that:</p>

4.3 Enforcement</h3> We reserve the right to review content and suspend accounts that violate these guidelines. Serious violations may be reported to law enforcement.</p>

5. Island Terms</h2>

5.1 Island Requirements</h3> To participate as an Island (compute contributor), you must:</p>

6. Credits and Payments</h2>

7. Intellectual Property</h2>

7.1 Our Property</h3> The Services, including software, APIs, documentation, and branding, are owned by Archipelag.io and protected by intellectual property laws. You receive a limited, non-exclusive license to use the Services in accordance with these Terms.</p>

7.2 Your Content</h3> You retain ownership of content you submit to the Services. By submitting content, you grant us a license to process that content as necessary to provide the Services.</p>

7.3 Output Content</h3> Content generated by AI models through our Services is provided as-is. You are responsible for ensuring your use of generated content complies with applicable laws.</p>

9.1 Service Availability</h3> The Services are provided “as is” and “as available.” We do not guarantee uninterrupted or error-free operation. We may modify, suspend, or discontinue features at any time.</p>

9.2 AI Output</h3> AI-generated content may contain errors, biases, or inappropriate material. You are responsible for reviewing and validating any output before use. We do not guarantee the accuracy, completeness, or suitability of AI-generated content.</p>

9.3 Third-Party Services</h3> The Services may integrate with third-party services (Stripe, GitHub, etc.). Your use of those services is subject to their respective terms and privacy policies.</p>

10. Limitation of Liability</h2> TO THE MAXIMUM EXTENT PERMITTED BY LAW:</p>

11. Indemnification</h2> You agree to indemnify and hold harmless Archipelag.io, its officers, directors, employees, and agents from any claims, damages, losses, or expenses (including reasonable attorneys’ fees) arising from:</p>

12. Dispute Resolution</h2>

12.1 Informal Resolution</h3> Before filing any formal dispute, you agree to contact us at hey@archipelag.io</a> and attempt to resolve the dispute informally for at least 30 days.</p>

12.3 Jurisdiction</h3> Any disputes not resolved informally shall be resolved in the courts of Amsterdam, the Netherlands.</p>

13. Changes to Terms</h2> We may modify these Terms at any time. Material changes will be notified via email or prominent notice on our website at least 30 days before taking effect. Your continued use of the Services after changes constitutes acceptance of the modified Terms.</p>

14. General Provisions</h2>

14.1 Entire Agreement</h3> These Terms, together with the Privacy Policy, constitute the entire agreement between you and Archipelag.io regarding the Services.</p>

14.2 Severability</h3> If any provision of these Terms is found unenforceable, the remaining provisions will continue in effect.</p>

14.3 Waiver</h3> Our failure to enforce any right or provision does not constitute a waiver of that right or provision.</p>

14.4 Assignment</h3> You may not assign your rights under these Terms without our consent. We may assign our rights at any time.</p>

14.5 Force Majeure</h3> We are not liable for delays or failures caused by circumstances beyond our reasonable control.</p>

15. Contact</h2> For questions about these Terms, contact us at:</p>

Introducing Archipelag.io

How We Share Your Information</h2>
We do not sell your personal information. We share information only in these circumstances:</p>

For Legal Reasons</h3>
We may disclose information if required by law, court order, or government request, or to protect rights, safety, or property.</p>

Business Transfers</h3>
If we are involved in a merger, acquisition, or sale of assets, your information may be transferred as part of that transaction.</p>

Children’s Privacy</h2>
Our Services are not directed to children under 16. We do not knowingly collect personal information from children. If you believe we have collected information from a child, please contact us immediately.</p>

Changes to This Policy</h2>
We may update this Privacy Policy from time to time. We will notify you of material changes by posting the new policy on this page and updating the “Last updated” date. Your continued use of the Services after changes constitutes acceptance of the updated policy.</p>

Vulnerability Disclosure</h2>
We welcome responsible disclosure of security vulnerabilities.</p>

Reporting a Vulnerability</h3>
If you discover a security issue, please report it to:</p>
Email</strong>: hey@archipelag.io</a></p>
PGP Key</strong>: A `security.txt</code> file with our PGP key is planned and will be available at /.well-known/security.txt</code></p>`

Contact</h2>
For security questions or concerns:</p>

Security Team</strong>: hey@archipelag.io</a></li>
General Support</strong>: hey@archipelag.io</a></li> </ul>

Security is an ongoing process. We continuously improve our security posture and welcome feedback from our community.</em></p>

4.3 Enforcement</h3>
We reserve the right to review content and suspend accounts that violate these guidelines. Serious violations may be reported to law enforcement.</p>

7.1 Our Property</h3>
The Services, including software, APIs, documentation, and branding, are owned by Archipelag.io and protected by intellectual property laws. You receive a limited, non-exclusive license to use the Services in accordance with these Terms.</p>

7.2 Your Content</h3>
You retain ownership of content you submit to the Services. By submitting content, you grant us a license to process that content as necessary to provide the Services.</p>

7.3 Output Content</h3>
Content generated by AI models through our Services is provided as-is. You are responsible for ensuring your use of generated content complies with applicable laws.</p>

9.1 Service Availability</h3>
The Services are provided “as is” and “as available.” We do not guarantee uninterrupted or error-free operation. We may modify, suspend, or discontinue features at any time.</p>

9.2 AI Output</h3>
AI-generated content may contain errors, biases, or inappropriate material. You are responsible for reviewing and validating any output before use. We do not guarantee the accuracy, completeness, or suitability of AI-generated content.</p>

9.3 Third-Party Services</h3>
The Services may integrate with third-party services (Stripe, GitHub, etc.). Your use of those services is subject to their respective terms and privacy policies.</p>

12.1 Informal Resolution</h3>
Before filing any formal dispute, you agree to contact us at hey@archipelag.io</a> and attempt to resolve the dispute informally for at least 30 days.</p>

12.3 Jurisdiction</h3>
Any disputes not resolved informally shall be resolved in the courts of Amsterdam, the Netherlands.</p>

13. Changes to Terms</h2>
We may modify these Terms at any time. Material changes will be notified via email or prominent notice on our website at least 30 days before taking effect. Your continued use of the Services after changes constitutes acceptance of the modified Terms.</p>

14.1 Entire Agreement</h3>
These Terms, together with the Privacy Policy, constitute the entire agreement between you and Archipelag.io regarding the Services.</p>

14.2 Severability</h3>
If any provision of these Terms is found unenforceable, the remaining provisions will continue in effect.</p>

14.3 Waiver</h3>
Our failure to enforce any right or provision does not constitute a waiver of that right or provision.</p>

14.4 Assignment</h3>
You may not assign your rights under these Terms without our consent. We may assign our rights at any time.</p>

14.5 Force Majeure</h3>
We are not liable for delays or failures caused by circumstances beyond our reasonable control.</p>