Archipelag is a decentralized compute marketplace that routes AI inference jobs to nearby, identity‑verified hosts (Islands) using real‑time latency measurements, delivering lower response times and data residency control. Hosts set market‑based rates and earn credits for actual compute usage, while workloads run in signed, sandboxed containers with zero‑trust security guarantees. The platform offers OpenAI‑compatible, streaming‑first APIs for chat, image generation, batch processing, and multi‑Island parallelism.
Funding
Funding not disclosed
Founders
Product
Problem
Developers and AI users often rely on centralized cloud providers for inference, leading to high latency, opaque pricing, and limited control over data residency. Simultaneously, owners of idle GPU resources have no straightforward way to monetize their hardware without complex setup or trust concerns.
Solution
Archipelag offers a decentralized compute marketplace where signed, sandboxed workloads—called Cargos—are dispatched to nearby, identity‑verified hosts called Islands. The platform measures real‑time round‑trip latency and routes jobs to the lowest‑latency Island, reducing response times and keeping data within specified jurisdictions. Hosts set their own rates and earn credits based on actual compute usage, while consumers pay market‑clearing prices that reflect supply, demand, and location. Security is enforced through cosign‑verified containers, seccomp profiles, and optional confidential inference using TEEs or encrypted pipelines. The service supports OpenAI‑compatible APIs for chat, image generation, batch processing, pipeline parallelism, and workflow orchestration, all with token‑by‑token streaming over WebSocket or SSE.
Target Audience
Primary customers are AI developers, SaaS providers, and enterprises that need low‑latency, cost‑transparent inference, as well as GPU owners and mobile device users who want to earn credits by contributing compute to the network.
Features
- Proximity routing based on measured RTT to deliver low‑latency inference from the nearest Island
- Market‑based pricing with host‑set rates and transparent clearing prices shown before execution
- Zero‑trust execution: signed Cargo containers run in sandboxed environments with seccomp filtering
- Optional data residency policies to restrict job placement to specific regions (e.g., EU, Switzerland)
- Streaming‑first API delivering token‑by‑token responses for LLM chat and progress updates for image generation
- Multi‑Island pipeline parallelism that automatically shards large models across 2–8 Islands
- Batch fan‑out allowing up to 100 parallel inputs in a single API call with real‑time progress tracking
- Confidential inference with hardware TEEs (Intel SGX, AMD SEV, ARM TrustZone) or software‑only homomorphic encryption