Eigen Labs released Darkbloom on April 15-16, 2026, a decentralized AI inference network that transforms idle Apple Silicon machines into a private, cost-effective AI computation platform. The project quickly gained traction on Hacker News with 379 points and 187 comments, highlighting developer interest in decentralized alternatives to traditional cloud AI services.
Four-Layer Privacy Architecture Ensures Secure Inference
Darkbloom implements a comprehensive privacy model with four distinct layers of protection. Requests receive end-to-end encryption before transmission, while hardware-verified nodes leverage Apple's Secure Enclave for attestation. The system employs a hardened runtime that blocks debugger attachment and memory inspection at the kernel level using PT_DENY_ATTACH. Each inference output is cryptographically signed by the specific machine that produced it, creating an auditable chain of computation. The inference engine runs entirely in-process with no subprocess, local server, or inter-process communication, minimizing attack surfaces.
Pricing Undercuts Major Cloud Providers by Up to 70%
Darkbloom offers text model inference at approximately 50% of OpenRouter's rates, representing up to 70% savings compared to Azure OpenAI and AWS Bedrock. Image generation costs $0.0015 per image, half of Together.ai's $0.003 pricing. Speech-to-text processing runs at $0.001 per audio minute. The platform provides an OpenAI-compatible API, allowing developers to migrate existing applications by simply changing the base URL while maintaining support for streaming, function calling, and multiple modalities.
Operators Earn $890-1,190 Monthly With 90% Profit Margins
Mac owners who join the network as operators retain 95% of token revenue, with only 5% covering network infrastructure costs. A Mac Studio M3 Ultra with 192GB RAM running 18 hours daily projects $900-1,200 in monthly revenue. At Apple Silicon's 30-watt power draw, electricity costs approximately $11 monthly, resulting in net profits of $890-1,190 per month—roughly a 90% margin. The network currently supports models including Gemma 4 26B, Qwen 3.5 27B and 122B MoE variants, MiniMax M2.5 239B for coding, FLUX.2 Klein 9B for image generation, and Cohere Transcribe for speech-to-text.
Experimental Status With Research Foundation
The project remains in experimental prototype phase, with developers warned to expect rough edges, breaking changes, and potential downtime. Installation requires a CLI terminal command, with a native macOS menu bar application currently in development. Eigen Labs published a research paper detailing the architecture, threat model, and economic analysis behind the network. The GitHub repository provides full technical documentation for developers interested in contributing or deploying operator nodes.
Key Takeaways
- Darkbloom uses Apple Silicon's secure enclave and kernel-level protections to provide hardware-verified private AI inference with end-to-end encryption
- Pricing undercuts major cloud providers by up to 70%, with text models at 50% of OpenRouter rates and image generation at $0.0015 per image
- Operators running Mac Studio M3 Ultra machines 18 hours daily can earn $900-1,200 monthly with approximately 90% profit margins after electricity costs
- The network provides an OpenAI-compatible API supporting streaming, function calling, and multiple modalities across text, image, and speech models
- Currently in experimental prototype phase with available models including Gemma 4 26B, Qwen 3.5 variants, MiniMax M2.5 239B, and FLUX.2 Klein 9B