Descriptions:
Matthew Berman walks through Dark Bloom, a newly launched distributed inference platform that turns consumer Macs into nodes of a peer-to-peer AI compute network. Rather than routing requests through centralized data centers, Dark Bloom aggregates spare compute from individual machines — currently requiring at least 48 GB of unified memory — and serves open-weight models including Qwen 3, Gemma 4, and GPT OSS at prices roughly 50% below comparable OpenRouter providers. Within its first week the network had already processed 4.5 billion tokens.
The video’s most technically interesting segment covers Dark Bloom’s privacy architecture: inference runs inside a single hardened Swift process using Apple’s MLX Swift LM engine on Apple Silicon, with no subprocesses, local servers, or interprocess communication. The design means the machine owner cannot observe a user’s prompts or the model’s responses — addressing the trust problem that typically makes peer-to-peer AI impractical.
Berman demonstrates a full installation using Codex to automate the setup, confirms 78 tokens per second on his test machine, and walks through the Stripe-based payment system that credits node operators. The video also briefly examines the broader framing: distributed inference as a structural alternative to large-scale data centers, analogous to rooftop solar as a decentralized alternative to grid power plants. Compensation for a Mac Studio M5 Ultra is estimated at roughly $37 per month.
📺 Source: Matthew Berman · Published August 27, 2026
🏷️ Format: Tutorial Demo
