Dark Bloom lets Mac owners enroll their machines as inference providersAn AI inference provider operates models as a service and returns outputs for application requests., combining many small nodes into a distributed alternative to centralized data centersDistributed AI inference uses multiple networked machines to serve model requests instead of relying on a single computer or centralized provider.. The host reports that its models were already available through OpenRouter and that the network claimed to have served 4.5 billion tokens during its first week.
The system currently targets Apple Silicon and uses MLX Swift LM inside a hardened Swift process. Its white paper claims to remove software paths that would let a machine owner inspect remote prompts and responsesPrivacy-preserving AI inference protects prompts, model inputs, outputs, and associated metadata from parties that are not authorized to inspect them., but the video does not independently verify that security model. Provider hardware must currently have at least 48 GB of memory.
Berman installs the provider software with Codex, completes device enrollment and measures GPT-OSS 20B at 78 tokens per second on his machine. He had not yet served a paid job or received revenue when recording, so the estimated earnings and electricity economics remained projections rather than demonstrated results.
The larger idea is to turn idle consumer hardware into shared open-model capacity while reducing dependence on centralized AI infrastructureDecentralized AI compute pools processing capacity from independent machines so model workloads are not limited to one cloud provider or data-center operator.. The project is still experimental, and the host advises using it at your own risk despite finding no obvious malware in an AI-assisted source review.
Watch on YouTube
