Eigen Labs · 2026-04-15 · notable
Darkbloom — Private AI Inference on Idle Macs by Eigen Labs
Eigen Labs launches Darkbloom: decentralized AI inference on idle Apple Silicon Macs with four hardware-backed privacy layers. Costs ~50% less than centralized providers; operators keep 100% of revenue. Now live in research preview.
Run AI inference on idle Apple Silicon Macs with hardware-backed privacy and half the cost of centralized providers.
Key specs
| Cost vs. centralized | ~50% less |
|---|---|
| Operator revenue share | 100% |
| Max model size | 239B parameters |
| Minimum mac os | macOS 14 (Sonoma) |
What is it?
Darkbloom is a research-preview inference network from Eigen Labs (the team behind EigenLayer) that routes encrypted AI requests to idle Mac computers instead of data centers. Mac owners install a provider agent; their machine starts serving inference during idle time. The full stack — coordinator, hardened provider agent, Secure Enclave integration, operator tooling, and web console — is open-sourced under the Layr-Labs/d-inference repository. Supports text (Gemma, Qwen, MiniMax), image generation (FLUX.2), and speech-to-text (Cohere Transcribe) via an OpenAI-compatible API.
How does it work?
Privacy is enforced through four independent layers: requests are encrypted on the user's device before transmission; each Mac node uses Apple's Secure Enclave for hardware-backed attestation proving it's genuine Apple Silicon; the hardened runtime blocks debugger attachment and memory inspection; and every response is signed by the specific machine that produced it. The coordinator routes traffic without reading it. The result is that neither Eigen Labs nor the Mac operator can see the contents of a request.
Why does it matter?
Centralized inference providers have both the compute and all the data passing through it. Darkbloom's hardware attestation model shifts this: the privacy guarantee is enforced by Apple's Secure Enclave rather than a policy document. With over 100 million Apple Silicon Macs idle most of each day and capable of running models up to 239B parameters, the distributed capacity is real. Cost cuts follow from eliminating data center margins — providers keep all revenue after electricity.
Who is it for?
Privacy-conscious developers and Mac owners who want to monetize idle compute.