Darkbloom: Decentralized AI Inference on 100M+ Idle Apple Silicon Machines

Eigen Labs has unveiled Darkbloom, a decentralized inference network that transforms idle Apple Silicon machines into a distributed AI compute platform. The system connects over 100 million Apple devices directly to inference demand, bypassing the traditional three-layer markup chain of GPU manufacturers, hyperscalers, and API providers.

The Problem: Three Layers of Markup

The current AI compute market operates through multiple intermediaries. NVIDIA sells GPUs to hyperscalers like AWS and Google Cloud, which mark up prices when renting capacity to AI companies. These companies then add their own markup before charging end users per token. This structure means users pay multiples of the actual silicon cost.

Meanwhile, Apple has shipped over 100 million machines with serious ML hardware since 2020. These devices feature unified memory architectures with 273 to 819 GB/s memory bandwidth and Neural Engines capable of running 235-billion-parameter models. Most sit idle for 18 or more hours daily, earning nothing from their compute potential.

How Darkbloom Works

Darkbloom eliminates every software path through which operators could observe inference data. The system employs four independent, verifiable layers:

Encryption: Requests are encrypted on the user's device before transmission. The coordinator routes ciphertext that it cannot read. Only the target node's hardware-bound key can decrypt the data.

Hardware Verification: Each node holds a key generated inside Apple's tamper-resistant secure hardware. The attestation chain traces back to Apple's root certificate authority.

Runtime Hardening: The inference process is locked at the OS level. Debugger attachment is blocked, and memory inspection is prevented. Operators cannot extract data from running processes.

Output Traceability: Every response is signed by the specific machine that produced it. The full attestation chain is published for independent verification.

OpenAI-Compatible API

The platform offers an OpenAI-compatible API that requires only a base URL change. All existing SDKs work without modification. The system supports streaming, function calling, image generation with FLUX.2 on Metal, speech-to-text with Cohere Transcribe, and large MoE models up to 239 billion parameters.

Cost Comparison

Idle hardware has near-zero marginal cost, allowing savings to pass directly to users. Darkbloom shows up to 70% lower costs compared to centralized alternatives:

Model	Input Cost	Output Cost	OpenRouter Cost	Savings
Gemma 4 26B	$0.03	$0.20	$0.40	50%
Qwen3.5 27B	$0.10	$0.78	$1.56	50%
Qwen3.5 122B	$0.13	$1.04	$2.08	50%
MiniMax M2.5 239B	$0.06	$0.50	$1.00	50%

Image generation costs $0.0015 per image compared to Together.ai's $0.0030. Speech-to-text runs at $0.0010 per audio minute versus AssemblyAI's $0.0020.

Operator Economics

Hardware owners contribute idle Apple Silicon and earn 100% of inference revenue. The only variable cost is electricity, which runs $0.01–0.03 per hour depending on workload. This translates to approximately 90% profit margins.

Installation is straightforward via a CLI Mac App that downloads the provider binary and configures a launchd service. The system auto-updates and runs in the background.

Available Models

Darkbloom offers curated models selected for quality:

• Gemma 4 26B: Google's latest fast multimodal MoE with 4B active parameters
• Qwen3.5 27B: Dense, frontier-quality reasoning (Claude Opus distilled)
• Qwen3.5 122B: MoE with 10B active parameters for best quality per token
• MiniMax M2.5 239B: SOTA coding with 11B active parameters, 100 tokens/second on Mac Studio
• Cohere Transcribe 2B: Best-in-class speech-to-text

The platform represents a fundamental shift in AI compute economics, similar to how Airbnb connected idle rooms to travelers and Uber connected idle cars to riders. By eliminating the hyperscaler middleman, Darkbloom offers lower prices while enabling hardware owners to monetize idle compute resources without compromising user privacy.