Compiled by: Felix, PANews
AI inference is gradually becoming a critical layer of internet infrastructure. However, most inference currently still relies on a centralized architecture, which is costly, has limited capacity, involves multiple layers of stacking, and carries certain security risks. At the same time, there are millions of powerful computers worldwide that remain idle for most of the day.
Eigen Labs recently launched the AI inference network Darkbloom, which explores performing distributed AI inference on idle Mac computers. By combining verified nodes, hardware-level privacy protection, and superior economic efficiency, it transforms idle Apple Silicon chips into a more efficient, privacy-first computing network.
The project was launched as a research preview around April this year, upgraded to a public alpha version in May, and is now available on the OpenRouter platform. In the alpha version, the available models are Google's Gemma 4 and OpenAI's GPT-OSS.
Core Architecture and Verifiable Privacy
The Darkbloom network consists of three parts: users, coordinators, and providers.
- Users can send inference requests through a chat interface or a compatible OpenAI API.
- The coordinator (operated by Eigen Labs) routes these requests to eligible Macs in the network.
- Providers (users who own these eligible Macs) run the models and return the output, but they cannot see the request content.
Darkbloom is built on a privacy-first distributed inference model. The provider process is hardened to resist common local inspection paths, including debugger attachment and external memory inspection. The integrity of the running binary is also part of the trust model, helping to ensure that the software serving requests conforms to network expectations.
The system also uses hardware-supported attestation based on Apple's security architecture. Secure Enclave keys, attestation signals, and periodic challenge-response checks are used to verify that participating nodes are running with the intended protections and software state, achieving truly verifiable privacy.
Economic Model and Daily Earnings
Darkbloom is fundamentally different in its business model compared to the vast majority of projects. In the traditional tech stack, costs include hardware, facilities, cooling, networking, operational overhead, and layers of profit margins. In Darkbloom's model, the hardware already exists, and the marginal cost is primarily driven by electricity. Darkbloom's benchmark pricing is only about 50% of current mainstream API aggregators. Providers (Mac hosts) can keep 100% of the inference revenue. Furthermore, Darkbloom has not taken the path of issuing tokens to subsidize early participants; node earnings come entirely from real AI inference demand.
It is worth noting that, given the project's early stage of development, earnings are relatively modest. Factors such as memory and hardware configuration, uptime, model demand, node health, and network demand can all influence earnings to some extent.
Current leaderboard data shows that the top provider earns less than $6 per day, and the fifth-ranked provider earns even less than $2. However, as the network opens up to large language models with high memory requirements and real user usage increases, this situation is expected to improve.
Regarding how to set up an idle Mac, the steps are as follows:
- Acquire a Mac with an Apple Silicon chip
- Ensure it runs macOS 14 or higher
- Install the Darkbloom provider
- Keep the Mac online with a stable internet connection
- Let the network route supported AI tasks
Related reading: A Roundup of Recent Stocks and Crypto Assets Worth Watching: AI, RWA, and Space Stocks...
