THE ESSENTIALS IN 30 SECONDS

  • Darkbloom, built by Seattle-based Eigen Labs, routes inference jobs from OpenRouter to Macs whose owners rent out their GPUs; The Register counts about 800 participants.
  • An Eigen Labs researcher told providers earnings have been low and, for most, are unlikely to recover soon, citing cheaper data-centre inference.
  • The team plans to steer providers toward in-demand models, test distributed CI/CD jobs, focus on high-end Mac Studio clusters and look for buyers beyond OpenRouter.

Darkbloom, a project that lets owners of Apple Silicon Macs earn money by running AI models for other people, has warned its contributors that payouts are low and may stay that way. The Register reported the message on October 30.

How the network works

Darkbloom is run by Eigen Labs, a Seattle-based company. Its website describes a private inference network powered by Macs: providers install an app or command-line tool, and customers call models through an OpenAI-compatible API. According to The Register, jobs are submitted through OpenRouter, the model gateway, where Darkbloom is listed as a provider for 12 open-weight models. The Register puts participation at about 800 people; the reporter, who is one of them, says he earned between a few cents and a few dollars a day.

Darkbloom's own best-case estimate is that a Mac mini with an M5 Pro chip and 64 GB of memory could make up to $192 a month if fully used around the clock. The median among the 50 people on its leaderboard was about $5.26 a day, according to The Register. These figures come from the company and have not been independently checked.

Why earnings fell

In a Slack message to providers, Eigen Labs researcher Kydo said earnings had been low and that, for most providers, he did not expect them to recover soon. He argued that data-centre inference has become much cheaper this year, removing the price advantage Darkbloom once had, and that cutting prices further would mostly reduce what providers earn. The Register cites Ramp's AI Token Price Index, which shows Anthropic's price per million tokens falling about 77% between May and late October.

According to the report, Kydo said owners of older or less powerful Macs, without high-end Max or Ultra chips, would probably find inference not worth running.

What comes next

Eigen Labs plans to analyse demand so providers can serve the most requested models, and to offer other work such as distributed CI/CD jobs, which require sandboxing and are due for testing soon. It also wants to focus on clusters of high-memory Mac Studios linked over Thunderbolt, improve time to first token and find token buyers beyond OpenRouter. Kydo told providers he expects many machines to leave the network. No timeline has been given for these changes, and The Register's report includes no comment from OpenRouter.