Turn your idle Mac into a tiny data center

Most people think running AI means paying a giant like OpenAI and trusting their prompts to a warehouse full of GPUs. This project flips that completely. In a recent video, AI creator Matthew Berman broke down Dark Bloom, and I was genuinely fascinated when I saw what the team behind it built. Instead of one massive data center in the middle of nowhere, they turned thousands of regular home computers into one giant distributed one.

Here’s the simplest way to picture it. You install Dark Bloom on your Mac, and while your machine is just sitting there doing nothing, it quietly shares some of its compute with other people who need to run AI models. In return, you get paid. Think solar panels, but for AI inference. Everyone generates a little, and together it adds up to something huge.

The old way vs the new way

Let’s talk about the contrast, because that’s the whole point.

The old way looks like this:

• A single enormous building packed with hundreds of thousands of GPUs.
• Placed far from cities, drawing enormous power from the grid.
• Increasingly unpopular, with plenty of public pushback lately.
• Controlled by a handful of very large companies.

The new way that the creators of Dark Bloom are testing:

• Break that data center into tiny pieces spread across ordinary homes.
• Each person owns their own compute and earns from it.
• No giant concentrated facility, no new power plant.
• Open source, open weight models running on distributed machines.

Matthew made a point that stuck with me. This isn’t really about the money. The exciting part is the shift in power. Instead of everyone depending on a few concentrated data centers, compute gets spread out to regular people. That’s a very different paradigm than what we’re used to.

What it actually runs

Dark Bloom serves smaller open source models that almost anyone can run, like Qwen, Gemma, and GPT OSS. On their own, these are modest. But when you aggregate thousands of nodes, you get serious inference power at a fraction of the usual cost.

And this isn’t just theory. According to the video, Dark Bloom is already live on OpenRouter, running roughly 50% cheaper than other providers there. The original poster mentioned it has already served around 4.5 billion tokens in about a week, and the project is only a few days old. That’s wild growth.

So there are two ways to plug in:

1. As a user, you connect and run models on someone else’s machine, even if your own computer isn’t powerful enough.
2. As a provider, you let your idle Mac serve other people and earn from it.

This is peer to peer AI. You’re almost paying another person directly instead of a corporation.

But what about privacy?

This was my first worry too. If your prompt runs on a stranger’s Mac, can they read it? The team addressed this head on in a full white paper. The machine’s owner has root access and physical custody, yet the design makes sure they still can’t see your prompts or the model’s responses.

The way the author explained it, the inference engine runs inside a single hardened Swift process. No subprocesses, no local server, no interprocess communication that could leak data. It uses Apple’s MLX Swift on the Apple Silicon GPU. In plain terms, they closed off every software path where your data could be observed. That’s a big deal if you’ve ever hesitated to send data to any cloud provider.

How to try it, per the video

Matthew walked through the setup, and it’s refreshingly short:

1. Go to darkbloom.dev and click the earn section.
2. Create a Dark Bloom account.
3. Install the CLI (a Mac app is coming soon). It’s basically one command.
4. Approve the provider enrollment under device management in macOS settings.
5. Connect a Stripe account so you can get paid.

He even had Codex handle the whole install for him, then just signed in with a code. Once enrolled, it tested his machine and clocked 78 tokens per second running GPT OSS 20B. After that, it runs in the background.

A few honest notes the creator shared:

• 💰 A Mac Studio with an M5 Ultra and 96GB might earn around $37 a month. Small, but close to free money since the Mac is idle anyway.
• 🔋 Apple machines are efficient, so the extra electricity cost is low. Your exact numbers depend on local rates.
• 🖥️ You currently need at least 48GB of RAM. They raised the bar because demand was so high.
• 🏦 Stripe sits in the middle, so Dark Bloom can send you money but can’t touch your bank.

The original poster was upfront about the risks too. A lot of comments called it a scam, so he had GPT 5.6 review the entire public codebase. The only real flag was that you currently keep 100% of revenue, which could change later. No hidden malware, no crypto mining, no credential theft. He still said use it at your own risk, since it’s very early and rough edges are expected.

Why this matters

I think this is one of the most compelling arguments yet for open source, open weight models. When the weights are public, regular people can build things like this and own their own compute. Distributed AI stops being a slide in a pitch deck and starts feeling real.

Want the full walkthrough, including the live install and the earnings breakdown? Watch Matthew Berman’s video and see it in action for yourself.