Darkbloom warns Mac owners their AI earnings are falling as token prices drop

Darkbloom is a project from Eigen Labs, a company based in Seattle. It lets anyone with a capable Mac rent out its graphics chip, the GPU, through a shared network. In return, they get paid.

The work is called inference, which just means running an AI model to get an answer. Someone types a question, a computer somewhere runs the model, and the reply comes back. Darkbloom lets Macs at home do that part.

The project began as an experiment in April. In August, The Register reported that it had become a paid provider on OpenRouter, after first serving jobs there for free. An Eigen Labs engineer said at the time that people were earning $120 to $200 a month per machine. That is a long way from today's warning.

The Register reports about 800 people take part, including its own reporter until two days ago. They earn anywhere from a few cents to a few dollars a day. The jobs come in through OpenRouter, an AI model gateway, which is a switchboard that sends each request to a provider.

In real life Think of it like renting out a spare room, except the guest is an AI question and it stays for a few seconds.

The company's own best-case estimate is up to $192 a month. That is for someone running one of Apple's recent Mac minis, with an M5 Pro chip and 64 GB of memory, around the clock. It also assumes a popular model, Qwen3.6 35B A3B, which sells for $0.065 per million input tokens and $0.75 per million output tokens.

A token is the small unit AI work is counted in, roughly a piece of a word. AI companies bill by the million tokens.

Not everyone gives their Mac to the network full time. The Register says the median earnings of the top 50 people on the leaderboard are about:

• Per day: $5.26
• Per month: $159.86
• Per year: $1,918.50

The Register says the best-case estimate looks plausible, but may not be sustainable. That last word is the whole story.

Because tokens are getting cheaper. The Register points to Ramp's AI Token Price Index, which tracks what businesses pay Anthropic and OpenAI. For Anthropic, the price per million tokens went from $2.65 on May 1, 2026 to $0.62 on October 22, 2026. That is a drop of 77%. The Register also notes another report of a 54% fall since May.

Tom's Hardware adds that token prices are falling so fast they have outpaced Moore's Law, the old rule that chips get twice as powerful every couple of years. Some reports it cites say AI is getting 50% cheaper every quarter.

If the 77% drop is applied to the median earner, Tom's Hardware calculates earnings would fall to about $36.77 a month, or $1.23 a day. That is where the headline number comes from.

The big picture: Cheaper AI is good news for people who use it. It is less fun for someone selling it from a Mac mini.

The Register says a researcher at Eigen Labs, who goes by Kydo, wrote to members on the company's Slack channel. He celebrated the network's trillionth token, then tried to set expectations.

"Earnings have been low. For most providers, I don't expect that it is going to recover soon."

He gave three reasons. First, data centers got cheaper to run:

"Output per NVIDIA cluster is up ~20x this year, and their margins are wide enough to cut prices 90 percent and still profit at 50 percent."

Second, the network lost its edge. "The market caught up to our price," he said. Going lower would leave most members with even less.

Third, OpenRouter is changing how it picks providers. Both outlets report it now cares more about speed, in particular decode speed (how fast words appear in a reply), prefill speed (how fast the model reads your question) and quantization (a way of shrinking a model so it runs faster). It cares less about the lowest price.

The older and smaller ones. Kydo said that for owners without a top-of-the-line Max or Ultra Apple chip, "inference won't be worth running."

That matters because plenty of members use ordinary machines. A fast, expensive Mac suits the new speed rules. A basic one fits the old price-first rules, which are fading.

Tom's Hardware says Darkbloom does not want some members to lose out without help. It plans to study which models are most in demand and tell members which to run.

The team listed several ideas, according to The Register and Tom's Hardware:

• Steer members to popular models: earn more by running what buyers want right now.
• Other kinds of work: The Register says this includes CI/CD jobs, the automatic checks that test new code before it goes live. The idea is that AI agents, which means programs that do tasks on their own, now write code changes faster than those checks can keep up.
• Sandboxing: that work needs a sealed-off space to run safely, which the company expects to start testing soon.
• Faster responses: improve time to first token, meaning how quickly a reply starts.
• New buyers: find customers beyond OpenRouter.
• High-end clusters: The Register says the hardware focus now looks like groups of top Mac Studios, which come with up to 512 GB of memory and link up over Thunderbolt.

The catch: none of this is a promise of higher pay. Kydo was blunt about that.

Honest, and a bit tough. Kydo wrote that the next stretch "will be a fight," and he expects "many machines will leave and Slack will get quieter."

But he also said those who stay will be first in line when new ways to earn show up. "We are a driven team and well capitalized. We will not give up."

Tom's Hardware adds a wider point. Companies are still blowing through their AI budgets in months. That is less about the price per token and more about using far more tokens, as AI agents grow busier. Cheap tokens do not always mean small bills.

• Do not buy a Mac to earn money: the best-case figure assumes a high-end machine, running all day, on the most popular model.
• Check the math first: electricity and noise count, and payouts can move quickly with token prices.
• Older Macs may earn little: the network itself says basic hardware may not be worth running.
• If you use AI: the same price drop that squeezes Darkbloom is part of why AI keeps getting cheaper.