Darkbloom, a service that rents out the unused computing power of Macs for AI work, has warned hardware providers that a quick recovery in sagging payouts is unlikely. According to The Register's report dated October 30, 2026, the operator delivered the grim outlook in a Slack channel for participants. Darkbloom has home Macs handle the "inference" processing needed to generate AI responses, but price competition has intensified, and simply keeping a machine connected to the network no longer brings in enough work. The figure of "$1.23 a day" that drew attention is not actual income; it is an estimate that applies a price decline from a different market. To understand real earnings, you need to look at the payout structure and the amount of work actually assigned, separately.
The $1.23 a day is an assumption-based estimate, not real income
According to The Register, about 800 people take part in Darkbloom, and the publication's reporter was one of them until two days before the article ran. In the participants' Slack, Kydo acknowledged the weak earnings, saying, "We don't expect most providers to recover in the near term." This was not an announcement that the service itself is ending, but a warning about providers' earnings outlook.
The median among the top 50 earners on the leaderboard, as reviewed by The Register, was $5.26 per day, or $159.86 per month. However, this is not the average income of all participants, nor is it profit after deducting costs such as electricity. Meanwhile, in an article dated October 31, Tom's Hardware applied a 77% decline in AI token prices to the $159.86 monthly figure, arriving at $36.77 per month, or about $1.23 per day. Still, actually observed earnings and a hypothetical amount calculated from a price decline need to be kept distinct.
| Figure | Basis / calculation conditions | What the figure does not tell us |
|---|---|---|
| $5.26 per day, $159.86 per month | Median of the top 50 earners on the leaderboard, as reviewed by The Register | Average income of all participants, or profit after expenses |
| From $2.65 to $0.62 per million tokens, a drop of about 77% | Ramp's Anthropic price index as reported by The Register; comparison of May 1 and October 22, 2026 | That payouts for the same Mac or the same model actually fell 77% |
| $36.77 per month, about $1.23 per day | Tom's Hardware's estimate applying the 77% decline to $159.86 per month; daily figure based on a 30-day month | Actual payment amounts or a forecast of future income |
The calculation is $159.86 × (1 − 0.77) = $36.7678, which divided by 30 days gives $1.2255933. In other words, the roughly $1.23 per day figure is merely the amount you get if you assume the $159.86 monthly figure fell uniformly by 77%. It is not a measured value reflecting Darkbloom's base pay or fluctuations in workload, so it cannot be concluded that providers' income actually fell to that level.
Nor does the 77% decline mean that the list price of a particular Claude model was cut by 77%. It is a change in a price index from Ramp reported by The Register, not the result of continuously tracking the same Mac, model and payout arrangement. In the first place, Darkbloom is not a service that runs Claude on individual Macs. According to the company's official GitHub repository, it runs models on Apple Silicon Macs using the machine learning framework MLX and makes that computing power available over the network.
Payouts have two parts: base pay and usage-based pay
According to Darkbloom's official explanation, provider payouts are split into "base pay" and "usage-based pay." Base pay comes from a fixed monthly reward pool and is conditioned on, among other things, a verified, healthy machine being online for at least 90% of each settlement period. Usage-based pay, on the other hand, arises according to the inference requests actually processed. In other words, the time a Mac is connected to the network does not necessarily match the time it earns revenue from inference work.
The official site gives an earnings example of $204.89 per month for a MacBook Pro with a 16-core M4 Max and 128GB of unified memory. However, this is an estimate that assumes a 60% generation utilization rate. Generation utilization refers not simply to the share of time the machine is powered on, but to the share of time it is actually generating output tokens. The estimate also assumes steady demand continues, and the company itself states that it is a projection and not a guarantee of earnings.
Base pay has eligibility conditions, and its funding is limited to a fixed monthly pool. Usage-based pay varies with actual demand and processing volume, so directly comparing the official $204.89 per month with the reported $159.86 per month cannot tell you whether earnings rose or fell. The two figures involve different devices and calculation conditions, and neither measures the change in earnings in the same environment.
Another important point for thinking about earnings is that the fees customers pay differ by type of processing. In Darkbloom's price list as of November 1, 2026, the per-million-token rates for "gpt-oss-20b" are $0.018 for input, $0.09 for output, and $0.009 for reading cached input back for reuse. Even for the same number of tokens, what customers are charged differs between input processing and output generation. However, these are customer-facing prices and do not translate directly into what providers take home per unit.
Therefore, even if token prices in another market fell 77%, overall earnings including base pay would not necessarily fall by the same proportion. To grasp actual income, you need to check the breakdown between base pay and usage-based pay, along with which models you handled and how much processing you did. Customer-facing prices may change in the future, and even if a model appears on the price list, there is no guarantee that work for that model will be assigned continuously.
Winning AI inference work depends on processing speed, not just price
Another factor shaping Darkbloom's earnings is how inference requests are allocated. OpenRouter, which connects users to multiple AI models and providers, does not necessarily pick the cheapest provider. Under its current official routing specification, the default allocation evaluates effective price, token generation speed, time to first response, and more. It also takes into account capacity limits, reliability and the quality of tool calling.
OpenRouter's official guide explains that providers that are more expensive but faster can still win requests. One key metric here is TTFT (Time to First Token), the time from sending a request until the first token comes back. In addition, generation speed, meaning how many tokens can be output per unit of time once generation begins, also determines how long users wait.
That said, OpenRouter has not stopped considering price. Effective price is still evaluated, and users can also choose providers by prioritizing price, generation speed or latency. Nor has it been confirmed that OpenRouter changed its routing mechanism in response to Darkbloom's warning. The key point is that both price and performance influence how work is allocated.
In its guide for providers, OpenRouter also explains that it allocates requests based on criteria such as price, latency, generation speed and reliability. It publishes uptime as well as TTFT and generation speed, and payouts are settled monthly based on actual usage. Under this system, a provider's usage-based pay is not determined by per-token price alone. How many billable requests it wins and actually processes also greatly affects revenue.
Darkbloom providers are therefore facing two problems: falling per-token prices and the difficulty of securing enough work. Even if they offer low prices, they may not receive enough work if their processing speed or stability lags competitors. At the same time, it is unclear how much of the earnings drop is due to falling unit prices versus performance disadvantages on individual Macs. At this point, there is also not enough data to conclude that buying a higher-performance Mac would improve earnings.
Can new revenue opportunities be found beyond inference?
According to The Register, Kydo outlined measures going forward: analyzing which models are in high demand so providers can choose suitable ones, shortening time to first response, and developing customers beyond OpenRouter. The company also plans to trial distributed execution of CI/CD jobs, which automatically build and test software changes, and a mechanism for running workloads securely in isolated environments. However, these are future concepts and trial-stage efforts; neither formal service offerings nor a recovery in earnings has been confirmed.
Even as AI use expands, revenue for those supplying computing resources does not necessarily grow at the same rate. For example, according to Ramp's AI usage data, token usage in the week ending October 23, 2026 rose 37.1% from the week ending September 25, while spending rose only 8.9%. This is an example showing that even when per-token prices fall, total spending can still grow as processing volume increases.
However, this tally is based on usage and billing data for destinations connected by companies using Ramp's "Token Spend Management," and does not represent Ramp's overall AI spending. Naturally, it is also not a figure showing that requests to Darkbloom increased. In the same weekly data, Anthropic's blended price fell 15.7%, from $0.73 to $0.61 per million tokens, but the comparison period and aggregation unit differ from the 77% figure mentioned earlier.
Darkbloom aims to link the surplus computing power of already-owned Macs with AI inference and other computing demand. But whether it pays to buy a new Mac and join is hard to judge at present, because there are few measured figures for earnings after deducting costs such as electricity. What matters is what kind of work, and how much of it, is assigned once the base pay eligibility conditions are met, and whether income can be earned continuously. If new uses such as faster inference and CI/CD translate into real demand, and post-expense earnings become clear, it should become possible to evaluate more concretely whether turning Macs' surplus computing power into income is financially viable.