Darkbloom pays Mac owners to serve AI inference from idle laptops. Should a business buy from it, or let staff join it?

In the last week of August an engineer at Eigen Labs, Gajesh Naik, posted that a project called Darkbloom had become a paid inference provider on OpenRouter, had served 4.5 billion tokens, was tracking $102,000 in annualised revenue, and had around 250 Macs online. He invited anyone with an idle Mac to enrol it. Within days the pitch had reached Instagram and TikTok in the form that spreads fastest: leave your laptop closed and earn up to $200 a month.

The idea is older than the project. Distributed compute networks have been tried for protein folding, for cryptocurrency and, several times, for AI. What is new is the combination of Apple silicon, which runs a 30-billion-parameter model on a desktop, open-weight models good enough to sell, and a privacy design that borrows from Apple's own Private Cloud Compute. The result is a supplier that undercuts the API market with hardware it does not own. Both halves of that sentence deserve a look from anyone who buys inference or manages a fleet of laptops.

How does Darkbloom work?

Darkbloom routes inference requests from a coordinator to privately owned Macs, which run the model locally and return the result, and pays the owner for the tokens served. A buyer calls an OpenAI-compatible API, so an existing SDK works unchanged. The request lands on a Go coordinator running in a confidential virtual machine, on AMD SEV-SNP hardware, which decrypts it, picks a provider machine and re-encrypts it for that machine alone. The provider's Mac, connected outbound over a WebSocket so no ports need opening, decrypts the job inside the Secure Enclave and runs the model on the GPU using Apple's MLX framework.

The models on offer are open weights that fit in consumer memory: Qwen 3.6 35B A3B, GPT-OSS 20B and Gemma 4 26B were the top earners on the network's public leaderboard in late August. Darkbloom's site says selected model prices are set around 50 per cent below typical API-provider rates for comparable models, with pay-per-token billing and no minimum.

To become a provider you create an account, install a command-line tool, enrol the machine through macOS device management and pass a throughput benchmark, run against GPT-OSS 20B, before the node goes live. The current minimum is 48 GB of unified memory, raised from an earlier threshold after more people applied than the network needed; the team has said it expects that bar to come down. Payouts go through a connected Stripe account, so Darkbloom never holds your bank details, and for now providers keep 100 per cent of what their machine earns, a split the project says may change. The terms of service allow personally owned devices only and bar renting cloud Macs to arbitrage the difference.

How much does a Darkbloom node earn?

Less than the reels claim and more than nothing, with a wide spread that depends on demand rather than on your hardware. Naik's own figure was $120 to $200 a month per machine on average. Darkbloom's earnings calculator put a Mac Studio with an M5 Ultra and 96 GB at about $37 a month. The company's published range for a year is $192 to $4,983, after an estimated electricity cost of almost $2 a month. The top three earners on the leaderboard in late August were averaging $26, $19 and $15 a day, which annualises towards the top of that range, and those are the best machines on a network where supply was already outrunning demand enough to justify raising the entry bar.

The honest summary for an individual is that a high-memory Mac that sits idle most of the day might cover its electricity and a streaming subscription, and might, if the network's demand grows faster than its supply, do better. Nobody should buy hardware for it. The economics for Darkbloom itself are the interesting part: it has no data centre, no power contract and no capital cost, and the supply side is people who already paid for the hardware and will accept a Stripe deposit for a share of its idle time. The Register's headline called it an inference Borg. The commenter on Hacker News who pointed out that the model has zero upfront investment was making the same point without the metaphor.

Should a business buy inference from Darkbloom?

For workloads that are already running on open-weight models and carry no personal or confidential data, yes, as one supplier among several and never the only one. For anything that touches customer data, not until your data protection review has read the white paper and decided the confidential-VM-plus-Secure-Enclave design meets your obligations, and probably not even then for regulated data.

The case for is price and the direction of travel. If your workload is one of the ones where an open model already beats an API, a provider at half the going rate with an OpenAI-compatible endpoint is a two-line change in your router configuration. OpenRouter already lists it, so a team that routes through OpenRouter may be buying from Darkbloom without having decided to. Check your provider preferences.

The case against is everything that comes with running your prompts on a stranger's laptop. The privacy design is well thought through, and independent code review at launch reportedly found no malware, no crypto mining and no credential theft in the open codebase, which is more than most new suppliers can say. But the protections are claims about software and hardware configuration, not a certification, and a reviewer with root on their own Mac is an adversary the design is built to resist rather than one it has been audited against. Availability depends on how many machines happen to be online. There is no service level agreement, no data processing agreement of the kind India's DPDP Rules or the EU AI Act will expect you to hold with a processor, and no enterprise contract on offer. Latency will vary with the residential connection of whichever machine picked up your request.

Where it fits, then, is the same place cheap spot capacity fits elsewhere: batch work, evaluations, synthetic data, internal tools, and any traffic you would be content to see fail over to a second provider when a node drops. Put it behind a router with a fallback, log which provider served each request, and keep it away from the paths where a customer is waiting or a regulator might ask.

Should staff run Darkbloom on a company laptop?

No, and the policy should say so before the first person asks. A company-issued Mac is a managed device under your security regime. Enrolling it in Darkbloom means installing a third-party agent through device management, opening a persistent outbound connection to a coordinator you do not control, running arbitrary inference jobs for unknown customers on hardware that also holds your data, and drawing power and fan noise around the clock. The privacy design protects the buyer's prompts from the machine owner. It says nothing about protecting the machine owner's employer from the software.

There is also the plain question of who owns the income from an asset the company paid for. Most acceptable-use policies already prohibit installing unapproved software and using company equipment for personal gain, and most managers have never had to enforce the second clause. Expect to. If the enthusiasm is real, the constructive answer is to point at the personal Mac at home, which is the machine Darkbloom's terms were written for anyway.

What does Darkbloom mean for inference prices?

It sets a floor under open-weight inference prices that no data centre operator can match on capital cost, and it will pull the price of running Qwen-class models down further than the frontier labs would like. We have written about frontier prices falling all year, and the driver has been competition among providers with broadly similar costs. A provider whose marginal hardware cost is zero changes the shape of that competition for the models it can serve, which today means anything that fits in 48 GB of Apple memory and tomorrow, as the bar drops and the fleet grows, means more.

The frontier models are safe from it, because nobody is running Fable 5.1 on a MacBook. The mid-market is not. A business planning its AI spend for 2027 should assume that the price of a good-enough open model served by someone else keeps falling, that the reliability and compliance story of the cheapest suppliers will lag their prices by a year or two, and that the value of owning a router that can move traffic between suppliers in an afternoon keeps going up. That last point was true before Darkbloom and it is the practical lesson of it.

Frequently asked

What is Darkbloom and how does it pay Mac owners? Darkbloom is a distributed inference network built by Eigen Labs that runs open-weight models such as Qwen 3.6, GPT-OSS 20B and Gemma 4 on privately owned Apple silicon Macs and sells access through an OpenAI-compatible API and OpenRouter. Owners install a command-line agent, enrol the machine, pass a throughput benchmark and are paid per token served through a connected Stripe account. As of late August 2026 the minimum requirement was 48 GB of memory, providers kept 100 per cent of revenue, and the network reported 4.5 billion tokens served and around 250 machines online.

How much can you earn renting out a Mac on Darkbloom? Reported figures range widely. The project's founder cited $120 to $200 a month per machine on average; Darkbloom's own calculator estimated about $37 a month for a Mac Studio with 96 GB; the published annual range is $192 to $4,983 after roughly $2 a month in electricity; and the top earners on the public leaderboard in late August averaged $15 to $26 a day. Earnings depend on network demand, which the project has said supply was outrunning, so social media claims of a guaranteed $200 a month should be discounted.

Is Darkbloom safe to use for business data? Darkbloom's design encrypts requests to a coordinator in a confidential virtual machine, re-encrypts them for one provider, and runs inference inside a hardened process on that Mac so the owner cannot read prompts or responses. The code is public and early reviews found nothing malicious. It is still an early-stage network with no service level agreement, no audit certification and no processor agreement of the kind data protection law expects, so it suits batch and internal workloads with no personal data, behind a router with a fallback, and not customer-facing or regulated traffic.

Related reading

• Qwen is now the open-model default. When an open model beats an API for your workload
• Frontier AI prices keep falling
• On-premise, sovereign cloud, or public cloud for AI: how to choose
• Agent permissions: how to decide what an AI agent may read, write and spend

A network of idle laptops undercutting the API market is a reminder that the cheapest inference and the most dependable inference are rarely the same supplier, and that a business needs the plumbing to use both. Building the routing, fallback and data controls that make a cheap supplier safe to try is part of what we do in a data platforms engagement.