What is Darkbloom

A new topology for private inference.

Darkbloom routes encrypted requests to verified Apple Silicon hardware, at roughly half the cost of typical API providers. Prompts stay hidden from operators. Mac owners earn from compute they already own.

Run a provider on hardware you already have. Darkbloom matches your Mac to demand it can actually handle.

Join the grid ↗

Why does it cost less

Darkbloom turns idle capacity into a private inference market

Typical inference passes through clouds, resellers, and API layers, each adding margin. Darkbloom routes requests to idle Apple Silicon instead - hardware its owners already paid for, where the marginal cost is mostly electricity. Pay per token, with no subscription or minimum.

Decentralized Inference Pricing

Model & SpecificationInput (1M Tok)Output (1M Tok)Typical APIvs typical API
GPT-OSS 20BMoE · 128K context$0.0145$0.07$0.1450% lower
Gemma 4 26B128K context$0.05$0.25$0.9050% lower
Qwen 3.6 35B A3BQwen3.5 MoE VLM with inline MTP · 256K context$0.08$0.75$1.5550% lower
Gemma 4 26BMoE · 128K context$0.05$0.75--
Gemma 4 26B 8-bit rollback128K context$0.05$0.25--

Why privacy matters*

Sensitive information shouldn't run on untrusted hardware

Every prompt carries information you'd never publish: personal reflections, customer conversations, source code, strategy. Darkbloom's privacy is more than a policy. The coordinator can route requests, the providers can serve them, but neither can get a usable view of the prompt.

How privacy is ensured*

Operator blind by design

Darkbloom removes software paths an operator could use to observe inference data. Requests are encrypted, the OS is sealed, memory is isolated, and the process itself refuses inspection.

01 / E2E Encryptionencrypted before it leaves your device
02 / OS IntegritySIP enforced · signed system volume · binary self-hash
03 / Memory IsolationHypervisor.framework · Stage 2 page tables
04 / Hardened Processdebugger blocked · no shell access
Your inference dataprompts · responses · model state

operator is here, every path inward is eliminated

What powers the grid

Decentralized inference on Apple Silicon mac hardware

The grid runs on fleet of consumer Macs, M1 through M5. High-memory machines powering a range of open-source models and the router assigns each request to the most capable available machine automatically.

GPU CoresApple Silicon
CPU CoresP + E cores
Unified Ram

How much can providers earn

Earn from compute
you already own

Darkbloom pays you a base reward just for staying online, plus usage earnings on top every time your Mac serves a real inference request. Earnings scale based on your chip and memory - the more capable your machine, the more it can run.

Estimated monthly earning

$204.89/mo

$2,458.64/yr at 60% duty cycle

Mac model
MacBook Pro
Chip family
M4 Max (16-core)
Unified memory
128 GB

This estimate assumes a 60% duty cycle - the share of time your Mac is actively producing output tokens. Actual earnings may vary.

Models your Mac can run

Models are ranked by estimated monthly earning at the selected duty cycle.

Qwen3.6 35B A3B Best current estimateFits in your 128 GB (22 GB of model weights)$204.89/mo
Gamma 4 26B A4BFits in your 128 GB (18 GB of model weights)$182.40/mo
GPT-OSS 20BFits in your 128 GB (14 GB of model weights)$144.15/mo