Apply
Tell us your GPU, idle hours, and country. Application takes 60 seconds.
Atlas Network · Invite-only beta
Earn while you sleep.
Your GPU, when you don’t need it.
Atlas routes paying inference traffic to approved consumer GPUs. You install a desktop app, your hardware joins the network when you’re not using it, and we pay you per token served. Open-weight models only. Sandboxed. Auto-pauses when you need your GPU.
Estimated earnings
What could your GPU earn?
Drag the sliders. Numbers are best-effort estimates against current Atlas Network pay rates and realistic peak / off-peak utilization.
Estimated monthly earnings
$704/ mo net
Gross$723.17
Electricity−$19.58
Net$703.58
Based on 24,105,600 tokens/day blended across the optimal workload mix (~50% embeddings at high throughput, ~35% small/medium quantized chat, ~15% large quantized chat). Effective partner pay rate ~$1.00 per 1M tokens; per-call rates vary by model class. Real earnings depend on demand, uptime, and actual workload mix Atlas routes to you; the desktop app surfaces running earnings once Phase 2 ships.
The apply form’s GPU field is free-text; the calculator is here to set expectations, not to bind us. Real partner earnings will vary with demand and your duty cycle.
How it works
Four steps from apply to first dollar.
Get accepted
We review applications in tranches. KYC kicks in once you've earned $600 (US 1099 threshold) or the equivalent abroad.
Install the desktop app
One signed installer per OS. Tauri shell, bundled inference runtime. Hardware probe + auto-fit picks the right model for your VRAM.
Start earning
Daemon runs in your tray. Pauses when you game, when GPU is busy, when your laptop unplugs. Weekly payouts ≥ $20 via Stripe Connect / Wise.
Hardware compatibility
Does your GPU qualify?
If you have a recent NVIDIA, Apple Silicon, or AMD workstation card, almost certainly yes. The desktop app probes your hardware on first launch and auto-picks the best model for your VRAM.
NVIDIARTX 30 / 40 / 50 seriesCUDA. Sweet spot: RTX 4090 / 5090.
Apple SiliconM2 / M3 / M4 (incl. Pro / Max / Ultra)MLX runtime. Unified memory makes 70B Q4 trivially loadable.
AMDRX 7000+ workstationROCm where supported. Otherwise CPU fallback.
OtherAnything elseApply anyway — we'll let you know if your hardware qualifies.
Trust & privacy
What partners see — and don't.
The whole partner program is built on the assumption that customers must be able to trust who handles their prompts. Here's how we make that work.
Open-weight only
The Atlas Network only ever serves open-weight models (Llama, Qwen, Mistral, DeepSeek, Gemma, Phi). Closed-weight models from OpenAI, Anthropic, and Google never touch a partner GPU — those calls always go to managed providers.
Sandboxed runtime
The bundled inference runtime runs in its own OS user / process. It cannot read your files or open inbound network ports. Prompts are held in memory only; we sign a ToS commitment to no on-disk logging.
Customer opt-out
Newmen customers see the partner-eligible flag in their console and can disable Atlas Network routing for their org. Pay-as-you-go defaults that flag off (managed providers only); the Reliability Loop plan defaults it on because customers there have bound evaluators that catch any partner regression before it reaches them. Per-call opt-out via `forbid_atlas_network: true`.
Prohibited content
Newmen runs a content classifier on every prompt before it leaves the scheduler. CSAM, weapons synthesis, and other prohibited categories are blocked at the scheduler — they never reach a partner.
Quality verification
A sample of every partner's calls is mirror-run on a reference provider with temperature 0. Sustained drift below tolerance retires the partner. Per-model accuracy is published openly on the pricing page.
FAQ
The questions we get most.
What's the realistic income?+
Use the calculator above for an estimate based on your GPU and idle hours. Always-on hardware clears materially more than nights-and-weekends setups. As a rough guide: an M4 Mac mini running 16h/day lands around $300/mo; an M5 MacBook Pro Max always-on lands around $900/mo; an RTX 4090 always-on lands around $1,000/mo; an M3 Ultra always-on can clear $1,200/mo; an RTX 5090 always-on can clear $1,500/mo. The math assumes the optimal Atlas workload mix — embeddings dominate raw token throughput, small/medium quantized chat dominates per-token revenue. Real numbers depend on demand and the workload Atlas routes to you; the desktop app shows actual running earnings once it ships.
When does payment kick in?+
Weekly, with a $20 minimum balance. US partners get paid via Stripe Connect; international via Wise (broader country coverage and cheaper FX). Below the minimum, your balance carries forward.
What about taxes?+
You're responsible for reporting income in your jurisdiction. We file 1099-NECs for US partners earning ≥$600/year via Stripe; equivalent forms abroad via Wise. KYC and W-8BEN flows are triggered at those thresholds.
What happens when I'm gaming or my GPU is busy?+
The desktop app monitors GPU utilization and pauses inference automatically. You can pause manually from the tray, set quiet hours, or pause when on battery. We never preempt your own workloads.
Won't this wear out my hardware?+
Sustained inference at moderate utilization is gentler on consumer cards than gaming — lower thermal cycling, no constant frame-rate stress. That said, 24/7 100% utilization absolutely shortens lifespan. The default duty cycle is 50%; you can dial it lower in app settings.
What model weights do partners host?+
Newmen-curated quantized open-weight models. We host the GGUF / MLX weights on a CDN, content-hashed and signed; the desktop app verifies the hash before loading. You don't need a Hugging Face account — the app handles distribution.
Can I run my own model alongside?+
Yes. The Atlas Network runtime runs in its own sandbox and ports — your local Ollama / LM Studio / etc. setups are untouched. We pause Atlas inference when other GPU processes spike utilization.
What if I want to leave?+
Uninstall the app. Pending balance is paid out at the next payout cycle. We delete your partner record (subject to legally-required retention for tax purposes). No notice, no penalties.
Why isn't this fully open yet?+
We're at the start of Phase 2 of the cost-tier rollout. We accept applications now, build a queue, and roll the desktop app out in tranches starting once we hit 10 paying customers on the demand side. Following the queue order helps us calibrate the scheduler before opening the floodgates.
Apply to be in the first cohort.
Sixty-second form. We’ll email you the moment your application is reviewed and the moment the desktop app is ready for your tranche.