put your gaming gpu to work: earn on every token it serves

August 4, 2026 5 min read ElephantPool team

If you own a decent GPU, you own a small piece of the most in-demand hardware on the planet — and it probably spends twenty hours a day doing nothing. That idle time is dormant capital: silicon you already paid for, sitting in a case, depreciating quietly while the world queues for inference capacity.

ElephantPool turns those idle hours into work. Your machine serves open-weight language models to real API traffic, and you get paid for every token it produces — 80% of the token revenue goes to the machine that did the work.

what a card can actually do

Let's stay honest and start from measured throughput, not marketing. In our benchmarks, a single Radeon RX 7900 XTX decodes at roughly 186 tokens per second on a quantized 1.5B model. Small models earn small money — their per-token price is low — but the same card holds its own on much larger models, where each token is worth considerably more. Serving bigger models, batching several requests at once, and staying online during demand peaks are what move the needle.

So how much is that in euros? It depends on three things we won't pretend to control: how much traffic the network routes to you, which models your VRAM can hold, and your local electricity price. A modern 24 GB card that stays online and utilized can plausibly earn a meaningful monthly amount; a card that's online two evenings a week will earn pocket change. Rather than quote a number here, use the earnings estimator on the homepage — it lets you set your own utilization and electricity assumptions and shows the net result, gross minus power.

getting started: three steps

1. Install mahout. mahout is our open-source host agent, written in Rust. One command on Linux x86_64:

curl -fsSL https://elephantpool.ai/install-mahout.sh | sh

2. Link your machine. Sign in at app.elephantpool.ai, grab the link code from your dashboard, and run mahout link <code>. mahout benchmarks your hardware so the network only routes work it can genuinely handle.

3. Go online. Run mahout up. Your card starts pulling model weights, joins the pool, and your dashboard starts showing live, receipt-backed earnings per token served.

amd and nvidia, equal citizens

Most GPU networks treat AMD as an afterthought. We didn't. NVIDIA cards run via CUDA and vLLM; AMD cards run via ROCm and vLLM, with a Vulkan/llama.cpp fallback for setups ROCm doesn't cover yet. Same features, same tiers, same 80% pay — from day one. Our own fleet includes AMD cards, so the ROCm path is exercised in production every single day, not once a quarter.

what about safety?

Reasonable question — you're letting strangers' workloads touch your hardware. Here's exactly what runs and what doesn't:

• Verified weights only. Every model file is checked against a published checksum before it loads. Your card runs the exact model everyone can inspect, nothing else.
• Sandboxed engines. Inference runs inside sandboxed containers with no network egress and read-only weights. Client code never executes on your machine — your GPU computes matrix math, full stop.
• Prompts never touch your disk. Requests and responses live in RAM for the duration of the job and are gone when it ends. You can't read your users' prompts, and neither can we.
• Your PC stays yours. mahout auto-pauses the moment you start using the machine — launch a game, and inference steps aside. One command (mahout down) takes you offline whenever you want.

And because mahout is open source, none of this is a trust exercise. Read the code.

the boring, important details

Every job your machine completes produces a signed receipt, and your balance accrues in a per-token ledger you can audit from the dashboard. Payouts run monthly by SEPA bank transfer once you've reached €50, processed by Stripe. Electricity is your cost — factor it in with the estimator before you commit a power-hungry card to 24/7 duty.

Your GPU is already plugged in. It might as well be working.