notes from the pool
Guides and deep dives from the team building Europe's distributed AI inference network — for GPU owners and for the builders on the other side of the API.
one model, every card: how a 27b runs on 8 gb
Qwen3.8 27B ships in five builds, from Q5_K_M on a 24 GB card down to IQ2_XXS on 8 GB. Why we split models from builds, what quantization actually costs, and how a quantized KV cache doubled our context overnight.
Read the article →we stopped projecting your earnings and started measuring them
Most GPU marketplaces advertise what a card could earn. Our dashboard shows what yours measured, and refuses to show a number until it has one. Here is why that choice cost us a nicer-looking page and was still right.
Read the article →prices should not need an engineer
Our price book used to be a file in a repository, which meant changing a price meant a deploy. It is now a dashboard control with guard rails that refuse an incoherent price — and the file is still the fallback you can return to.
Read the article →our website and our app were two products. now they are one.
The marketing site and the dashboard had different navbars, different footers and different ideas of what we look like. Merging them into one Aurelia app meant solving the problem every SPA has with search engines — and finding two bugs the moment we looked.
Read the article →the container that wasn't there
A friend with a 12 GB card asked what Docker images we'd install on his PC. The honest answer was none — and working out why took us through engines, image digests, registries and what a home GPU should really be asked to run.
Read the article →meet mahout: the open-source agent that puts your gpu in the pool
mahout is ElephantPool's open-source host agent: one small Rust binary that probes your hardware, verifies every download, serves open-weight models, signs a receipt for every job — and pays you 80% of token revenue. How it works, how to install it, how to contribute.
Read the article →why we pay gpu owners 80% — and how the math works
Every job on ElephantPool produces an ed25519-signed receipt, the ledger counts in micro-cents, and hosts keep 80% of token revenue. Here's why the split works and how monthly SEPA payouts happen.
Read the article →your prompts don't belong in virginia — the case for eu-sovereign inference
Prompts are some of the most sensitive data a company produces. CLOUD Act vs GDPR, Schrems II, and what an EU-Sovereign inference tier actually guarantees — attested EU machines, no transfers, standard DPA.
Read the article →ship with open-source ai: one api, the best open models of 2026
Qwen3.8, MiniMax M3, GLM-5.2, DeepSeek V4-Flash and Gemma 4 behind one OpenAI-compatible endpoint. Prepaid credits, spend caps, SSE streaming, transparent per-token pricing — 2–3× cheaper than US APIs.
Read the article →put your gaming gpu to work: earn on every token it serves
Your GPU sits idle most of the day. Install mahout, link your machine, and get paid for every token it serves on the ElephantPool network — NVIDIA and AMD, supported equally from day one.
Read the article →EU-only should be a routing rule, not a sentence on a website
You can now require EU processing per request, and the scheduler refuses to route it anywhere else. Here is how a machine's location is established — and what that proves.
Read the article →Your GPU, your hours
Availability windows, power caps and per-machine limits — set from the terminal or from your phone, with a sync rule that never silently loses an edit.
Read the article →Every token per second is money
Hosts are paid per token, so engine performance is not a benchmark score — it is the hourly rate. What we tune, why we measure instead of upgrading, and what it is worth.
Read the article →Your machine's address never leaves our infrastructure
EU-only routing needs to know where a machine is. Asking a geolocation API would have meant handing a third party the address of every host on the network. So we built the lookup ourselves.
Read the article →One model, nine prices
Everyone else quotes one price per model and never says which version answered. We price each quantization, prove which one ran, and let you set a floor.
Read the article →