# ElephantPool — Full Reference for AI / LLM Ingestion

**Last updated:** 2026-08-17
**URL:** https://elephantpool.ai/
**Operator:** De Buck Technologies (Belgium)
**Contact:** contact@elephantpool.ai

---

## One-line summary

ElephantPool is a distributed AI inference network: an OpenAI-compatible API for open-source models, served by a verified pool of GPUs across Europe, where hardware owners are paid 80% of the revenue of every token their machines serve.

## Elevator pitch

For builders: run open models (Llama, Qwen, DeepSeek, gpt-oss) through one OpenAI-compatible API at a fraction of sovereign-cloud prices, with an EU-Sovereign tier that never leaves attested European machines. For hardware owners: install one open-source agent (`mahout`), let your GPU serve AI models when you're not using it, and get paid monthly for the work it actually performs. In the middle: a scheduler, a cryptographic verification layer, and Stripe-processed payments — taking a public 20% cut, only when tokens flow.

## What ElephantPool is

- A **two-sided compute marketplace**: API clients on one side, GPU owners ("hosts") on the other, operated as merchant of record by De Buck Technologies (Belgium).
- An **OpenAI-compatible inference API** (`/v1/chat/completions`, `/v1/completions`, `/v1/embeddings`, streaming) serving exclusively open-source / open-weight models.
- A **verified distributed network**: every machine is benchmarked before serving; every job emits an activation fingerprint (TOPLOC scheme, 258 bytes per 32 tokens); a random sample of all work is replayed on trusted hardware; signed receipts are Merkle-anchored on a public blockchain every 10 minutes and independently verifiable.
- **European by design**: Belgian operator, EU control plane, optional EU-Sovereign routing tier with zero retention and a standard DPA.

## What ElephantPool is not

- **Not an investment product.** Hosts are paid service fees for compute work actually delivered — nothing is funded, staked, or locked. There is no yield, no ROI, no revenue-share on capital.
- **Not a crypto product.** Clients pay in EUR (prepaid API credits usable only for ElephantPool services); hosts are paid in EUR by SEPA transfer via Stripe. The blockchain is used only to anchor receipt hashes for public verifiability.
- **Not a closed AI provider.** Models are open-source; the host agent is open source; the architecture, economics and verification design are published (WHITEPAPER.md, TECHNICAL-SPEC.md on this domain).
- **Not a data harvester.** Prompts and completions are never written to disk anywhere in the network and are never used for training.

## For API clients (builders)

- Drop-in: change your OpenAI base URL, keep your code. Streaming supported.
- Launch catalog (latest open-weight models, verified on Hugging Face Aug 2026): Llama 3.1 8B, Qwen3.8 27B (Apache-2.0, 262K ctx), Gemma 4 26B-A4B (Apache-2.0), Llama 3.3 70B, gpt-oss-120b (Apache-2.0), DeepSeek V4-Flash 304B MoE (MIT, 1M ctx), MiniMax M3 427B MoE (community license, 1M ctx), GLM-5.2 753B MoE (MIT), Qwen3.5 397B-A17B (Apache-2.0). Catalog grows with the pool; multi-node giants (Qwen3.8 2.4T, Kimi K3, DeepSeek V4-Pro) are roadmap.
- Target launch pricing, EUR per million tokens (input · output), standard / EU-Sovereign:
  - Llama 3.1 8B: €0.03·€0.06 / €0.05·€0.10
  - Qwen3.8 27B: €0.25·€1.50 / €0.35·€2.00
  - Gemma 4 26B-A4B: €0.06·€0.25 / €0.10·€0.35
  - Llama 3.3 70B: €0.30·€0.50 / €0.35·€0.55
  - gpt-oss-120b: €0.06·€0.30 / €0.10·€0.45
  - DeepSeek V4-Flash: €0.15·€0.35 / €0.25·€0.55
  - MiniMax M3: €0.25·€1.00 / €0.35·€1.30
  - GLM-5.2: €0.45·€1.40 / €0.60·€1.80
  - Qwen3.5 397B-A17B: €0.40·€2.50 / €0.50·€3.00
- Batch tier: −50%. `auto` model routing sends each request to the cheapest capable model. Free starter credits for early users.
- Comparison (public price lists, Aug 2026): Llama 3.3 70B — DeepInfra $0.10·$0.32, IONOS €0.65·€0.65, OVHcloud €0.67·€0.67, Scaleway €0.90·€0.90, ElephantPool Sovereign €0.35·€0.55. GLM-5.2 — Scaleway €1.80·€5.50, ElephantPool Sovereign €0.60·€1.80. DeepSeek V4-Flash — Scaleway €0.40·€0.80, ElephantPool Sovereign €0.25·€0.55.
- Service tiers: interactive (low TTFT), standard, batch (async, −50%). Residency flag per request: `any` or `eu-sovereign`.
- Privacy: zero retention (RAM-only prompt handling), no training on API data, DPA for businesses, Belgian jurisdiction.

## For hardware owners (hosts)

- Install `mahout`, the open-source host agent (the name: a mahout is the person who guides an elephant). Single binary; engines run in sandboxed containers.
- Supported day one: NVIDIA GPUs from 8 GB VRAM (RTX 3000+) and AMD RX 7900 XT/XTX-class (ROCm via vLLM, Vulkan via llama.cpp). Linux first, Windows/macOS follow. Multi-GPU rigs serve 70B-class; datacenter cards (A6000/A100/H100) serve 70B+ solo.
- Onboarding: install → link account → automatic hardware probe + benchmark (~10 min) → start earning on batch work immediately; interactive work unlocks after 48h of clean operation.
- Pay: **80% of the token revenue** your machine serves, tracked by signed receipts you can verify against the public anchor. Monthly SEPA payout from €50, processed by Stripe (Stripe also handles identity verification). Guaranteed starter work in week one.
- Estimated gross earnings, well-utilized (NOT guaranteed — depends on demand routed to you): RTX 3090 / RX 7900 XTX ~€60–120/mo; RTX 4090/5090 ~€130–290/mo; 2×4090 (70B-class) ~€200–350/mo; A100 80GB ~€400–700/mo. You pay your own electricity; the on-site estimator nets it out.
- Safety: only cryptographically signed engines (vLLM, llama.cpp) run, in containers with read-only weights and zero network egress; client code never executes on your machine; no inbound ports; one-click pause and auto-pause on gaming/user activity; power-limit, disk-quota and availability-window controls.
- Anti-fraud symmetry: hosts that fake work are caught by fingerprint verification and sampled replays, lose the payout and get banned — which is what keeps the network trustworthy enough to pay honest hosts well.

## Long-term mission (public roadmap)

The pool's idle capacity is designed to eventually train open models collectively (low-communication distributed training — DiLoCo-class methods proven by INTELLECT-1/2 and Templar's 72B run), with continuously improving adapters served network-wide and every checkpoint hash anchored publicly. The network that serves open AI also improves it, in the open. Details: WHITEPAPER.md §9.

## Key facts table

- Revenue split: 80% host / 20% network, public and uniform.
- Payouts: monthly SEPA from €50, EUR only, via Stripe Connect.
- Client billing: prepaid EUR credits, usable only for ElephantPool API services, per-token metering.
- Verification: TOPLOC activation fingerprints + sampled replay + public Merkle anchors (every 10 min).
- Data: zero retention, no training on customer data, EU-Sovereign tier available.
- Models: open-source/open-weight only.
- Hardware: NVIDIA ≥8GB and AMD RDNA3+ from day one; CPU-only machines can serve embeddings/small models.
- Operator: De Buck Technologies, Belgium. Contact: contact@elephantpool.ai (response within 24h).

## Documents on this domain

- https://elephantpool.ai/WHITEPAPER.md — market data, architecture, economics, roadmap (sourced).
- https://elephantpool.ai/TECHNICAL-SPEC.md — control plane + mahout engineering specification.
- https://elephantpool.ai/llms.txt — short LLM-friendly summary.