why we pay gpu owners 80% — and how the math works

August 18, 2026 4 min read ElephantPool team

Marketplaces love vague revenue-share language: "competitive rates", "up to", asterisks pointing at asterisks. We'd rather show the machinery. This post explains exactly how a token served on your GPU becomes euros in your bank account — and why the split is 80/20 in your favor.

every job leaves a signed receipt

When your machine completes an inference job, it doesn't just report "done". It produces a receipt signed with an ed25519 key: which model ran, how many input and output tokens were processed, when, and at what price-book rate. The scheduler countersigns its side. That receipt — not a log line in our database, not a number we assert — is the unit of accounting.

Because receipts are cryptographically signed by both parties, neither side can quietly rewrite history. Your dashboard's earnings figure is the sum of your receipts, and you can audit every one of them. If we ever displayed a number your receipts don't support, you could prove it.

a ledger that counts in micro-cents

Per-token prices are tiny — fractions of a cent per thousand tokens — so rounding at job granularity would systematically leak value in someone's direction. The ledger therefore accrues in micro-cents (millionths of a euro cent). A 340-token completion on a small model credits you exactly what it earned, down to the last micro-cent, and rounding happens once: at payout, in your favor's worst case by less than a cent per month.

why 80/20 is the honest split

The logic is about who carries which costs. In a classical cloud, the provider owns the building, the hardware, the power contract, the cooling and the ops team — so the provider keeps essentially all of the revenue. On ElephantPool, that stack is inverted:

• You bring the capital. The GPU, the rig around it, the depreciation — already yours.
• You pay the operating cost. Electricity and the internet line are on your meter.
• We run the network. Scheduling, verification, the API gateway, model distribution, billing, fraud defense, and payment processing — that's the 20%.

Since the host carries both capex and opex of the compute itself, the majority of the revenue belongs to the host. 80/20 isn't generosity; it's cost accounting. And the split is public and uniform — no negotiated side-deals, no house hardware paid differently, no spread hidden between what builders pay and what the price book says.

compare that with the classical cloud

Run the same numbers a datacenter does. A hyperscaler rents a GPU-hour at a multiple of its amortized cost — margins on GPU compute are notoriously the healthiest line in their earnings. The gap between what an API customer pays and what the silicon costs to run is the datacenter margin. ElephantPool's model deletes most of that layer: builders pay roughly 2–3× less than US API prices, and the machine that did the work keeps 80% of what remains. Both sides of the marketplace pocket the deleted margin.

from ledger to bank account

Once a month, your accrued balance is settled — by SEPA bank transfer, from €50, processed by Stripe, which also handles the identity verification that keeps the pool clean. Below €50 the balance simply rolls over; nothing expires. No crypto wallet, no gift-card points, no withdrawal fee schedule: a bank transfer, in euros, with a statement line you can hand to your accountant.

the point of all this

A distributed pool only works if thousands of strangers trust the accounting without trusting each other. Signed receipts, a micro-cent ledger, one public split, and boring euro bank transfers are how we make trust unnecessary. The math isn't exciting — that's the feature.