Using CoresRent
How metering works
Rent is only as honest as the meter. This page explains what is measured on each sled, how often, who signs it, and how a reading gets from a GPU to the public ledger.
Rent is only as honest as the number it is based on. At CoresRent that number comes from the hardware itself. The tenant does not report it and neither do we. This page explains what is measured, how often, and how a reading becomes a signed, public record.
What is measured
Each sled has eight GPUs. For every one of them, and for the sled as a whole, the meter takes one sample every 60 seconds.
| Reading | Source | Used for |
|---|---|---|
| Lease state per GPU | BMC, from the scheduler's allocation | GPU time billed to the tenant |
| GPU activity and power | NVIDIA DCGM on the host | Cross-check that a leased GPU is working |
| Sled power | Metered outlet on the rack PDU | Energy in Wh, the power line of the split |
| Health | BMC and DCGM (temperature, ECC errors, faults) | Maintenance, and flagging a faulty GPU |
Rent is charged for allocated time, the same way cloud GPUs are billed: a GPU counts as rented for a sample if it is assigned to a paying lease at that moment. Activity and power are recorded alongside so the two can be compared.
The meter module
The meter is a small module fitted to the sled's board at install. It reads the BMC and the host's DCGM over the board's management bus, and the PDU outlet over the site's management network. It keeps its own clock and buffers up to seven days of samples if its link to the custodian drops.
The module is sealed. The custodian photographs the tamper seal at install and at every visit, and the photos are linked from the machine's page. Replacing a module, or finding a broken seal, is logged in the Registry as a status change on that machine.
From sample to signed record
- Every 60 seconds, the module takes a sample. For each GPU: is it on a paying lease, is it drawing power, is it healthy. For the sled: watts at the PDU outlet.
- Each sample stands for 60 GPU-seconds per leased GPU. An hour has 60 samples, so a fully rented sled logs 8 × 3,600 = 28,800 GPU-seconds. That is the ceiling, and records above it are invalid.
- At the top of the hour, the samples are closed. The custodian's signing service adds up the hour: GPU-seconds, energy in Wh, the contracted rate and the gross rent that follows from them.
- The custodian signs the record. It signs EIP-712 typed data with a secp256k1 key that is listed for that site in the Registry contract. The record also carries the hash of the same machine's previous hour, so hours form an unbroken chain.
- The hour's records become one root. All records for the site and hour are hashed into a Merkle tree. The root is posted to the Ledger contract within 10 minutes of the hour.
- The full records are published. The records and their proofs go to IPFS, and the content ID is stored with the root. Anyone can fetch them and recompute the root. Verifying a record shows how.
Why usage cannot be inflated quietly
No single party can change a number without it showing. The tenant never reports usage. The custodian signs, but its signature is public and tied to one key, so a false record is evidence against it. And CoresRent cannot edit anything after it is signed:
| If someone tried to | It would show as |
|---|---|
| Change a number in a record | A signature that no longer matches the custodian's key |
| Delete or reorder an hour | A break in the prevHash chain for that machine |
| Swap a record after the root is posted | A Merkle proof that no longer reaches the on-chain root |
| Bill a GPU that is not working | A leased GPU with idle DCGM activity and low PDU power, flagged on the machine page |
| Sign with a key from another site | A signer that does not match the Registry entry for that site |
When the meter misses time
If the module misses samples, for example during a network fault it could not buffer, those minutes are not billed and not paid. The record shows it honestly: samples reads 57 instead of 60. A machine that is fully offline still gets a record for the hour, with zero samples, so the chain never skips.
This errs against owners, and that is deliberate. A meter should never bill time it did not see.
What the meter does not prove
A record proves that GPUs were leased and drawing power for a given number of seconds, at a given rate, as signed by the custodian. It does not prove the tenant paid. Tenants settle their invoices with the SPV, and each daily settlement lists both the gross rent in that day's records and the USDC actually received. Any gap is visible there.
Last updated 5 October 2026.