What can go wrong

Owning GPUs can lose money. This page lists the ways it can go wrong, what each would do to your rent or your capital, and what, if anything, limits the damage.

Read this first#

Market#

1. Demand falls and machines sit idle. An hour nobody rents earns nothing, and the fixed costs of that hour still come off. Tenants move to newer GPUs, cheaper clouds or their own hardware. Pools spread this across many machines, but they cannot remove it.
2. Rental prices fall. The price of a GPU-hour for a given chip has tended to fall as newer chips arrive and supply grows. Rental contracts are short, so a lower market price reaches your rent within weeks.
3. Your share is exposed to one kind of asset. Every machine in the pool is the same sled: eight NVIDIA L40S GPUs. A shift away from that hardware, or a sharp change in how AI models are trained and served, would hit all of them at once.

The same sled on a bad month#

Here is the day from Where the rent goes again, then the same sled with lower utilisation, then with a lower price. Space ($12.44 a day) and insurance ($4.15 a day) stay fixed. Power falls with load; at 50% utilisation it is assumed to be $13.00.

Utilisation falling from 80% to 50% cuts net rent by 46%. A price 22% lower cuts net rent by 32%. Net rent always falls faster than gross, because some costs do not fall at all.

The machines wear out#

A GPU is not a building. An L40S sled that cost $93,750 is written down over five years, about $18,750 a year or $51.37 a day. On the base day above, that is most of the $82.94 of net rent. Part of every payout is your capital coming back.

Machines are usually sold in their fourth or fifth year and the proceeds go to holders. What a four-year-old GPU will sell for is not known today and could be very little. The custodian sells retired machines through at least two independent hardware brokers and takes the higher bid. The sale and the price are published on the ledger, and the proceeds are paid out in USDC within 30 days.

Operations#

4. Hardware fails. GPUs, memory and power supplies fail. A failed GPU earns nothing until it is replaced, and a warranty return can take weeks. The maintenance reserve pays for repairs, and insurance covers larger damage, but neither pays the rent that was lost.
5. A site goes down. Power, cooling or network faults at a data center stop every machine in it. Insurance covers some outages. All of a rack is in one building, and the fleet is in two: IAD-4 holds 32 of the pool's 64 GPUs and DFW-2 holds the other 32. An outage at either site stops half the fleet until it is fixed.
6. A custodian fails. If a custodian goes bust or breaks its agreement, the machines remain the SPV's property. Moving them to another site costs money and weeks of downtime.

Technology#

7. A signing key is stolen. Someone holding a custodian's key could sign false records for that site. The key can be revoked in the Registry, after which its signatures are rejected, and the power cross-check makes invented usage stand out. Records signed before revocation would need to be reviewed.
8. A contract has a bug. The contracts were audited, which lowers the risk but does not remove it. A bug could lock or lose funds. See The audit.
9. USDC loses its peg. Rent, payouts and redemptions are all in USDC. If USDC is worth less than a dollar, so is your rent. USDC can also be frozen by its issuer.
10. Ethereum gets expensive. When gas is high, withdrawing small amounts can cost more than they are worth. Letting rent build up in your balance and withdrawing less often helps.

Structure#

11. Rules change. The law on owning shares of income-producing assets through blockchain contracts is still moving. New rules could limit who may hold shares, add costs, or force the pool to wind down and sell its machines.
12. You cannot always get out. Rent in your balance can be withdrawn at any time, but your capital cannot. Shares cannot be passed to anyone else, so the monthly redemption window is the way out, with 30 days' notice, and requests can be filled pro rata when cash is short.
13. The pool is concentrated. 8 sleds, 2 sites, 2 custodians and one GPU model, with half the fleet in each site. That is diversified compared with one sled, and concentrated compared with most funds.
14. CoresRent Labs fails. The machines and the pool's USDC belong to the SPV, not to CoresRent Labs, and the custodians keep running the sites. Finding tenants and running the platform would need a new manager. The SPV's independent directors can appoint a replacement under the servicing agreement, and a standby servicer has agreed terms to take over within 60 days.