Skip to content
← Insights
Nov 14, 20264 min read

Count paid GPU hours before calculating the hardware payback

GPU availability and paid occupancy are different measurements. Check hardware eligibility, then use whole-system power and received payouts in the estimate.

A GPU server beside an electricity meter and a clock, with blank accounting sheets underneath.

A GPU listing earns money only when a customer pays for work under that listing's terms. Keeping the machine online makes it available; the paid hours determine how much of that availability becomes revenue.

A hosting estimate needs both observations. An advertised hourly rate with assumed full occupancy can make an idle server look profitable before anyone has rented it.

Check whether the hardware is eligible first

Different compute networks accept different systems. COCOON's current GPU-owner prerequisites describe a confidential-computing server with an Intel TDX-capable CPU and an NVIDIA GPU supporting confidential computing, identified as H100+ in the guide. It also specifies supporting system software.

That requirement rules out using an ordinary RTX 4090 gaming-rig example as a COCOON earnings forecast. A rate advertised for that card on another marketplace does not establish that the same machine can join COCOON.

Before doing the income calculation, match the actual machine to the actual service. Include the host system in the check. A compatible accelerator inside a server that lacks another required component may still fail the network's setup requirements.

Keep an eligibility result separate from a commercial result. Passing a hardware test establishes that the machine can meet that test. It leaves demand for its work unmeasured.

Availability and occupancy use different denominators

Vast.ai's host optimisation guide explains the relationship between price per hour and occupancy. Its marketplace also uses configuration and verification information that can affect whether a listing appears to a customer.

Availability describes the time a machine can accept the promised work. Occupancy describes how much of the relevant period it is rented. A host can have excellent availability while receiving few bookings. Calling a low booking rate uptime would conceal which problem needs attention.

Use one stated period for the calculation. For a 30-day month, there are 720 hours. If a hypothetical host is paid for 144 hours, its paid occupancy over that whole month is 20%. If it was offered for only part of the month, record that shorter availability period separately.

The measurement should come from completed rentals under the platform's billing rules. A busy GPU is not necessarily performing billable work. Your own maintenance task can consume the device while producing no hosting revenue.

Put every assumption beside the arithmetic

Consider an illustrative host receiving $0.30 per paid hour, with the amount defined here as revenue after marketplace commission. At 144 paid hours, the month's hosting revenue would be $43.20. These are scenario inputs, not observed market rates or an expected booking forecast.

Assume the complete host draws 300 watts during paid work and 100 watts for the remaining 576 hours. That produces 100.8 kWh of electricity use for the month. At an illustrative electricity tariff of $0.20 per kWh, the energy cost would be $20.16.

Subtracting that energy cost leaves $23.04. The remainder still has to cover any additional operating expense and the cost of owning the hardware. A meter reading for the whole system is a better input than an accelerator's advertised maximum power.

Change one assumption at a time to see where the estimate is fragile. A lower occupancy rate reduces paid revenue while the host may continue drawing idle power. A different tariff changes the cost of every available hour. A short promotional payment can raise one month's receipts without establishing a repeatable rate.

A platform balance has a payout process

Vast.ai's host-payout documentation describes payout methods and eligibility conditions. It explicitly says earnings depend on hardware, availability and customer demand. Its market statistics also have host-account access requirements.

A displayed balance and a received payout occur at different steps. Record the settlement period and any charges that apply to your actual route. A business estimate should use funds you can receive under those conditions.

If the service pays in a token, record the token amount first. Its eventual value in the currency used for your bills depends on the conversion you actually make. Keeping both units prevents a change in token price from being mistaken for a change in compute demand.

Promotional rewards need a separate line in the record. A reward for joining the network can be real income while answering little about what a customer will pay for repeated inference jobs. Do not merge the two into one utilisation figure.

Use an operating record before scaling the estimate

A trial with suitable existing hardware can measure paid work without assuming that every offered hour will sell. Retain the meter readings with the payout record. Note periods when the host was unavailable so a lower payment has an explainable denominator.

A short successful trial can justify a longer observation. It cannot establish a full hardware lifetime or future demand. Before expanding the system, inspect whether the same customers and terms could support the added capacity.

The useful output is a ledger that connects eligible hardware to billable work and received payment. Once those measurements exist, the hosting estimate can be revised with evidence instead of a borrowed hourly price.

GPU computingInfrastructureBusiness