What customers will pay for
A customer can buy space for its servers, a way to connect them, or a service that runs on yours. Decide who owns the equipment before you quote. The equipment owner pays for replacements and carries the risk of idle servers. The site operator still pays to keep unused space and power available.
| Service | How you bill | Cost to price into the contract |
|---|---|---|
| Powered rack space | Monthly fee for space and reserved power. State whether electricity is included or billed separately. | Cooling, backup power and capacity you must keep available. |
| Private connections | Installation charge and a monthly fee for each link to a carrier, cloud provider or another customer in the building. | Fibre, network equipment and the cost of keeping a second route available. |
| Backups and disaster recovery | Monthly fee for storing a copy, running recovery drills and keeping enough spare capacity to take over when the main site fails. | A separate site, data transfer, on-call staff and credits if recovery takes too long. |
| GPU rental | Charge by the hour for use of a graphics processing unit (GPU), or sell a block of hours the customer commits to buy. | Chips, power while they sit idle, repairs and replacement. |
| Wholesale data hall | Take a multi-year commitment from an operator or cloud provider for a dedicated hall and a specified amount of power. Agree when billing begins if capacity opens in phases. | Upfront fit-out, power reserved for one tenant, the effect of that tenant leaving and the cost to refit the hall. |
| Managed AI service | Charge for work the customer can count, such as documents processed. Use a monthly minimum if you must reserve capacity. | Cost per completed job, including repeat runs, review and support. |
Equinix reports separate revenue lines for powered space, private connections and managed infrastructure. Its 2025 filing says customers are generally billed “on a fixed and recurring basis each month for the duration of their contract.” The required investment depends on the service. You can sell managed backups from rented space. A dedicated hall needs enough power and cooling for the tenant before it brings in its first server. Digital Realty’s 2025 filing shows how an operator can track minimum payments from leases separately from the repayment of operating expenses. On a large lease, check how the customer’s minimum payment covers the capital you commit to its hall.
Equinix 2025 Form 10-K, pp. 50–51 and Note 18 · Digital Realty 2025 Form 10-K, Note 3
Match the service to the market
In smaller Caribbean markets, start with a service that a local organisation already buys. Digicel lists rack rental in Barbados and offers colocation, cloud backups and disaster recovery to customers in the Cayman Islands. In Jamaica, the Urban Development Corporation reported using Digicel’s Caymanas facility for off-site backups in 2011–12. eGov Jamaica reported operating its own hosting service in 2024–25. An operator could lease space and sell tested recovery to banks, hospitals or hotel groups. Before choosing a second location, check whether a storm, cable fault or shared carrier could cut off both sites at once. Confirm where the customer allows copies of its data to be stored.
Digicel Barbados colocation · Digicel Cayman Islands business services · Jamaica UDC annual report, p. 11 · eGov Jamaica annual report, p. 17
In a network hub, proximity to other tenants can be worth money. Equinix’s SP4 site in the São Paulo area sells cabinets and private cages alongside internet exchange access and direct links between networks. Equinix also offers these direct links at its Singapore sites. A new operator selling links needs to show which carriers and cloud services are reachable, how much the connection costs and whether both routes share the same duct. A building close to a cable landing station still needs agreements with the networks that use it.
Equinix SP4 service list · Equinix cross-connect locations, p. 4
For a GPU cluster serving international buyers, sell blocks of time you can deliver. Amazon EC2 lets customers reserve GPU capacity for specific dates and charges for the block up front. That matters when a customer needs many chips together to finish a job on time. A smaller operator can apply the same pricing idea to a few paid batches of research or model training. Put the chip type, memory, network speed, storage, start time and rescheduling terms in the quote. Ask the customer to run a sample job before committing to the full order.
AWS Capacity Blocks service · AWS reservation and billing terms
These offerings show what operators sell in different places. To estimate revenue for your site, get local power and fibre quotes and ask buyers what they would pay.
Price an empty rack before you sell a full one
If you reserve 50 kW for a customer, the power plant and cooling system must be ready for that load. The contract should state the monthly capacity charge, how electricity is metered and when the bill changes if fuel costs rise. When power is billed at cost, that part of the bill pays the supplier. Count the margin on service fees separately. If you pay suppliers in US dollars and bill customers in local currency, specify when you can adjust those prices. Run a lower-sales case and a higher-cost case before signing.
Recovery has a different bill. Ask the customer how much data it can afford to lose and how long it can leave its system offline. Then price a recovery drill with those limits. The drill should test that the second site can take over when the first site and its main network connection fail together.
How many GPU-hours do you need to sell?
Say you install 64 GPUs and charge US$3 for one hour on one chip. Assume each billable hour costs US$0.45 to deliver. Budget US$750,000 a year for fixed costs, including idle power and an allowance to replace chips. This is a worked example. Every input is an assumption.
How paid use changes the annual result
| Share of theoretical hours sold | Revenue | Balance after assumed costs |
|---|---|---|
| 25% | US$420,480 | −US$392,592 |
| 50% | US$840,960 | −US$35,184 |
| 75% | US$1,261,440 | US$322,224 |
Each paid hour leaves US$2.55 for fixed costs. Divide US$750,000 by US$2.55 and you need about 294,118 paid hours a year, or 52.5% of the theoretical total. Cut the price to US$2.40 and break-even rises to about 68.6%. Compare your offer with published cloud prices for the same chip, memory and contract term, then check what regional buyers will commit to.
Lambda public GPU price list · Jamaica study, GPU price stress test
Give the buyer a test it can pay for
Pick one service and one buyer. For an insurer that needs disaster recovery, use approved test data to run a restore at a second site. Measure how long it takes the system to come back, how much data it loses and what the operator bills. A university with a queue of research jobs needs a different test. Rent capacity for a known batch of jobs and measure the cost and time to finish them.
If the test works, ask for a written order that states the monthly minimum, service hours and who pays when the cost of power or data transfer changes. Use that minimum to price the first phase. Compare the cost of leasing space, buying compute as needed and building your own facility, where each option can fulfil the order. Use the same availability and service terms in every quote.
Run your first customer’s numbers
Use the customer’s minimum order and your supplier quotes as the starting case. Then test what happens if the next customer takes longer to sign, the price falls or your largest buyer leaves. For a mixed rack and GPU project, the Portmore calculator lets you enter your own prices, paid use, power costs and staffing. The study’s preset values describe one Jamaican case. Replace them with inputs for your market.
Model your own prices and costs