Adrian Dunkley
Research newsletter · Jamaica

What would a data centre bring to Portmore?

Adrian Dunkley and Nicholas Dunkley
Maestro AI Labs · 19 November 2026

A two-megawatt centre could support about eight local full-time jobs under our assumptions. Recovering its cost would take customers willing to pay more than our starting price.

Test your own assumptions
Illustration of the server hall, cooling units and service road
Construction time-lapseOpen for service

The completed scene shows an example layout in Portmore.

Who proposed Portmore?

St Catherine East Central MP Alando Terrelonge asked the Government to place a technology park and AI data centres in Portmore. Cabinet authorised an exploratory agreement with overseas investors. No buyer, parcel, construction budget or operating capacity had been confirmed in the public record we reviewed by 19 November.

On 18 November, Information Minister Dr Dana Morris Dixon set out two conditions. “We’ve also said as a Cabinet that it cannot be pulling from the grid,” she said. The project must also have no impact on Jamaica’s water systems. A developer must show how it can meet both conditions through a dry season and a storm.

Terrelonge’s proposal · Post-Cabinet briefing coverage

There is a local customer to learn from. The Urban Development Corporation used Digicel’s Caymanas centre for off-site backups. A public buyer also sought co-location space in 2026. Those records show a use for local hosting. Neither tells us what a customer would pay to train an AI model in Portmore.

UDC annual report, 2011–12 · ICT Authority procurement record

What the operator would have to earn

We modelled two megawatts (MW) of computer capacity. The operator rents rack space, runs existing AI models, trains models, and reserves some space for public-sector customers. The model sells 72% of capacity once the centre reaches normal use in its fourth operating year. All money figures use constant 2026 US dollars.

$50.6mInitial construction, site and computer equipment
−$27.9mNet present value at a 10% real discount rate
6.22%Share of 10,000 trials with a positive net present value

In operating year four the centre bills $13.4 million and spends $7.1 million on operations. That surplus has to cover construction and two rounds of computer replacement. At a 10% required return, discounted cash receipts fall $27.9 million short. At an illustrative J$160 per US dollar, the initial investment is about J$8.1 billion. The simple return on initial and replacement capital is −7.9% over the model period. The project does not repay its investment within 15 operating years.

A graphics processing unit (GPU) is a chip used to run many calculations at once. A GPU-hour buys one hour on one chip. The base case starts at $2.65 per GPU-hour for running an existing model. Holding the other terms fixed, that starting price has to reach $3.85 to break even. The training price moves with it at 82% of the rate. At the starting prices, even 100% mature use leaves a $5.24 million gap in present value. Raising all service prices by 25% moves break-even to 79.65% mature use. A signed agreement would need to cover volume and term as well as price.

How often did a trial recover its cost?

Number of trials per $5 million band · horizontal axis in US$ millions of net present value
0300600900-100 to -95 million USD: 2 cases-95 to -90 million USD: 14 cases-90 to -85 million USD: 36 cases-85 to -80 million USD: 72 cases-80 to -75 million USD: 144 cases-75 to -70 million USD: 250 cases-70 to -65 million USD: 350 cases-65 to -60 million USD: 533 cases-60 to -55 million USD: 618 cases-55 to -50 million USD: 788 cases-50 to -45 million USD: 854 cases-45 to -40 million USD: 873 cases-40 to -35 million USD: 876 cases-35 to -30 million USD: 810 cases-30 to -25 million USD: 761 cases-25 to -20 million USD: 635 cases-20 to -15 million USD: 583 cases-15 to -10 million USD: 492 cases-10 to -5 million USD: 397 cases-5 to 0 million USD: 290 cases0 to 5 million USD: 216 cases5 to 10 million USD: 142 cases10 to 15 million USD: 106 cases15 to 20 million USD: 67 cases20 to 25 million USD: 50 cases25 to 30 million USD: 23 cases30 to 35 million USD: 6 cases35 to 40 million USD: 9 cases40 to 45 million USD: 3 cases-100-75-50-25025Break-even
The middle 90% of modelled values run from −$70.4m to +$2.5m. Half fall below −$37.3m. The distributions describe the assumptions in this study. They do not give measured odds for a real Portmore project.
Compare the five service choices
2 MW serviceYear-four revenueNet present value
Mixed workload$13.4m−$27.9m
Rack space$5.3m−$32.4m
Public-sector rack space$6.0m−$28.2m
Running existing AI models$22.6m−$17.1m
Training AI models$18.5m−$40.7m

Rack-space customers supply their own computer equipment. The operator buys GPU systems for the AI services. Prices and paid use are assumptions.

Who supplies the power and water?

The dry-cooled site needs 2.76 MW at full load. It uses 18.73 gigawatt-hours (GWh) in a normal year at the modelled level of sales. The starting electricity price is $0.23 per kilowatt-hour for reliable, independent supply, including the cost of the power plant. We have no supplier offer at that price.

A separate test paired 14 MW of solar panels with a battery that could store 34.2 megawatt-hours. We then held the site at full load for 72 hours and cut sunlight to 10% of its usual modelled level. The battery began with 19.35 megawatt-hours. The site still needed 162.5 megawatt-hours from another independent source. That source had to supply up to 2.76 MW. A solar array that produces enough electricity over a year can still leave the servers without power on a particular night.

Power through three days of low sunlight

Hourly solar, storage and additional firm supply in a constructed storm test
Hourly dispatch chart. A 72-hour low-sunlight test needs backup on 66 hours and up to 2.76 megawatts.
The test uses a modelled daylight pattern. A project engineer would need hourly site weather, measured loads, a fuel plan, maintenance records and a backup test.

Dry cooling moves heat into the air. A hybrid system uses some water to assist dry cooling. An evaporative tower uses more water to carry heat away, which reduces electricity use in this model. Water for staff and site services is included in every row.

Cooling systemAnnual site electricityAnnual site waterNet present value
Dry18.73 GWh570 m³−$27.9m
Hybrid17.64 GWh4,912 m³−$27.0m
Evaporative16.83 GWh20,517 m³−$24.4m

The dry case still uses about 1,560 litres of water a day on average. Its assumed power supplier emits 0.45 kg of carbon dioxide per kilowatt-hour, or roughly 8,427 tonnes a year for the site. This excludes construction and off-site water use. Require a source agreement, water balance and discharge plan for each design. Test how much extra power dry coolers need on a hot, humid day.

Jamaica Information Service on drought supply · Operating solar plant comparator

What residents would gain and carry

The model calls for 13 full-time equivalent operating jobs. If Portmore residents fill 65% of those hours, their share is 8.45 full-time jobs. To reach a target of 20 resident jobs at that hiring rate, employers would need at least 31 full-time posts. A contractor who works on the hall for six months accounts for half a construction job-year. The build supports about 128 job-years in total, of which 77 go to residents under a separate 60% hiring assumption.

At an assumed $36,000 annual wage, gross resident payroll comes to $304,200 a year before deductions. If those workers could already earn 80% of that amount elsewhere, the additional gross earnings would be $60,840. Neither figure is council revenue. Any proposal for a public grant should show which revenues the council can collect and which new services it must fund.

The 2022 census counted 57,927 people in Greater Portmore and 13,479 in Waterford. These are community counts. They tell us who lives near a project, not how many people have the skills or want the work. Ask employers to specify roles, training, apprenticeships and payroll by place of residence.

STATIN community census counts · Portmore Community College programmes

Portmore sits on a coastal floodplain. During Hurricane Ivan in 2004, researchers reported surge reaching 3.3 metres on parts of the Hellshire and Port Henderson–Fort Augusta coasts. During Beryl in 2024, crews had to clear roads and repair lines even where the initial assessment reported limited damage. A dry server hall still loses service if fuel, staff or fibre cannot reach it.

Before a land commitment, survey the building site, drainage outfall and road elevation. Test flooding from the river, heavy rain and the sea together. Review future sea level, wind, heat and drought. The historical coastal observation cannot set a floor height for an unknown parcel.

Climate Studies Group Mona, Portmore case study · Government storm surge map · Beryl reporting from Portmore

The next investment should buy evidence

Start with a paid trial on existing infrastructure. Invite a bank, a public agency or a regional customer to buy a defined service. Record the hours used, final invoice, response time and recovery time. A leased trial gives Portmore a way to test demand before it buys land or equipment.

Stress caseNet present value
Anchor contracts at premium rates$11.2m
Firm power cost rises to $0.35/kWh−$42.4m
30% capital overrun and one-year opening delay−$51.5m
Year-four storm. 90 extra days lost and 30% shell repair−$33.3m
Combined delay, low demand, storm, high power price, hot operation−$93.0m

The premium-contract case clears the required return. It assumes 90% paid use and a 25% increase in all service prices. A letter of interest cannot establish either figure. In a separate buyer comparison, a 2 MW service running existing models at 25% use costs less to rent than to own when the initial rental price is below $7.55 per GPU-hour. At 72% use, that threshold falls to $2.88. Those thresholds compare equal quantities of computing service with the same start date.

What would change the recommendation?

Bring signed customer volumes and prices, an independent power bid, a water plan that meets Cabinet’s condition, parcel-level engineering and a resident hiring plan. Rerun the model with those terms. Compare the result with a lease and with leaving the site undeveloped.

Open the Portmore calculator ↗

Assumptions, sources and papers

The model tracks two construction payments followed by 15 operating years. It excludes corporate tax, debt service and resale value. Net present value discounts each year’s cash flow to the investment date. Return on investment divides cumulative net cash by initial and replacement capital. The calculator uses the same code as the papers.

Reference case inputs
InputStarting assumption
Workload by installed computer capacity40% commercial rack space, 35% inference, 15% training, 10% public-sector rack space
Use and pricing72% mature use. Commercial and public rack space earn $310 and $350 per kW-month. Inference starts at $2.65/GPU-hour. GPU prices fall 1.8% a year in real terms.
Construction and equipmentBase facility cost $10.7m per IT MW before cooling and site allowances. Each installed GPU system share costs $32,000 per GPU. Replacement in years 5 and 10 costs 70% of the initial GPU price each time.
Power and coolingDry cooling. Facility energy is 1.38 times computer energy. Process water use is 0.03 litres per kWh of computer energy. Independent power is $0.23/kWh. Water is $4/m³.
Hiring and output65% local share of operating jobs. 60% local share of construction job-years. Output is sold GPU-hours and rented kW of rack space.
Risk trials10,000 draws vary paid use, service prices, capital cost, power price, future GPU rates and disruption. Sales and use move together. Costs and energy prices also move together.
Storm lossesIllustrative annual event chance 10%. Each event removes seven days of paid service and requires repairs worth 1.5% of facility shell cost. No parcel-level flood probability is available.
Model checksIndependent arithmetic checks; repeat runs at 1,000–20,000 draws; 300 bootstrap resamples of the 10,000-draw result. The resampling measures simulation noise. It cannot validate commercial assumptions.
Evidence and further reading

Read or reproduce the study