By early August 2026, Cuba's national grid had collapsed six times this year, four of them within a single month. Millions of Cubans went without electricity for 18 hours or more a day, and in Havana blackouts ran past 22 hours. Storms disrupted restoration within a day of the sixth collapse, before technicians had finished bringing the grid back.
Two failures are compounding. Oil supply has been cut to almost nothing, and the generating fleet is too old and poorly maintained to run reliably on what fuel remains. AI cannot supply oil. It can help with the second failure, which is what turns each shortfall into a national blackout.
How an oil shortage and one plant took the grid down
Venezuelan crude, Cuba's main energy source for decades, fell to effectively zero in January 2026 after the United States cut off shipments and Mexico suspended planned exports. The grid needs 90,000 to 110,000 barrels of oil a day to run at full capacity, and domestic production covers about 40,000. The reported generation deficit peaked at 2,113 megawatts in mid-2026. On bad days, about 1,230 megawatts were available against demand of 3,250.
The Antonio Guiteras Power Plant, Cuba's largest thermoelectric station, has had at least 17 unscheduled outages in 2026, most traced to economiser faults and boiler water leaks. On a system already running short, losing Guiteras is often enough to trigger a nationwide collapse. The same plant, failing for the same mechanical reasons, keeps taking down a grid that serves about eleven million people.
Four tools that would reduce the damage
Each of these runs in grid operations somewhere today. They differ in cost and in how much new hardware they need.
| Tool | What it prevents | What it needs |
|---|---|---|
| Predictive maintenance at Guiteras | Unplanned trips of the largest plant | Plant sensor data and a model trained on its failure history |
| Cascade prediction and automatic islanding | One plant's failure spreading to the whole grid | Grid-wide sensors, fast switching equipment, capital investment |
| Priority load shedding | Hospitals and water pumps losing power in rotating cuts | A map of critical loads on each circuit and remote switching |
| Solar forecasting and balancing | New solar parks adding instability to a weak grid | Weather forecasts, output data from each park, storage or fast reserve |
1. Predictive maintenance for a plant that fails the same way
Economiser faults and boiler leaks give warning. Pressure drifts, temperatures move outside their usual range and vibration patterns change over days or weeks before the equipment fails. A machine learning model trained on the plant's own sensor history can flag that drift while the unit is still running.
A flagged fault becomes a planned shutdown. Operators can stage the repair crew, bring other units up and shed load in an orderly way before taking Guiteras offline, instead of losing it without warning. That does not make up for decades of underinvestment, but it would likely have turned some of this year's collapses into managed outages.
2. Cascade prediction and automatic islanding
Cuba's grid has little ability to contain a local failure. A cascade model runs on the grid's real layout and live load data, spots the moment a fault is about to spread, and splits the grid into self-supporting sections before it does. This is standard where the sensors and switching equipment exist. Cuba's grid lacks much of that equipment, so this is a capital project, not a software update. It is also the fix aimed most directly at how six collapses happened.
3. Load shedding that protects hospitals first
Current rationing cuts power by district for up to 22 hours, with no distinction between a residential street and a hospital ward on the same circuit. Cuba's health system has reported strain from the fuel shortage, and unplanned power loss at medical facilities is among the worst effects.
A priority system maps hospitals, water pumping stations, refrigerated medicine stores and telecoms relays onto the grid, forecasts supply and demand by the hour, and cuts load in the order that does least harm. Where critical sites share a circuit with homes, the practical first step is to move them to their own feeder or give them on-site backup.
4. Forecasting for the solar parks already being built
With Chinese financing and equipment, Cuba connected 49 solar parks between early 2025 and early 2026. That added more than 1,000 megawatts and raised solar from 5.8 percent of national generation at the start of 2025 to over 20 percent. China has committed to 92 parks and about 2,000 more megawatts by 2028. The US embargo restricts American technology, not Chinese panels, so the build-out has continued through the oil cut-off.
Solar output swings with cloud cover and drops to zero at night. A stable grid absorbs that; Cuba's cannot. Without forecasting and balancing across all the parks, adding solar to a grid that already fails when one plant trips can make it less stable. With forecasting, plus storage or fast reserve to cover the swings, the solar build-out is Cuba's most credible route away from oil.
The Core Argument
Cuba's oil shortage is a political problem no software can solve. The way that shortage becomes a national blackout, with one ageing plant failing unpredictably and taking the grid with it, is a problem predictive maintenance, cascade prediction and priority load shedding are designed for. The solar hardware is already being installed. What is missing is the control and forecasting layer that makes it work as a system.
The same weakness exists across the Caribbean
Puerto Rico's grid has been unstable since Hurricane Maria in 2017. Its Integrated Resource Plan aims to cut expected unserved energy from 154 hours a year to under 2.4 hours by 2032, relying heavily on better outage prediction and automation. Jamaica, the Dominican Republic and Haiti depend on imported fuel for much of their generation and face their own mix of hurricane exposure, single-supplier dependence and deferred maintenance.
Any grid where one plant carries a large share of national generation can fail the way Cuba's has. The four tools above apply to all of them.
Who could build and pay for this
For Cuba, the embargo rules out American vendors but not the technology. The Chinese, European and open-source channels supplying the solar hardware can supply sensors, models and control systems. The Ministry of Energy and Mines, which has coordinated the solar build-out with Chinese partners, is the obvious owner of an extension into grid controls.
For the rest of the region, CARICOM's energy bodies and the Caribbean Development Bank could fund a shared grid-resilience platform, built once and licensed to any member facing fuel shortages, storm damage or an ageing plant. Caribbean AI companies, now numbering dozens of active ventures, could build it.
What I cannot judge from outside is the state of Guiteras's own sensors and records. Predictive maintenance needs years of reliable readings from the plant, and if those were never collected or were lost, the model has little to learn from. The first months of any project there would go on instrumenting the plant, not on AI.
What to do next
- Unión Eléctrica engineers at Guiteras, first: gather every available sensor log and maintenance record for the economiser and boiler since the last major overhaul, and note where readings are missing.
- Cuba's Ministry of Energy and Mines: add a forecasting and balancing system to the specification for the next tranche of Chinese-financed solar parks, alongside the panels and inverters.
- Hospital and water utility managers in Cuba: list every critical site that shares a circuit with residential load and give the list to the utility, so those feeders can be separated or given backup first.
- Caribbean utilities with one dominant plant (in Jamaica, the Dominican Republic and elsewhere): run a study this year of what happens to grid frequency if that plant trips at peak demand, and whether current automatic load shedding contains it.
- The Caribbean Development Bank and CARICOM energy officials: scope a shared grid-resilience platform with two pilot utilities before the 2027 hurricane season.
Frequently Asked Questions
Why does losing one power plant black out a whole country?
Because supply and demand on a grid must match every second. When a large plant trips, grid frequency falls, and if other generators cannot make up the shortfall within seconds, their own protection systems disconnect them to avoid damage. On a small grid with few spare units, that chain reaction can reach every plant within seconds.
Doesn't every grid already have automatic load shedding?
Most do, in the form of under-frequency relays that disconnect blocks of customers when frequency falls below set points. These relays are fixed in advance and do not know what is on each circuit. AI-based schemes add forecasting and priorities, choosing which loads to drop based on current conditions, but they depend on the same switching equipment being in place and maintained.
What data does a predictive maintenance project need to start?
It needs time-stamped readings from the equipment, such as pressures, temperatures, flows and vibration, plus a record of past failures and repairs. Many plants already store readings in a control system historian even if nobody analyses them. Where sensors are missing on the parts that fail, installing them is the first step, and useful models usually need several months of data after that.
Can solar power alone keep Cuba's lights on at night?
No, not without storage. Solar parks produce nothing after sunset, so evening demand still falls on thermal plants, imported fuel or batteries. Solar can cut daytime fuel burn and leave more fuel for the evening peak, which is why forecasting and storage decide how much it helps.
Can Caribbean companies legally work on Cuba's grid?
Many can, but the rules depend on where a company is based and its links to the United States. The US embargo applies directly to US persons and can create risks for firms with US operations, banks or investors. Any Caribbean firm considering Cuban energy work should take legal advice before signing.