AI arrives as a customer: its demand fits the hours Spain is already throwing away

It was twenty past eleven on 2 June and the electricity price was heading for zero. David Muñoz, regional delegate at Naturgy, said as much from the stage. "It's 11:20. If we go onto OMIE now, well, we'll see that prices may already be close to zero or at zero." Until four in the afternoon, he added, we would probably even see negative prices. A voice from the panel pushed his forecast out to seven.

Then came the comparison. "Do you remember mortgages a few years ago? Negative Euribor. I didn't go and complain to the bank." "I think right now the abundance of energy shouldn't penalise us," he added. Where AI is concerned, the Spanish power system's problem is one of matching. The sun decides when power is generated, and the market spreads consumption across other hours. A data centre's demand is the closest thing to a solution that has come the system's way.

The arithmetic that makes a data centre the system's best customer

Muñoz made the case in four numbers. "3.8 gigawatts." "8,760 hours a year." Then the industry's ceilings, measured in equivalent full-load hours. The best wind farm he knows: "3,000 hours". The best solar PV plant: "2,000".

Those 3,000 and 2,000 hours describe the best site imaginable. The installed fleet delivers less. Divide 2025 generation by the year-end installed capacity published by Red Eléctrica, Spain's transmission system operator, and wind comes out at about 1,800 equivalent full-load hours and solar PV at around 1,200, against the 8,760 hours in a year. The real figures only strengthen Muñoz's argument, because no renewable technology in Spain can replicate that consumer's profile.

The 3.8 GW is the share that the proposed 2030 grid plan, presented in September 2025, sets aside for data centres. It sits within the 27.7 GW of new access the plan offers to demand on the transmission network, fourteen times what the plan now in force provided for.

The system needs a buyer like that. In 2025 Spain's day-ahead market logged 798 hours with the price at zero or below, according to data from Red Eléctrica and Pexapark. In 2026 the count stands at 906 as of 29 September, and the full-year 2025 total was passed on 19 July. Nobody publishes the complete series, so the 906 is our own count from the daily files of OMIE, the Iberian power market operator, taking every hour whose average price is at zero or below. Run on 2025, the same method reproduces the 798 to within one hour.

Curtailment has gone from anecdote to line item. In the first half, 2.5 TWh of renewable generation was not integrated into the system, 50% more than in the whole of 2024, according to an estimate by Aurora Energy Research. June was the peak, with 1.2 TWh curtailed by the system operator, and it was also the month in which solar PV took its largest share of the mix, at 29.4%.

The average is misleading too. From the same files, the average price so far in 2026 is €73.8/MWh, while the first twenty-nine days of September came in at €142.5/MWh, the most expensive stretch of the year. That average lumps together hours that have nothing to do with one another, but you pay for them one hour at a time.

Alezeia González, director of operations at Fortia Energía, which buys power for large electricity-intensive industry, put it this way. "Right now the problem we have in the electricity sector is that there's no more room for solar PV generation." Solar, she said, is a technology "heavily depreciated because there is so much solar PV penetration", and one that also generates all at once. "We've all heard about zero prices, negative prices in solar hours. So demand going up isn't such bad news," she concluded.

The bottleneck lies at the grid node and in flexibility

Demand is welcome. Where it can get in is another matter. González set out the two limits in order. The first: "that there is grid capacity to be able to connect so much demand at a single node". The second: flexibility, because cheap energy arrives when the sun is up and a data centre draws power around the clock. "It's not so much a question of resources as of the system's capacity and flexibility," she said, summing up.

The numbers from the connection queue bear out that order. Spain's Ministry for the Ecological Transition acknowledges in writing that data centres have been granted more than 6 GW of access capacity on the transmission network and around another 6 GW on distribution networks. In the same document it puts the computing capacity that the national Artificial Intelligence Strategy envisages for 2030 at 2.5 GW, or between 3.5 and 4 GW of electricity demand. The access granted is triple the planned deployment. What is scarce is firm capacity at the specific node where you want to put the machines.

González named the contradiction. A data centre, she said, "provides a supply with an SLA, with a quality of supply, it guarantees that it's not going to have any interruptions […] and yet when we talk about demand flexibility we're talking precisely about accepting that there may be supply interruptions or drops in consumption". She sees the way out in the contract. "We also have to innovate in the business model of data centres to be able to activate this flexibility."

Activating demand needs two-way communication with the system operator, and on the factory floor that signal reads as an intrusion. González voiced it as an industrialist would. "Who's going to come into my facility here and send me a signal that's going to set processes running? […] they're going to stop a cement mill on me."

The regulator reached the same conclusion. On 31 July 2026 the CNMC, Spain's competition and energy regulator, approved the flexible access permit regime, with effect from 1 September. New demand and storage will be able to connect in areas with no firm capacity, in exchange for accepting cuts when the system is tight. On the transmission network and above 1 MW, they must also agree to reduce power automatically when the operator asks, without affecting consumption that holds a firm access permit. It is the conditional capacity González described in June.

What to do about it if you develop, sell or buy energy

The value now lies in knowing how to operate capacity in the right quarter-hour. That changes three of your decisions.

The first is siting. Muñoz pointed to an undervalued asset: the generation plants already operating and scattered across the country, "with access points with substations, with lines that aren't in need of planning". Put your machines next to existing generation and you connect years before a new node arrives. "Artificial intelligence will increasingly go on site […] near the solar PV plant, near the nuclear plant, near the different power stations there will be small data centres," said Manuel Crespo, founding partner of Nodos.

The second is the contract. If your asset can shift load, you already have a product the system pays for: the system operator allocated 1,775 MW of active demand response for the second half of 2026. What you are selling is hourly shape. A supply contract that does not specify the hours in which you accept cuts leaves out its most valuable lever.

The third, data, is the one you budget for least. "This real-time data is gold when it comes to making decisions about energy management," said González, in markets "that are moving closer and closer to real time". Here AI's job is arithmetic. Matching forecast, price and shiftable load in fifteen-minute windows is a continuous calculation that cannot be kept up by hand.

Crespo provided the counterbalance. "The forecasts for artificial intelligence demand are absolute madness. I mean, you take the worst of the forecasts, the most pessimistic one, and there's no human, or physical, or material, or bureaucratic capacity, or anything to supply that much energy." The global numbers bear him out: the International Energy Agency puts data centre consumption worldwide at 485 TWh in 2025 and close to 950 in 2030. Crespo himself pointed to a way out. Generative AI took off in 2022, it is now 2026 and, he said, "I think we haven't reached the stage where they're really working on optimising the systems and making them more efficient".

Spain has spent twenty years learning to build megawatts. A megawatt is installed once and then produces whenever the sun pleases. The hour has to be matched every single day. That is where your margin lies: in knowing what gets switched on at two in the afternoon on a Sunday in May, and who pays for it.

RENMAD UsefulAI 2026 panel, Madrid, 2 June 2026. David Muñoz, regional delegate, Naturgy; Alezeia González, director of operations, Fortia Energía; Manuel Crespo, founding partner, Nodos; Álvaro Díaz del Río, founder and CEO, BluMind. Moderated by Raúl García, ASEALEN.