6,329 MW granted to data centres and none in service

The figure has been public since this morning and nobody in IT is going to read it, because it ran in the energy press: of all the demand capacity Spain's grid operator has granted on the transmission network, data centres hold the largest slice — 6,329 MW, 34.2% — and have exactly zero of it in service.

We come at this from the boring side: we own hardware in a datacentre and we run our own network, so when somebody asks us for a colocation quote the conversation always ends up in the same unit, which is no longer rack units but kilowatts. That is why this number strikes us as the most interesting thing published this month, and it is not about the geopolitics of AI: it is about what you sign next year when you ask for four more kilowatts.

What exactly was published

Since February 2026 Spain's grid operator has kept a public map of access capacity for new demand on the transmission network, node by node, in PDF, CSV and XLSX. What it has added now, with data as of 1 October, is the breakdown by consumption type: seven categories, so for the first time you can see how much of that capacity is reserved for data centres and how much for factories, railways or hydrogen. The company itself describes it as "a voluntary transparency exercise", and that is exactly what it is: it obliges nobody to do anything, they have simply opened the drawer.

The table, with the two columns that matter. The first one is paper. The second one is copper.

* What the source publishes is the 18.52 GW granted and the 11.05 GW not yet operating. The 7.47 GW in service is our own subtraction.

The arithmetic that is not in the release

These are divisions, we did them ourselves and we flag them as such. If anyone wants to redo them, the two published starting numbers are 18.52 GW granted and 11.05 GW not yet operating.

The 0% is not an accusation: building a data centre and running a line to it takes years, and part of those 6,329 MW was granted recently. But it is worth saying what an access permit is: a right to connect at a specific node, not a finished build. Between holding the right and having voltage there is a substation, a line, civil works, transformers with long lead times and a commissioning. And in the meantime that capacity shows as occupied for anyone else who comes knocking at that same node.

"Available" is also a bookkeeping entry

Here is the part that caught our attention, and it takes reading two publications back to back. In February 2026, when the map launched, the grid operator summed things up by saying the transmission network was already 75% taken, with room for new demand at only one node in four. In the 1 October update there is margin left to grant at more than 35% of nodes. Seven months, ten points.

The grid operator does not attribute that movement to new construction but to a change in the calculation scenario: after revising the criteria set by the CNMC resolution of 1 December 2025 and accounting for how photovoltaic generation has evolved, firm demand capacity rises at 204 nodes, stays flat at 658 and falls at 75. At 204 nodes the headroom grew without a truck turning up, and at 75 it shrank for the same reason. If your business depended on one of those 75, you would have found out from a PDF.

Of the three figures in circulation — granted, available and in service — only the third is physical. The other two are the output of a criterion, and criteria get revised.

And what does this have to do with you and your three racks

Almost nothing, today. Quite a lot, the day you ask to grow. A while back we wrote that colocation is no longer negotiated in rack units but in kilowatts, and what this map implies is that a kilowatt now has two different prices depending on where it comes from: the one already in the room, or the one that arrives when the extension is energised.

The outage this produces is not spectacular, which is exactly why it is hard to see coming. It is not a blackout. It is that you ask to go from 4 to 8 kW per rack to fit two more virtualisation nodes, and you are told they can do 6 for now, or that you can have the 8 but in a rack four aisles away, or that you can have them down a single feed until the extension lands. That third answer is the dangerous one: A+B redundancy that sits at A for a few months, documented nowhere, and that nobody revisits once B finally arrives.

It is the same mechanism we described when writing about disk lead times: an administrative or logistical delay enters your repair time with the same weight as a technical fault, and it is almost never written into the plan. The fault is inevitable; the outage is not. The difference between the two is usually one line somebody should have asked for in writing.

The one-page annex

This is a proposal of ours, not an official template. It is not a questionnaire for the meeting: it is a sheet to staple to the contract, written as a supplier declaration and dated. In that format because a verbal answer during a sales visit does not survive a change of salesperson.

The supplier signs the first three lines. You sign the fourth, and it is the only one entirely in your hands.

What this figure is NOT

It is not a reason to move. If you have two or three racks, stable draw and no growth planned, this number is a curiosity and you should not touch anything; moving rooms is expensive, risky and comes with a cutover window that no news item justifies. Nor is it an argument for or against public cloud: the stay-or-go sum is still the five-year one, and a map does not change it.

And if what actually worries you is continuity, moving rooms is not the lever either. The lever is cheap and boring: keep the copy and the restart in another physical place, timed with a stopwatch, so that the room stops being an irreversible decision. That costs a fraction of renegotiating power and solves a bigger problem. Anyone selling you a move before asking where you restart if that room is gone is selling you the expensive part.

The limits of what we have just said

These figures cover the transmission network. Many mid-sized rooms and almost every corporate rack hang off the distribution network, which has its own queue and is not on this map; back in February the talk was of roughly 30 GW more in projects with a distribution access point. We give the railway and hydrogen figures as percentages because that is how the source broke them down. "In service" means the connection has been commissioned, not that it is being used: an energised data centre may be drawing a small fraction of what it has contracted. And the 6,329 MW are not a forecast of anything: some of those projects will get built, some will not, and this post does not know which is which.

The conflict of interest, up front: we sell colocation with our own hardware in a datacentre, so we have an obvious stake in this conversation existing. That is why the annex above is written so that any supplier can sign it, ourselves included, and why the four lines apply equally to one of our rooms and to somebody else's. If your current provider signs them without blinking, you do not need us for this.

Sources (verified on 1 October 2026): the breakdown by consumption type with data as of 1 October 2026 — data centres 6,329 MW (34.2%) and 0 MW in service; industrial 5,838 MW granted and 4,346 MW in service; railways 19.2%; hydrogen and renewable gases 13.2% with 27 MW in service; total 18.52 GW granted and 7.47 GW in service — and the movement in firm capacity (up at 204 nodes, flat at 658, down at 75, with margin at more than 35% of nodes) after revising the scenarios from the CNMC resolution of 1 December 2025 — El Periódico de la Energía, corroborated by Press Digital; the map's first publication (February 2026), the network being "75% taken" and the ~30 GW in projects with a distribution access point — El Periódico de la Energía, 20 February 2026; the tool and its calculation criteria — Red Eléctrica, "Conoce la capacidad de acceso". The 57%, 74%, 58% and 0% percentages are our own divisions over those figures, not published numbers. The four-line annex is a proposal of ours, not anybody's official template. A note on rounding: one of the two sources gives 6,328 MW and the other 6,329; we use the latter.