There is one figure in the European data centre report CBRE published three days ago that is worth reading twice. In the first half of 2026, AI-specialist cloud providers — the neoclouds — signed 420 MW of colocation capacity in Europe. In the first half of 2025 they signed 89 MW. Nearly five times more in twelve months, in a market measured in decades of amortisation.
The tempting move here is the easy headline: "AI is eating your rack." It isn't true, or at least not yet, and certainly not in the facility where you keep four rack units. Something has changed, though, that does affect you and rarely gets discussed: the unit of measurement. For fifteen years, renting space in a data centre has been a conversation about square metres and rack units. Now it is a conversation about electrical power, and there are live contracts â many signed years ago, when the sums worked â still written in the old unit.
The numbers, and the two consultancies that disagree
Two different reports, published two weeks apart, look at the same European market and do not say the same thing. It is worth showing that as it is, because the industry habit is to take the scariest number and publish it without the footnote.
| Figure | CBRE (Q2 2026) | JLL (3 Aug 2026) |
|---|---|---|
| Colocation vacancy across FLAP-D | Just below 8% (down from 8.3%) | 6,4% |
| Frankfurt | 4.6% (wholesale) | 3,1% |
| Capacity delivered, Europe, Q2 | 685 MW | — |
| City leading the half-year | — | Paris, 72.5 MW |
| Where the growth is going | 66% of neocloud capacity, to the Nordics | "Milan, Iberia and the Nordics", not leaving the core |
Six point four per cent and eight per cent are not the same number. Each firm measures a different perimeter — which facilities count, what counts as available, what counts as wholesale — and neither publishes the full recipe. Our reading, offered as what it is, a reading: doubt the decimal, not the sign. Both point downwards, both have pointed downwards for years, and JLL puts its finger on the part that actually matters to whoever signs: "pre-leasing now defines occupier strategy. Committing well ahead of requirement is a necessity rather than a choice…"
What got expensive is not the rack: it is the grid connection
Here is, for us, the figure of the half-year, and the word "AI" appears nowhere in it. JLL puts prime powered land in the FLAP-D markets at €2.26 million per MW, against €1.24 million in 2021: 82% more in five years. It is not the square metre that went up. It is the square metre with current behind it.
And behind that rise sits a lead time, which is the part money does not fix. The same report describes Frankfurt with 311 MW in development and 687 MW planned, with delivery paced by connection lead times of 24 months or more; and in primary markets, for latency-bound workloads, it cites "lead times of up to 10 years" limiting new supply. Ten years. A grid timeline does not accelerate when you pay more: you plan for it or you lose it.
Why this is NOT your problem yet
Let us say it before anyone accuses us of selling fear to a client with six servers: none of these numbers describes your contract. FLAP-D means Frankfurt, London, Amsterdam, Paris and Dublin. These are wholesale deals of tens of MW between hyperscalers and developers. Your half-cage in a facility in Barcelona, Madrid or Bilbao does not trade in that market, does not compete for those MW, and is not renegotiated with those figures on the table. Anyone showing you JLL's 82% to justify a 30% rise on your two-rack invoice is playing games with scale.
What does travel downwards is something else, and it arrives late: the cost of powered land and the cost of a kWh are the input for everyone who builds, including the regional facility where you actually are. It does not compete for the same contracts, but it buys in the same electricity market and hires the same contractors. When its turn comes to expand a hall, that sum will be its sum. And there is a second, slower and more interesting effect: CBRE places 66% of neocloud capacity in the Nordics, and JLL describes capital "broadening into Milan, Iberia and the Nordics rather than leaving the core" — the qualifier is theirs and worth respecting: it describes growth spreading out, not relocating. Iberia sits on the good side of that sentence. For a Spanish client, the reasonable three-year reading is more local supply, built later and priced higher.
From U to kW: the six contract lines that now matter
This is the actionable part, and it depends on no consultancy. Take your hosting contract and look for these six things. If three of them are not written down, your contract measures in the old unit:
- 1Contracted power, in kW per rack, and what happens if you exceed it. This is the line that is almost never there. A 42U contract with a 3 kW feed runs out of electrical room long before it runs out of physical room: you can have twenty free U and still be unable to plug in the new chassis. If your contract only states units, you do not know what you bought.
- 2Flat rate or metered consumption, and who carries kWh variation. Two different businesses dressed as the same invoice. On a flat rate the provider carries the electricity risk and charges for it in the premium. Metered, you carry it. Either works, as long as you know which one you signed.
- 3A right to grow, with notice periods. Not "we'll talk when the time comes": how many additional kW you can request, with how much notice, and at what reference price. In a market where committing before you need it has become the norm, a written growth option is worth more than a discount.
- 4Density the aisle allows, not the rack. A single rack at 15 kW can be cooled almost anywhere. A whole row at 15 kW is a different facility. Ask about the hall, not your cabinet, because the thermal limit belongs to the neighbourhood.
- 5Relocation clause. Many contracts let the provider move you to another hall with notice. That is legitimate and sometimes unavoidable. What must be written down is who pays for the outage, who moves the hardware and which window is respected.
- 6Term length, seen for what it is: insurance with two faces. In a tight market a long term protects your price and ties your flexibility. A short one does the opposite. The answer depends on whether you know your three-year electrical draw or are guessing it — and that sum is yours, not the salesperson's.
"If they squeeze me, I'll move to the cloud"
It is a reasonable answer, and it has a plumbing problem. Public cloud comes out of the same building and the same substation. In CBRE's 2026 forecast, Europe reaches 13 GW of total supply: 8.7 GW of colocation and 4.3 GW of hyperscaler self-build. Those are two columns of the same market, competing for the same powered land and the same connection lead times. Moving to the cloud does not take you out of the physical constraint; it changes who invoices it to you, and in what unit.
We already did the honest comparison between hosting your own hardware and renting compute, with the maths on the table, in colocation versus public cloud over five years, and that conclusion does not change with these reports: it depends on the workload, not the fashion. And on the building as a single point of failure, in July we wrote about what a Google Cloud outage caused by power and cooling left behind: everyone's hardware, yours and the giants', ends up depending on an electrical panel and a chiller.
When you should do nothing about this
- ✗Do not sign five years out of fear of a real-estate report. If your draw has been flat for three years and your contract has a fixed price to 2028, the correct move is none. Committing to a long term only makes sense if you know what kW you will need; otherwise you are buying certainty with someone else's budget.
- ✗Do not buy density you will not use. Reserving 10 kW per rack "just in case" is paid every month even if you draw 2.5. A written growth option costs less than reserved power and does the same job.
- ✗And do not switch providers over a headline. Changing data centres costs outages, transport, readdressing and weeks of nerves. A tight market is exactly the wrong moment to move without a specific, measured reason.
What we do with this
First, the conflict of interest, which here is a double one and belongs on the table: everyWAN provides colocation and is also a tenant — we run our own hardware hosted in data centres and we have Proxmox VE with Ceph in production spread across several facilities. So we are neither writing this from the sidelines nor writing it without an interest: if you end up renting a rack, we could be the ones renting it to you. Read it knowing that.
That said, the criterion we apply to our own estate is the one we are proposing here, and it involves buying nothing. Plan capacity on the horizon supply actually moves on, which per the connection lead times these reports describe is twenty-four months, not twelve. Measure real draw per rack instead of adding up power-supply labels, which state the maximum a unit can deliver rather than what you are pulling. And keep contracted power in the inventory, next to addressing and the rest of the infrastructure, so it is a queryable field and not a PDF in a drawer; we use NetBox for that. On where a mid-sized company's hardware ends up living, we wrote a few weeks ago about what leaves the building not necessarily being the cloud.
In short
The neoclouds' 420 MW will not knock on your provider's door this year. The 82% rise in powered land does not reach you tomorrow either. But both reports describe the same thing through two different windows: the scarce resource has stopped being space and has become power, and the lead times for making new power are measured in years. If your colocation contract still speaks only in rack units, it is not badly drafted: it is drafted in last decade's unit. Changing that costs an afternoon reading the contract and one conversation with your provider, and it is better done now than on the day you need to plug something in. It is the same question we ask when designing any infrastructure: not how much space it takes, but how much it draws and how far it can grow.
Sources (checked on 22 Aug 2026): the 420 MW signed by neoclouds in H1 2026 against 89 MW in the same period of 2025, the forecast of 13 GW of European supply for 2026 (8.7 GW colocation and 4.3 GW hyperscaler self-build), the 66% of neocloud capacity in the Nordics, the 685 MW delivered in the quarter and FLAP-D colocation vacancy, in European Data Centres Figures Q2 2026 (CBRE) and its accompanying press release; the 6.4% FLAP-D vacancy and Frankfurt's 3.1%, powered land at €2.26M/MW against €1.24M in 2021 (+82%), Frankfurt's 311 MW in development and 687 MW planned, connection lead times of 24 months or more and up to 10 years in primary markets, Paris's 72.5 MW in the half-year, and the quotes on pre-leasing and on capital broadening into Milan, Iberia and the Nordics, in the EMEA data centre mid-year 2026 report (JLL, 3 August 2026). The two firms measure different perimeters and publish different vacancy figures; both are reproduced as published.
How many kW do you have contracted per rack?
If you had to go and find the contract to answer, that is the conversation. At everyWAN we host our own hardware in data centres and help size our clients': real power draw, growth horizon and what the small print says. Without selling you one rack too many, because the spare one is on your invoice every month.
Talk to everyWAN
Microsoft 365
ES 
CAT 
EN