The Register reported on 19 October 2026 that Microsoft has reportedly abandoned plans to anchor Nscale’s proposed Loughton AI supercomputer in Essex, after learning the site might not get enough grid power until the early to mid-2030s. The Telegraph first reported the withdrawal on Sunday 18 October. For anyone buying datacenter power capacity, the lesson is not about one UK site: a megawatt in an announcement and a megawatt you can draw are different things.

Start with what is not confirmed. Neither Microsoft nor UK Power Networks (UKPN) has stated the reason for the exit, and the grid cause comes from press reporting. Nscale says it remains committed, is exploring alternative power sources, and that the “London-wide grid issue” does not affect its other contracts.

UKPN told The Register it cannot discuss individual connections. It did say that many datacenters can connect to its distribution network but are “awaiting works on the national transmission network” it does not own. That is a general statement, not a Loughton confirmation, but it describes a pattern: the party selling you space may not control the thing your date depends on.

The numbers around the project show how loose “megawatts” can be. The September 2025 plan was 23,040 Nvidia GB300 GPUs in a 50MW facility, expandable to 90MW, with first chips expected in early 2027. Sources disagree on which figure applies, and none we read shows 90MW as secured, so treat it as a plan. The Register also reports that in March the site was still a scaffolding yard.

The queue itself is inflated. Omdia’s Alan Howard describes developers securing several land parcels and requesting a grid connection for each, to see what gets approved. Eurelectric has said speculative projects slow connections in Italy and Spain. Grid connection delays follow from a queue full of requests that were never meant all to be built.

AWS EMEA’s Pamela MacDougall said in February 2026 that securing power can take up to seven years, against about two years to build a datacenter. Treat the grid date as a planning input, not an edge case.

Five things a megawatt can mean

This is our own vocabulary, not a standard, but we find it separates most announcements from reality. Ask which of these a figure refers to:

  • Announced: a number in a press release or a plan.
  • Applied for or offered: a position in a connection queue, or a utility’s offer.
  • Contracted: capacity in a signed agreement with a date attached.
  • Energised: utility supply that is live at the site today.
  • Deliverable to your racks: critical IT load, after cooling and losses, at the rack you will rent.

Only the last two are capacity. The first three are intent, and intent has a date that can move by years.

A site total tells you how big the building is. The per-rack limit tells you whether your hardware runs in it.

Confirmed vs announced MW: the questions to ask

Put these to every operator before you sign, and get the answers in writing:

  1. Which of the five states is the figure you quote, and for which phase of the site?
  2. Is that capacity energised and serving customers today, or in build?
  3. If any of it is not live, what is the date, and what does it depend on that the operator does not control, such as transmission works?
  4. Is the figure gross utility load or critical IT load?
  5. Is the same capacity already promised to other customers?
  6. What is the contract remedy if the date slips?

Policy may start to help. Ofgem has proposed a Datacenter Commitment Fee, refunded once the facility draws power and forfeited if the project exits the queue early. The consultation closed on 16 September 2026 and there is no final decision yet. If it lands, a queue position would cost something, which is the point.

Why per-rack limits matter more than site totals

Even confirmed, energised megawatts do not help if they cannot reach your hardware. In our experience the ceiling per rack depends on the site’s power connection, its power distribution and its cooling, and that ceiling decides how many accelerators you can fit, whatever the building total says. A large site with a low per-rack ceiling can serve you worse than a small one with a high ceiling.

For colocation power planning, ask for the per-rack figure and what it assumes. In our Colocation you rent a third of a rack, two thirds or a full rack, up to 20 kW per rack where the site supports it. The power promise in our 99.9% SLA holds for equipment fed from both the A and the B feed whose real draw stays under 80% of the rack’s power connection.

A worked example with our own numbers: take a rack with a 20 kW power connection. The power promise holds while your real draw stays under 16 kW, which is 80% of 20 kW. Size your hardware against that 16 kW, not against the nameplate.

AI datacenter capacity planning on small, known sites

We operate four datacenters, and our installed power is known and stated here: MAD1 0.5 MW, AMS1 2 MW, MUN1 2 MW, HEL1 0.5 MW, 5 MW across the four. That is site capacity, not capacity for sale, and it is not a 50MW campus. If your plan needs tens of megawatts in one place, you need a different kind of provider; if it needs racks, ask us for the per-rack figure for your site in writing.

The GPUs behind our AI services are currently in Amsterdam only. AMS1 has six carrier uplinks to tier-1 carriers and sits in the Benelux gateway, with high-density fibre and UK and Ireland connectivity around it. Bare Metal GPU servers are in all four sites, and Colocation is sold in all four, with short-term capacity in Madrid and Amsterdam only. The facility operators provide space, power and cooling; the network and the transit are ours.

If your own date depends on a grid connection you do not control, plan a bridge: GPUs by the hour, or Bare Metal GPU servers now and your own racks when the power is live. Bare Metal has a six-month minimum term, and lead times are 4 hours for a server in stock and up to 14 days for a custom order, so build that into the schedule. Then run the six questions above on us as well.

Sources