How a quarter becomes a year

Most published dates do not slip in an announcement. They widen, restated a few times until the original commitment has quietly dissolved.

15 September 2026

The slip you can see before it is announced

A schedule date is a point. The commitment around it is a window, and the window is where the information is.

When a program narrows its window, it believes the date. Somebody has looked at the remaining work and decided a quarter is defensible where a year used to be. That signal usually passes unremarked, because narrowing sounds like confidence rather than news.

When a window widens back, that is the slip. Not the announcement of one. The slip itself. And it rarely gets reported, because nothing was withdrawn, nobody was contradicted, and each statement was accurate on the day it was made.

This is not deception. It is how organizations talk. A person answering a question on stage in December is describing a quarter. A press office in September is describing a year. Both are being straightforward. Only the sequence carries the finding, and nobody is responsible for reading the sequence.

It is common enough to have a number

Currence, which tracks 190 GW across 777 data center projects announced since 2024, reported in February 2026 that in 2025 twenty six percent of expected capacity slipped6, and a further ten percent of projects pushed back their commercial operation dates "without much notice"6. Their outlook for 2026 is that thirty to fifty percent of planned capacity ends up delayed.

Two things follow. Predicting that a large program will be late is a free bet: it proves nothing and deserves no credit. And that ten percent is the phenomenon this piece is about, already measured by somebody else. Dates moving quietly is not an anecdote. It is roughly one project in ten, every year, in a sector where a year of capacity is worth tens of billions.

Those slip figures are analyst assessments rather than audited counts, and should be read as such.

Four statements about the same 200 megawatts

Stargate UAE is a one gigawatt AI compute cluster in Abu Dhabi, built by G42 with OpenAI, Oracle, NVIDIA, Cisco and SoftBank. The first phase is 200 megawatts.

Four statements about the same 200 megawatts, from the launch announcement, the developer, a named executive on stage, and a wire report nine months later.

The December statement deserves credit rather than suspicion. A named executive narrowed a year to a quarter, on stage, at a named event. Most programs get vaguer as a date approaches. This one got sharper, which is the behaviour of people who believe their own plan.

Nine months later the same phase is a year again, and G42's own comment to Reuters is that "the project's specifics are subject to continuous review."3 A quarter became a year. Nothing marks it.

The other half: read the verb

Look at the table again and ignore the dates. Go live. Delivery. In place and ready. Come online.

Four different events, and on a program of this kind they sit months apart. A hall can be finished and energized while serving nothing. Commissioning can be complete while integrated testing continues.

Crossrail is the case that settles it, because the program published its own record. The tunnels were finished in 2015. The railway opened in 2022. For most of those seven years a completed railway existed and a working one did not.

The same program already did this, twelve months ago

You do not have to reason from a British railway. Stargate's own flagship American site supplies the precedent.

On 30 September 20254, the last day of a quarter, Crusoe announced that the Abilene campus was "live"4. The release says the first phase is "up and running on Oracle Cloud Infrastructure," that the first two buildings "were energized," and that the campus "has already begun running early training and inference workloads."

It states no megawatt figure.

What that omission was worth became clear later. Epoch AI, tracking the site carefully from outside, estimated 0.6 GW operational. Then in April 2026 Oracle noted that 200 MW was operational, and Epoch revised its own figure down to roughly 0.3 GW5.

A careful independent analyst, reading public announcements, overestimated a site by a factor of two, because "live" and "energized" and "running early workloads" are not a load figure and were never meant to be.

Why Stargate UAE is a good test rather than an easy target

Score it against the seven flows that enable any construction activity, and five are supplied better here than on almost any comparable American project.

Materials were procured early: G42 reported all long lead equipment secured and first mechanical systems delivered2 by October 2025, which is precisely the failure point that strands US campuses for years. Labour is available at scale. The site is greenfield, 19.2 square kilometres1, with no existing operations to work around. Capital is sovereign. Permitting and interconnection, the constraints that bind Texas and Virginia, are not constraints here.

This is close to a best case, which is exactly what makes it worth watching. It isolates the variable. If the date moves on a program with five flows fed on demand, the cause is one of the two nobody outside can see: the information flow that turns a built hall into a proven one, and the precedence chain inside commissioning.

Then a sixth flow moved, from outside

On 11 September, Reuters reported that the wider campus is being redesigned. A single site is likely to become a network distributed across the UAE, with air defence, blast resistant concrete and underground construction under consideration, following attacks on Gulf data centres earlier this year. The first phase is expected to continue, with modifications3.

That is the external conditions flow, and it came from nothing inside the program. No planning competence anticipates it and no replanning absorbs it.

Which is the distinction worth taking away. Some flows belong to the program and some belong to a system above it.

Crossrail's binding flow was information, which was internal, and that is why the fix worked: after the August 2020 replanning, milestone attainment went from roughly 30 percent to roughly 90 percent7 with the same contractors on the same railway. The repair was available because the flow was theirs.

An interconnection queue is not theirs. A regional security threat is not theirs. When the binding flow belongs to a shared system, a program's own competence stops being the variable that decides the date. Projects stop failing one at a time and start failing together, and nothing about any individual project looks wrong beforehand.

The prediction

Published 15 September. The quarter ends 30 September.

Not a prediction about the schedule, because twenty six percent of capacity slipped last year and betting on lateness costs nothing. A prediction about the language, grounded in what the same program did on the last day of a quarter twelve months ago.

If an announcement lands by 30 September, it will use a completion verb rather than a load verb. Ready, delivered, complete, energized, live. Not serving, carrying, in production. It will not state a figure for compute actually being served. And no mainstream coverage will separate a finished building from served compute.

Checked on 1 October, stated now so it cannot be adjusted later:

The follow-up publishes in the first week of October either way. If the first 200 megawatts are announced as carrying production load inside the quarter, this reading was wrong, and that will be the headline of the second piece rather than a footnote to it. A gigawatt class program hitting a publicly stated date is rare enough to be the more interesting outcome.

What to take to work

Track the window, not the date. Keep the last four public statements about any commitment that matters to you and note the width of each. Direction of travel tells you more than the date, and tells you earlier.

Ask what the verb means. Ready for what. Complete to what standard. Live serving whom. Epoch AI missed by a factor of two on that question. Your supplier's answer is often a different month than the one you were given.

Neither is sophisticated. Both work because almost nobody does them, and because the information is already public, sitting in documents that no single review reads side by side.

by Ahmed Schrute