In March I published an analysis of what I called the Gulf's two-hundred-billion-dollar execution gap. The thesis was that the region had committed more capital to AI infrastructure in eighteen months than any comparable sovereign bet in history, that the money was not the constraint, and that what would decide whether it converted into working systems was execution capacity — qualified people, proven methodologies, architectural discipline, production-grade institutional knowledge.
Five months of data later, I still think the constraint was named correctly. I put it in the wrong place.
I wrote the execution gap as though it would bind on the buildout — as though a shortage of qualified people would slow the conversion of committed capital into functioning infrastructure. That is not what happened. The infrastructure has moved faster than my framing implied was achievable, and the gap binds one layer downstream, where a bank or a hospital or a ministry has to take general-purpose capacity and turn it into a governed workflow that produces a countable outcome. This piece is the correction, with the numbers that forced it.
The buildout has executed roughly to plan
The Saudi position is the clearest test, because HUMAIN is the vehicle where a capacity ceiling would show up first. It has secured 211 land parcels for data centre development, with access to fourteen gigawatts of power. It has tendered infrastructure works for a six-gigawatt campus in east Riyadh, spanning roughly twenty-four square kilometres and developed in phases across six one-gigawatt plots. Facilities in Riyadh and Dammam at a hundred megawatts each are slated to come live around now, against a target of 1.9 gigawatts installed by 2030. The joint venture with AMD and Cisco announced in November 2025 begins operations this year with a hundred-megawatt first phase, and plans up to a gigawatt by 2030.
In the UAE, the first two-hundred-megawatt tranche of Stargate UAE is tracking to the third quarter of this year, with long-lead equipment procured and the first mechanical systems already on site — the opening phase of a gigawatt cluster inside a five-gigawatt campus.
None of that is a story about a capacity ceiling. It is a story about procurement and construction executing roughly to plan, which is a thing this region has been extremely good at for fifty years and had no particular reason to stop being good at because the payload changed.
The gap binds at the adoption layer
The conversion numbers tell a different story, and they have not improved.
McKinsey's GCC survey work, run with the GCC Board Directors Institute across 139 senior executives and board directors in all six Gulf states, shows adoption of AI in at least one business function rising from 62 percent in 2023 to 84 percent in 2025. Only 31 percent describe themselves as scaling or fully scaled. And only 11 percent qualify as what the study calls value realizers — organisations that have adopted AI in at least one function, are scaling or have scaled it, and can attribute at least five percent of earnings to it. Eighty-nine percent plan to increase their AI budgets in the coming year.
The distance between 84 and 11 is the entire argument, and it is not a money problem. Nothing in that chain is blocked by a shortage of compute.
The workforce picture underneath it has, if anything, deteriorated. The Hays GCC Salary Guide for 2026 reports that ninety percent of regional employers faced skills gaps during 2025. That is consistent with the thesis I argued in March. What it is not consistent with is the prediction I attached to it, which was that the shortage would show up as concrete not getting poured.
Building a gigawatt and governing a claims process are different problems
The distinction that does the work here is between a capital problem and an institutional one.
Building a gigawatt is a capital and construction problem. It has a tender, a critical path, a contractor with a track record, and a definition of done that everyone in the room agrees on. It is expensive, it is complex, and it is fundamentally a category of work the Gulf has industrialised.
Getting an insurer's claims process onto agentic infrastructure — with a stated authority envelope, an enforcement point where an action can still be refused, and an evidence trail a supervisor will accept — is an institutional problem. There is no contractor. The definition of done is contested inside the organisation. It requires someone senior to state, in writing, what the system may do on the firm's behalf, and that is a conversation most enterprises have never had about any actor, human or otherwise. The 31 percent figure lives there, and no amount of capacity moves it.
Where the survey data disagrees with my emphasis
There is a second correction, smaller and more uncomfortable, and I would rather state it than let it sit.
I led on human capital as the binding constraint. The most recent regional breakdown does not. A survey of more than a hundred organisations across the six GCC states, published in May 2026, puts technology integration as the most-cited challenge at 61 percent, ahead of talent gaps at 44 percent and uncertainty about return on investment at 37 percent.
I could argue that integration difficulty is largely what a talent shortage looks like when you ask the question differently — you cannot integrate what nobody on the team has integrated before — and I do believe that. But I do not get to have it both ways. I named a first-order barrier and the survey data disagrees with my ordering, and a retrospective that quietly reinterprets its own thesis until the data fits is worth nothing.
The supporting evidence in the McKinsey work cuts the same way. Among organisations that are not realising value, the weakest self-assessed capability is technology and data foundations at 37 percent, below change management at 41 percent and talent and operating model at 43 percent — while strategic alignment sits at 72 percent. Intent is not the missing piece. Substrate is.
What the numbers on both sides are not
A correction that presents its own supporting data as clean is repeating the original error, so it is worth being explicit about what these figures can and cannot carry.
The infrastructure numbers blur five different states. Committed, tendered, under construction, energised, and in productive use are not the same thing, and public reporting routinely runs them together. A six-gigawatt tender is a procurement document, not a data centre. Bids closing is not steel in the ground. A 2030 installed target is an intention with a plausible plan behind it. The gap between announced gigawatts and capacity an enterprise can actually rent this quarter is real and under-reported, and I would treat every headline figure in this space — including the ones I have used here — as an upper bound.
The adoption numbers have the opposite weakness. They are survey instruments, self-reported by executives with an interest in the answer, and the definitions vary between studies in ways that make cross-comparison shaky. The 84 percent adoption figure counts an organisation using AI in a single business function, which is a very low bar and flatters the region. The 11 percent value-realiser figure is the sturdiest number in the set precisely because it is defined tightly — five percent of earnings attributable to AI — but a tight definition on a self-reported base is still a self-reported base.
And the two-hundred-billion figure deserves the same scepticism. It aggregates commitments across entities with different time horizons, different funding certainty, and some overlap. It was defensible as an order of magnitude in March and remains so. It was never a balance sheet, and anyone quoting it as one — including me, if I do it — is over-claiming.
A smart critic would push further and say the two datasets do not even measure the same population: the infrastructure numbers describe a handful of sovereign vehicles, the adoption numbers describe a broad cross-section of enterprises, and there is no reason a lag between them should be read as a causal gap rather than an ordinary sequencing effect. That is a fair objection. My answer is that the sequencing effect has a name and a location — it is the governance and integration work at the enterprise boundary — and that where it has been done, the conversion has followed.
The structural advantage nobody is pricing
There is a compensating advantage in the Gulf position that I did not weight heavily enough in March, and it is the most interesting thing on the table.
Much of this estate is new-build. In a retrofit, governance is a negotiation with twenty years of accumulated systems, each with its own identity model and its own reasons why the authority boundary cannot move. In a new-build, the four properties that matter — entitlements granted by someone accountable, context assembled per call from entitled sources, policy enforced where the action crosses into the system of record, and evidence written at execution — cost a fraction of what they cost as a retrofit, because nothing has to be unpicked first.
The CBUAE's sovereign financial cloud with Core42 is the clearest instance of the pattern, and it is why I have been more confident about that piece of my March analysis than any other. Governance embedded at the infrastructure layer rather than added at the application layer is the property that distinguishes it. A supervisor that owns the substrate can specify what may happen on it; a supervisor that owns only the rulebook has to hope.
That advantage has a clock on it. It is available for exactly as long as the estate is still new. Every quarter that a gigawatt of capacity carries workloads whose authority model was never stated, the Gulf converts a structural advantage into the same retrofit problem everyone else is paying for.
What a Gulf enterprise should actually do
None of what follows is a purchase, and none of it waits on a supervisor.
Take the three workflows where an agent would touch a system of record, and for each one write down the authority envelope before anything is wired in: what class of action it may take, up to what value, on which accounts, granted by which named person, expiring when. If nobody in the business will put their name to that sentence, you have found the actual blocker, and it is not a technology-integration problem or a talent problem. It is an ownership problem, and it is cheaper to discover it now.
Then put the enforcement point where the action crosses into the system of record, in the gateway or service layer you already operate — not as an instruction in a prompt the agent may choose to follow. Then write the evidence synchronously with the action: what the agent knew, from where, under which entitlement, evaluated by which policy version. Those three decisions are ordinary engineering. They are also, in almost every programme I look at, the difference between the pilot and the thirty-one percent.
There is a timing argument on top. The revised interagency model risk guidance issued in North America on 17 April 2026 placed generative and agentic AI expressly outside its scope, with separate guidance promised. Whenever that arrives, it will be written against whatever the industry has already built. The same is true of the CBUAE's expectations and of every other supervisor now watching this space. Firms building the evidence layer this year are not complying with a specification. They are becoming one.
So: one call that held, one that was aimed at the wrong layer, and one emphasis the data has argued with. The Gulf did not have a capital problem and it does not have a construction problem. It has the problem everyone else has — converting capacity into a governed workflow somebody will sign for — with the unusual advantage of an estate new enough to design that in rather than bolt it on, and a shrinking window in which that advantage is still free.
If you are on the enterprise side of that gap in the Gulf and the authority model has not yet been written down, that is the conversation I have most weeks — compare notes with me.