>
State of Play is a read on a sector: its size, its structure, where the money moves and what moves it. Each edition takes one market and sets out what is known, what is claimed, and what the difference between the two is worth to anyone deciding where to commit.
Twelve gigawatts have been announced across the Gulf and a few hundred megawatts energised: the distance between the two is where this read sits. Sovereign artificial-intelligence compute across the six Gulf states: what has been announced, what has been financed, what is energised, and why those three numbers differ by a factor of about three.
Figures carry their basis and their date. Where published estimates disagree, the range is given and the basis named rather than averaged away, and where a widely quoted number does not survive checking it is corrected here rather than repeated. Every chart plots a figure stated in the text, and nothing is projected beyond what a named source projects.
The Gulf’s capacity divergence is not a disagreement between analysts. It is the difference between measured load and press-release nameplate.
The two numbers in circulation for Gulf artificial-intelligence capacity in 2030 are not two estimates of one thing. The low family, of the order of 3.3 gigawatts across the six states, is built on a measured operational base and projected forward (FTI Consulting via Arab News, 21 August 2025). ValuStrat echoed it in its GCC Data Centre Market Review 2025. The high family, of the order of twelve gigawatts, is the arithmetic sum of announcements. Some are dated 2034 and some are multi-country. None carries a documented grid connection. The ratio between them, about 3.6 times, is a definitional artefact rather than a forecasting dispute.
The two families are reconcilable, and the reconciliation factor is about three. In the Gulf’s most mature market, the live-to-total-tracked multiple can be derived from two figures published by DC Byte. Self-built capacity of approximately 65 megawatts represents 5.44 per cent of total tracked United Arab Emirates supply, which implies a tracked denominator near 1,195 megawatts. That sits against Knight Frank’s live print of 376 megawatts of information-technology load (DC Byte via Data Center Knowledge, 19 February 2026; Knight Frank, 14 October 2025). That multiple of
roughly 3.2 times, applied in reverse to a twelve-gigawatt announced pipeline, implies about 3.7 gigawatts eventually live. The two prints describe one market.
Three corrections frame what follows. The flagship Saudi vehicle changed hands on 1 November 2026. The Public Investment Fund and Aramco closed the HUMAIN transaction at 51 per cent and 49 per cent respectively. Their joint statement gives a planned deployment of three gigawatts of artificial-intelligence data centre capacity by 2030, not the six gigawatts widely reported (Aramco, 1 November 2026). Announced capacity and operational capacity are different ledgers. They are kept separate throughout. The binding constraint on delivery is power rather than silicon,
a position stated plainly by Cushman and Wakefield in its 2026 global comparison of 107 markets.
A megawatt figure without a stated basis and without evidence of a grid connection is not a capacity estimate, and any party contracting into this market should treat it as a headline. Applied consistently, that test removes most of the Gulf’s announced pipeline from the installed column without asserting that any single project has failed.
In the Gulf, sovereign artificial intelligence mostly means in-country data residency and state-adjacent ownership, not a domestic frontier-model capability.
The term sovereign artificial intelligence is used in the Gulf to describe four distinct things, and the differences are consequential for anyone sizing the market. The first is state or state-fund ownership of compute assets, as with HUMAIN under the Public Investment Fund and Aramco, or Qai under the Qatar Investment Authority. The second is in-country data residency on commercial cloud, which is what Oracle Alloy delivers for Bahrain’s Information and eGovernment Authority and what the Kuwait launch arrangements with Google describe. The third is domestic model development. The fourth is domestic silicon. Only the first two are materially present in the region.
Qatar illustrates the distinction cleanly, and corrects a common misdescription. Syntys, the Doha-headquartered carrier-neutral platform carved out of Ooredoo Group and launched to market on 11 March 2025, is a hosting landlord. It owns buildings and sells capacity. It is neither a model developer nor a graphics-processing-unit owner. Qatar’s national artificial-intelligence company is Qai, wholly owned by the Qatar Investment Authority, launched on 8 December 2025 under chairman Abdulla Al-Misnad,
and Qai has stated explicitly that it is not building frontier large language models. The vehicle with the capital and the vehicle with the floor space are different companies.
The fourth definition, domestic silicon, has in one instance been abandoned. Alat, the PIF company launched in February 2024 with a stated hundred-billion-dollar budget, removed its founding chief executive and abandoned semiconductor manufacturing in April 2026, redirecting resources towards the data centre build-out rather than towards owning facilities. It holds no capacity. The reporting on that retreat was not retrievable at source and is treated here as unverified detail around a verified direction of travel.
A further qualification applies to the word sovereign where it is used by operators. Ooredoo Kuwait described its facility with NVIDIA, announced on 9 November 2025, as the nation’s first sovereign artificial-intelligence data centre. In that usage sovereign denotes in-country data residency rather than state ownership. The operator is Syntys, so Kuwait’s single
artificial-intelligence facility is majority Qatari-group owned. Label and ownership diverge. The label is the company’s own.
For the purposes of this read, sovereign artificial-intelligence capacity is therefore defined narrowly: compute plant inside a Gulf state, owned or anchored by a state entity or state fund, and energised. On that definition the regional installed total is measured in hundreds of megawatts. Live training capacity is close to nil. Saudi Arabia’s operating base is colocation, two public cloud regions, two research clusters and one inference region. Google’s live Dammam region offered no A2, A3 or A4 accelerator machine families when its region listing was read on 31 July 2026.
Six different definitions of a data centre megawatt are in circulation in the Gulf, and the spread between them is wider than most of the growth being debated.
The single most useful discipline in this market is to ask what a megawatt figure measures before asking whether it is large. Six bases are in live circulation. Information-technology load, the power delivered to computing equipment, is the reference at one times. Gross or facility power, which adds cooling and lighting as well as losses, runs at roughly 1.2 to 1.5 times that reference and is rarely labelled. Grid or utility supply,
meaning megawatts requested or contracted from a utility, runs at roughly 1.3 to 2.0 times. Operational capacity, as published by Saudi Arabia’s Ministry of Communications and Information Technology, carries no published definition at all. Contracted capacity at full occupancy is saleable design capacity rather than energised load. Announced capacity is not capacity.
Two further traps appear in the live record. One is the presentation of solar generation megawatts adjacent to data centre megawatts. Beyon’s Bahrain Solar Park Phase 5 at 1.5 megawatts-peak and Oman Data Park’s 1.4 megawatt captive plant are generation, not information-technology load, and must not be added to capacity. The other is double counting through the ownership stack, which Part II takes in detail. Both traps inflate totals without any party making a false statement.
Where a publisher does define its basis, the definition should be used. Knight Frank’s glossary defines a data centre megawatt as the amount of
information-technology load or computing power that can be supported. It defines live information-technology capacity as power currently operational and available for use. Arizton defines power capacity as the total information-technology load a facility is designed to support. Those two definitions, taken together, make the United Arab Emirates and the small-state colocation figures in this read comparable with each other. Saudi Arabia’s national series is not comparable with either. Its basis is unstated.
The consequence for the Saudi dataset in particular is large. If the Ministry’s 467 megawatt print for the first quarter of 2026 is gross facility power rather than information-technology load, every derived Saudi figure shifts by 20 to 50 per cent. No methodology note and no definition have been published alongside the series. Nor has any load-versus-gross statement. That is the most consequential single measurement gap in the Gulf dataset, and it sits at the centre of the region’s largest programme.
Facility counts corroborate the concentration. There were 112 operational data centres in the Gulf and 78 upcoming developments as reported on 27 October 2026. A separate count puts the national distribution at the start of March 2026 at 61 in Saudi Arabia, 57 in the United Arab Emirates, 16 in Oman, 11 in Qatar, 8 in Bahrain and 5 in Kuwait. It is also reported that
Khazna, stc and Gulf Data Hub together represent more than half of total Gulf information-technology power capacity. That concentrates counterparty risk as much as it concentrates capacity.
Gulf operational capacity in 2026 is of the order of 880 megawatts of information-technology load, with Saudi Arabia and the United Arab Emirates holding more than 85 per cent of it.
The measured regional base can be constructed two ways. The two agree. Taken from the top down, Arizton via Research and Markets reported on 27 October 2026 that the United Arab Emirates holds more than 400 megawatts and over 45 per cent of total operational information-technology power in the Gulf. That implies a regional total of the order of 880 megawatts. The same report puts Saudi Arabia and the Emirates together at more than 85 per cent of operational capacity, leaving under about 130 megawatts for the other four states. Taken from the bottom up, Qatar sits at roughly 40 to 45 megawatts and Kuwait at 2 to 10 megawatts, with Bahrain in the low tens and Oman at 10 to 20 megawatts. Those sum to roughly 75 to 105 megawatts, inside the top-down envelope.
One independent series applies a single method to both leaders and should be noted with its weakness. The International Data Center Authority, published on 21 July 2026 via Arab News, gives Saudi Arabia 440 megawatts against the Emirates at 340 megawatts, the pair holding roughly 80 per cent of a regional total near one gigawatt. That series is cited here through a secondary aggregation, because the article itself returned navigation content only and could not be opened. It is not independently
echoed elsewhere. Its value is as a cross-check on the shape of the region rather than as a level.
The honest statement of precision matters more than the level. Research and Markets published United Arab Emirates figures of more than 400 megawatts and more than 430 megawatts on the same day, 27 October 2026, both described as current information-technology load capacity in 2026, in releases of different geographic scope. The precision of these datasets is therefore not better than plus or minus 10 per cent. They should not be quoted to three significant figures. Usable capacity, in the formulation of Data Center World Middle East, is what has been built, energised and cooled, and then certified and connected. It is still counted in hundreds of megawatts.
Export control is the subject of most commentary on Gulf compute. Grid interconnection is the subject that actually gates delivery.
Power delivery is the binding constraint on Gulf artificial-intelligence capacity, and the evidence for that proposition is both direct and structural. Cushman and Wakefield compared 107 markets across 24 variables in its 2026 global data centre comparison, published on 18 May
2026. That report states that power is still the primary constraint across global data centre markets, and that an artificial-intelligence infrastructure bubble appears unlikely because structural frictions constrain overbuilding. It also finds that in Europe, the Middle East and Africa the proportion of realistically deliverable pipeline capacity is declining even as pipeline capacity expands. Those are the publisher’s own stated takeaways. They cut against an oversupply thesis.
The structural evidence is that the Gulf’s largest commitments entered the public record as power procurement. The only grid-connection datum in the entire Saudi record is a tender issued on 4 May 2026 (MEED, 4 May 2026). It covers a 380/132/33 kilovolt network, 500 megavolt-ampere and 200 megavolt-ampere substations and a 2,000 megavolt-ampere bulk supply point for HUMAIN’s Al-Saad campus in east Riyadh. In Kuwait, three 132/11 kilovolt main transformer stations estimated above 22 million dinars, about 71.5 million dollars, gate the country’s hyperscaler regions entirely. That construction contract was not confirmed as awarded as at 4 November 2026.
The counter-example that proves the point is the United Arab Emirates. Energy consumption is not currently a binding constraint on growth there, in DC Byte’s assessment of February 2026. Abu Dhabi and Dubai were named by Cushman and Wakefield in September 2026 among markets benefiting from faster power delivery, available land and strong connectivity. Emirati power costs approximately 0.11 dollars per kilowatt-hour,
nearly double Saudi Arabia and more than double Qatar. The Emirates buy speed at a price. Where power is not the constraint the market grows; where it is, the pipeline stalls irrespective of capital.
Chip access, by contrast, is a political risk rather than a physical one. Part II takes it separately. A facility without silicon can be filled later, while a facility without a connection agreement cannot be energised at all. New capacity in the region arrives in increments of 100 to 200 megawatts, gated by grid interconnection and generation rather than by land or capital. Contract against energisation dates rather than campus masterplans, because that distinction decides when revenue starts.
Saudi Arabia has added about 94 megawatts a year for four years. Every published 2030 target requires between two and six times that rate.
The run-rate test disciplines the Gulf’s 2030 range, and nothing else does. Saudi Arabia’s own ministerial series gives 68 megawatts of operational capacity in 2021, 440 megawatts in 2025 and 467 megawatts in the first quarter of 2026 (Ministry of Communications and Information Technology, 5 May 2026). That is 399 megawatts added over four and a quarter years, or roughly 94 megawatts a year. The most recent quarter annualises to about 108 megawatts a year. Against that delivered rate, the
national target of about 1.5 gigawatts by 2030 requires some 218 megawatts a year sustained for five years. That is twice the historical rate.
The higher targets require step changes that no published delivery schedule supports. The three-gigawatt figure stated jointly by the Public Investment Fund and Aramco on 1 November 2026 implies about 533 megawatts a year, 4.9 times the historical rate. HUMAIN’s own 6.6 gigawatt pipeline by 2034 implies about 700 megawatts a year, 6.5 times the historical rate, sustained for nine years. A forecast that requires a sixfold step change in delivery, with no published delivery schedule and no disaggregated connection queue, is not a forecast. The arithmetic for the three-gigawatt row is this practice’s own. It is applied to the ministerial base.
Two moderating observations are owed. Run rates are not physical laws: the 2021 to 2026 period largely predates the artificial-intelligence capital cycle, and a market can re-rate. The increments now being built are also larger than those historically delivered, which is precisely why grid connection rather than construction has become the gate. Neither observation closes a 6.5 times gap, and neither comes with a published schedule that would allow it to be tested.
The test applies with equal force outside Saudi Arabia. Organic growth rates there are smaller still. Knight Frank recorded 19.2 megawatts of new Emirati capacity coming online during 2025 against a live base above 376 megawatts. DC Byte recorded nearly 20 megawatts of artificial-intelligence information-technology power going live in the Emirates in the year to February 2026. Those are the observed rates in the region’s most capable market. They are the correct starting point for any projection, and they are an order of magnitude below the announced pipeline.
Physical security entered the Gulf data centre column in 2026. One market’s installed capacity went backwards and one masterplan is being revised.
The 2026 regional war has changed the Gulf asset base in ways that no capacity forecast published before it reflects. United States and Israeli strikes on Iran began on 28 February 2026. On 11 March 2026 Iran’s Islamic Revolutionary Guard Corps declared that targets included infrastructure of Google, Microsoft, Palantir, IBM, NVIDIA and Oracle in Gulf states. Iran also published a list of 29 technology targets across Bahrain, the Emirates, Qatar and Israel. Iranian drone strikes damaged three
Amazon Web Services data centres from early March 2026, two in the Emirates and one in Bahrain (CSIS, 15 July 2026; Al Jazeera citing Reuters, 24 March 2026).
The operator’s own description of the damage is specific. Two Emirati facilities were directly struck. In Bahrain a drone strike in close proximity to a facility caused structural damage, disrupted power delivery and fire suppression activity that resulted in additional water damage. On 15 September 2026 Amazon Web Services declared that it was unable to restore access to resources and data hosted exclusively in the Bahrain region, me-south-1, and in the Emirati availability zone mec1-az2. That region had launched on 30 July 2019 with three availability zones in Manama. The damage spanned multiple availability zones and exceeded what regional and multi-zone services are designed to withstand. A further update was promised for early 2027.
Two consequences follow for the numbers used here. Bahrain is the one Gulf market whose installed capacity went backwards during 2026, and any pre-March-2026 Bahrain total that included Amazon Web Services is now overstated. The Emirates’ five-gigawatt artificial-intelligence campus
masterplan is meanwhile under revision. Reuters reported on 11 September 2026 that the original single-campus plan in Abu Dhabi may be split into a network of sites, with officials weighing air defences and partly underground construction. G42 told Reuters that campus work remains on schedule. Reuters could not determine how close officials were to approving a new plan or what the effect on cost and timing would be.
The stated cause of that revision is physical security and infrastructure resilience following the war, not power. It is the one place in this read where the constraint is something other than the grid. Amazon Web Services advised clients to consider migrating workloads out of the Middle East, which is a demand-side effect as consequential as the supply-side loss.
Saudi Arabia’s exposure is different in kind. No facility inside the Kingdom has been reported struck. On 27 July 2026 Saudi air defences intercepted drones aimed at oil infrastructure in the Eastern Province and Riyadh, all destroyed, with no damage, and the targets were energy assets rather than data centres. A further claim by the Revolutionary Guard Corps of a July 2026 strike on the same Bahrain site is unverified and is not used here: it could not be confirmed because Bahraini services had already been offline for months.
The war also repriced the asset base. That matters to anyone underwriting Gulf compute. Before the fighting the regional market would write as much as a billion dollars of political-violence cover on a single risk, at an annual premium of between 0.1 and 0.3 per cent of the policy limit. By
October 2026 available capacity had fallen to between 100 and 300 million dollars, and premiums ran from 2 to 10 per cent of limit (Marsh, 2026). Critical infrastructure and large data centre projects sit at the upper end of that range. Cover contracted as demand for it grew.
The insurers' own losses explain why the repricing is likely to hold. Political-violence claims from the early rounds of missile and drone strikes came to about 2 billion dollars (Marsh, 2026). Global annual premium for that entire class of cover is roughly 1.5 billion dollars. The broker's head of global placement for the region expects two years from any lasting ceasefire before pricing approaches its pre-war level, and allows that it may never fully return. One insurer once wrote 100 million dollars on a single risk. It now offers about 10 million. A programme of any size must therefore be placed across many more balance sheets.
The consequence for this read is a new entry in the cost of a Gulf megawatt. Insurance was a rounding item in a data centre operating budget before 2026 and is now a significant one. For the largest campuses it has become a condition of financing rather than a line within it. Gross insured political-violence losses across the region ran above 3 billion dollars between the start of March and late July 2026 (Marsh, 2026). Pre-war forecasts carry none of this. A megawatt costs more than it did, and the premium is the measure of how much more.
The Gulf’s flagship compute vehicle is no longer wholly state-fund owned, and its official 2030 target is half the figure most commentary carries.
The Public Investment Fund and Aramco closed the HUMAIN transaction on 1 November 2026, with the Fund holding 51 per cent and Aramco 49 per cent. Consideration was approximately 9.9 billion riyals, about 2.6 billion dollars, in 2026 cash contributions from Aramco, together with the transfer of artificial-intelligence, compute and digital assets. The joint shareholder statement gives the planned deployment of three gigawatts of artificial-intelligence data centre capacity by 2030. That is the only capacity figure for HUMAIN carried in a formal release naming both shareholders. It is preferred throughout this read.
Trade press and company briefings carry six gigawatts by 2030 and 6.6 gigawatts by 2034 (Middle East AI News, 4 September 2026). On the available evidence the six-gigawatt figure is the site envelope of the Al-Saad campus, six plots of one gigawatt each, rather than a delivery commitment. Where the larger number is used it should be quoted alongside the three-gigawatt shareholder figure and labelled as a campus envelope. The distinction matters. The two numbers imply delivery rates of 533 and more than 1,000 megawatts a year respectively against a national historical rate near 94.
HUMAIN’s verified live capacity is modest. The GroqCloud Dammam inference region, built with Groq and Aramco Digital, runs at roughly 50 megawatts with 19,000 language processing units and its service is live. The
company’s Riyadh and Dammam artificial-intelligence factories, each stated at at least 100 megawatts, both broke ground in August 2025 with go-live stated as the second quarter of 2026. That quarter closed on 30 June 2026 with no launch announcement, and independent trackers continue to class both facilities as under construction. The honest formulation is that no public evidence confirms that they opened. This is the largest single unresolved item in the Saudi ledger.
Two documented refusals to disclose bear on how this company should be read. On 18 June 2026 HUMAIN declined to detail locations or projects when disclosing its 211 land plots. On 24 July 2026 it declined to reveal the specific requirements of its NVIDIA arrangement or to say how many chips had been delivered. The company’s own domain is rendered in client-side script and returns nothing to a plain fetch, so the flagship Saudi artificial-intelligence vehicle publishes nothing machine-verifiable. A separate trap should be noted, since NVIDIA’s own May 2025 release contained it. The domain humain.ai belongs to an unrelated Belgian consultancy.
The structural reading is that HUMAIN has become an anchor tenant as much as a developer. Its 36 megawatt anchor commitment inside DataVolt’s 44 megawatt Riyadh East facility represents more than 80 per cent of that facility’s planned capacity and converts HUMAIN, in that instance, from builder to customer. That is not a criticism of the model. It is a correction to the assumption that HUMAIN’s target capacity will sit on its own balance sheet.
The largest line item in the Saudi pipeline entered the public record as a procurement for switchgear. That is the most instructive fact in the dataset.
HUMAIN’s Al-Saad campus in east Riyadh carries a nameplate of 6,000 megawatts, six plots of 1,000 megawatts each, and its status is the tendering of enabling works. The only grid-connection datum in the entire Saudi record attaches to it: a tender issued on 4 May 2026 for a 380/132/33 kilovolt network, 500 and 200 megavolt-ampere substations, and a 2,000 megavolt-ampere bulk supply point (MEED, 4 May 2026). The 2,000 megavolt-ampere figure is utility apparent power. It is not information-technology load and it is not megawatts.
The sequencing implied by that tender is the useful part. A bulk supply point of that scale is a multi-year utility undertaking, and it precedes rather than follows the delivery of compute. Read forward, it means the Al-Saad
envelope cannot energise in any material quantity inside the current decade on the evidence available. Read backward, it means the project is real. Enabling works are being procured, which is more than can be said for several larger nameplates elsewhere in the region. Both readings should be held at once.
The same discipline applied to the Emirates yields a comparable result. The five-gigawatt UAE-US AI Campus in Abu Dhabi is verified from two independent official sources, G42 on 15 May 2025 and the UAE Embassy, whose current text states that construction is underway with the first 200 megawatts ready in 2026. Only the one-gigawatt first-phase cluster is specified, and of that only the 200 megawatt tranche is dated. The remaining four gigawatts has no phasing, no financing detail and no commissioning dates in the public record. The basis of the five-gigawatt figure is not stated. It is aspirational.
The operational status of the most-cited Emirati project is zero verified megawatts. Stargate UAE was announced on 22 May 2025 as the first international Stargate deployment and the first partnership under OpenAI for Countries. It comprises a one-gigawatt cluster with a 200 megawatt first phase expected to go live in 2026, firmed to the third quarter of 2026 by Mubadala’s managing director in December 2025. No commissioning, energisation or go-live announcement exists from any of its six named partners. The G42 newsroom index, read current to 2 November 2026, carries no Stargate UAE operational release at any point after the launch.
G42’s own November 2025 inventory of live supercomputing assets does not include it.
Two further corrections apply to that project. The Stargate brand has been retired from the campus phase one. OpenAI has stepped back from other international Stargate locations and dropped the label from newer efforts, and the UAE-US AI Campus no longer uses it for phase one (reported 4 September 2026). Conflicting secondary claims of a February 2026 commissioning should be rejected in favour of the absence of any partner commissioning release. So should one tracker that simultaneously lists the project as under construction and as 1,000 megawatts operational. Whether the first phase uses GB200 or GB300 silicon is unresolved in the public record and is not asserted here.
An independent audit of thirteen Saudi and Emirati projects found none at the firm-connection gate and one in seven defining its own load basis.
The Gulf’s announced pipeline has been audited once, and that audit carries much of the argument here. Madar Strategy, in research note 01 of 22 July 2026, examined thirteen Saudi and Emirati cases against a gross announcement universe of 4,000 megawatts, a figure the author labels explicitly as not a market total. Of seven headline projects, one in seven defines information-technology load, three in seven meet the project-site test, and none publishes firm connection terms. Across all thirteen cases, eleven sit at the lowest power-evidence gate with no project-specific power evidence, one at the next and one above that. None reaches the firm-connection or energised gates.
The individual gradings are instructive. Stargate UAE at 1,000 megawatts is graded at the lowest gate on both disclosure and power. The Microsoft and G42 Emirati expansion at 200 megawatts is graded at the lowest gate on disclosure, site and power. DataVolt at Oxagon, carrying a 1,500 megawatt campus nameplate, is site-identified only, with its latest milestone dated 10 February 2025. The author’s own caveat must be carried with the finding, and is carried here: the audit does not establish that any project is
cancelled, delayed, unfunded or unable to secure power. It establishes only what is evidenced.
The author is an independently published individual analyst rather than an institution, and the finding is cited as such. Its value is not authority but method: it is the only published attempt to grade Gulf announcements against power evidence rather than to aggregate them. Its conclusion is corroborated structurally by the absence of a published connection queue in Saudi Arabia, by the fact that official Saudi statistics do not disaggregate data centre consumption at all, and by the unawarded Kuwaiti substation contract.
A reader who applies those five gates to any Gulf announcement will arrive at the low-mid family of estimates without needing a model. The practical instruction, in the words of Data Center World Middle East, is to contract against energisation dates rather than campus masterplans. It should be noted alongside that an artificial-intelligence training rack draws above 100 kilowatts against 5 to 10 kilowatts for traditional equipment. That is why floor space counts tell a reader little about load.
The preamble adds the condition that determines what follows. If the two entities fail to become United States companies on or before 6 April 2027, they will need to apply for authorisation under the licensing process at section 748.3(c). Against that, eight United States-headquartered companies and their subsidiaries receive licence-free items and full Strategic Trade Authorization with no expiry: Amazon, Apple, Google, Meta, Microsoft, OpenAI Group PBC, Oracle and X.AI. The asymmetry is the
point. American operators hold durable access; the Emirati national champion holds a five-month-dated authorisation conditional on a change of domicile.
Only the Emirates holds a country rule, it names two companies, and its authorisation expires on 6 April 2027.
United States chip access to the Gulf is entity-based rather than country-based, and the one country framework that exists carries an expiry date. The Bureau of Industry and Security final rule titled Enhanced Favorable Treatment for the United Arab Emirates was published at 91 Federal Register 43034 and took effect on 10 July 2026. It creates a new supplement naming exactly two Emirati commercial entities, Group 42 Holding Ltd, trading as G42 and including G42 Cloud Technology LLC, and Core42 Technology Projects LLC. Neither receives Strategic Trade Authorization eligibility. The rule states verbatim that absent subsequent notice the authorisation for those two entities shall automatically expire on 6 April 2027.
Saudi Arabia has no country framework at all. That finding was established by exhaustive query of the Federal Register across all Bureau of Industry and Security documents published from 1 January 2026. The
query returned 45 documents, of which one is a Gulf rule, being the Emirati rule above. Qatar, Kuwait, Bahrain and Oman have no rule, no published authorisation and no named entity. The practical reading is that the Kingdom’s access runs through licences issued to its American counterparties and through case-by-case authorisations rather than through a published national framework. That is a weaker position than is generally assumed of the region’s largest programme.
One widely used planning figure should be discarded. The claim that 500,000 chips a year would flow to the Emirates originates in Reuters reporting of 14 to 15 May 2025, citing two unnamed sources on a preliminary agreement still under negotiation. It appears in no published government document: not in the Commerce framework text of 15 May 2025, not in the Bureau’s press release of 10 July 2026, and not in the final rule itself. It should not be used as a planning figure.
The AI Diffusion Rule still sits in the Code of Federal Regulations, unenforced by executive discretion and reactivatable without new rulemaking.
The strongest single fact supporting the political reversibility of Gulf compute access is that the instrument most commentary treats as repealed has never been rescinded in the Federal Register. A full-text search for Artificial Intelligence Diffusion from 1 January 2025 returns five documents only. They are the two rules of 15 January 2025 at 90 Federal Register 4544 and 90 Federal Register 5298, and the amendment of 14 February 2025 at 90 Federal Register 9604. The others are the Emirati rule of 14 July 2026 and the Unified Agenda introduction of 14 August 2026 at 91 Federal Register 52792. No published rescission or replacement for 90 Federal Register 4544 exists as of the last Bureau rule indexed, 24 September 2026.
What exists in place of a rescission is a press release of 13 May 2025 together with a non-enforcement posture. The Emirati rule itself states that the Bureau is only enforcing the relevant licence requirement to destinations in Country Groups D:1, D:4 or D:5. That is enforcement discretion rather than repeal. The instrument therefore sits in the Code of Federal Regulations, unenforced by executive choice, and reactivatable without new rulemaking. Any Gulf compute plan with a horizon beyond the current United States administration should be stress-tested against that fact.
The statutory position reinforces the same conclusion. The GAIN AI Act was not enacted. It passed the Senate as an amendment to the fiscal year 2026 National Defense Authorization Act by 77 votes to 20 on 9 October 2025, was excluded by the House, and was stripped in conference. It does not appear in the enacted Act. The White House opposed it. The net statutory position is that there is no enacted statutory chip-export constraint on the Gulf, and the regime rests entirely on executive discretion. A reported follow-on, the AI OVERWATCH Act in the Senate fiscal year 2027 bill, is unverified and is not relied upon here.
Two assumptions about the regulatory environment inside the region should also be corrected. Saudi Arabia repealed its government cloud
localisation mandate, moving from the National Cybersecurity Authority’s cloud controls of 2020 to the 2024 revision, which changes the residency calculus for any party assuming a hard localisation requirement. And xAI’s Gulf money came from the Qatar Investment Authority rather than from MGX, a distinction that matters when mapping which sovereign fund is exposed to which model developer.
The same institutions appear as investor, landlord and customer in the same transactions. That circularity is a structural feature of Gulf compute, not an anomaly.
Gulf artificial-intelligence capital is circular in a way that complicates both demand signals and risk assessment. HUMAIN invested three billion dollars in xAI’s Series E, announced in February 2026, in a round
that raised twenty billion dollars at a valuation of 230 billion. xAI is simultaneously the anchor customer of HUMAIN’s flagship. A separate agreement between the two parties covers a data centre above 500 megawatts, whose status, basis and commissioning date are all unverified. The investor is also the offtaker. That holds on both sides of the trade.
The pattern repeats at the asset level. HUMAIN anchors 36 megawatts of DataVolt’s 44 megawatt Riyadh East facility, more than 80 per cent of its planned capacity, supporting a package of about 500 million dollars with banks expected to fund roughly three quarters of it. DataVolt is wholly owned by Vision Invest, so the capital is private rather than government, and NEOM is landlord rather than compute buyer in the Oxagon project. A single anchor commitment is therefore doing the work of a financing covenant, and the anchor is the state-owned champion.
The Emirates presents the same structure through equity rather than tenancy. Microsoft holds a 1.5 billion dollar minority stake in G42 and is separately a customer of Khazna, a G42-linked operator, through a 200 megawatt expansion announced on 5 November 2025. Mubadala speaks publicly for Stargate UAE timing. Core42, also G42-linked, is both a sovereign cloud operator and a buyer of the chips its parent is named for in the Bureau of Industry and Security supplement. Each relationship is disclosed; the aggregate exposure is not disclosed anywhere in one place.
Qatar shows a cleaner but still related-party pattern. Syntys acquired Q Data QFZ LLC from Doha Venture Capital, the venture subsidiary of the
Qatar Free Zones Authority, which makes the transaction a state-linked asset transfer into an Ooredoo-carved-out vehicle rather than third-party acquisition. The price was not disclosed. Kuwait’s only artificial-intelligence facility is operated by that same Qatari group.
None of this is improper, and none of it is hidden. The analytical consequence is narrower. In a market where the investor, the landlord and the offtaker are frequently the same institution or its affiliates, a signed capacity commitment is weaker evidence of independent demand than it would be in a market with arm’s-length tenants. Demand should therefore be tested against third-party contracted capacity, of which Qatar holds the clearest examples, rather than against intra-group commitments.
Close to 95 per cent of Emirati capacity is delivered through colocation, and more than 85 per cent of hyperscaler power in the country sits inside someone else’s building.
Close to 95 per cent of total United Arab Emirates capacity is delivered through colocation models, and self-built data centres account for just 5.44 per cent of total capacity, approximately 65 megawatts. More than 85 per cent of hyperscaler information-technology power in the country is deployed via colocation rather than self-build. Except for Amazon Web Services, hyperscalers active in the Emirates operate and expand through leased space (DC Byte via Data Center Knowledge, 19 February 2026). These are the most useful structural numbers in the Gulf record.
Those two figures also yield the most useful reconciliation key in this read. Dividing 65 megawatts by 5.44 per cent gives a total tracked denominator of approximately 1,195 megawatts, roughly three times Knight Frank’s 376 megawatt live print. The two are not in conflict. DC Byte measures total tracked supply across all phases, while Knight Frank measures live information-technology load only. The live-to-total multiple in the Gulf’s most mature market is therefore about 3.2 times. That arithmetic is this
practice’s own, derived from two figures DC Byte published, and is labelled as such.
Three further structural facts follow from the same source. They bear on cost and siting. Emirati power costs approximately 0.11 dollars per kilowatt-hour, nearly double Saudi Arabia and more than double Qatar. Renewable energy accounts for only 5 per cent of the Emirati energy mix, and sustainability has not yet shaped Emirati site selection or design at scale. Energy consumption is not currently a binding constraint on Emirati growth, unlike Kuwait. The Emirates has, in effect, chosen speed and reliability over price.
Geography is shifting inside the country. Dubai remains the primary commercial hub, but land and zoning constraints in premium locations are
pushing the market outward. Abu Dhabi is the fastest-growing strategic region for sovereign and artificial-intelligence-grade capacity, clustering in KEZAD, Masdar City and the Mafraq Development Zone. Ajman is emerging as a secondary artificial-intelligence-ready node, and Fujairah matters for subsea connectivity. Cushman and Wakefield ranked Abu Dhabi third and Dubai eighth among primary markets in Europe, the Middle East and Africa in May 2026. Dammam ranked fourth and Riyadh sixth among secondary markets, with Doha second and Jeddah third among tertiary markets.
The implication for a party seeking Gulf capacity is that the counterparty is usually a landlord rather than a cloud. The capacity question is therefore a leasing question, and the relevant covenants are delivery dates, power pass-through and the landlord’s own connection position rather than a vendor’s regional roadmap.
The same physical plant is routinely counted twice: once by its owner and once by its operator. Two Gulf examples are documented.
Double counting through the ownership stack is the most common arithmetic error in Gulf capacity aggregation, and it occurs without any party misstating its own position. In Qatar, Syntys owns buildings and
counts 26 megawatts of live information-technology capacity, while Quantum Switch leases and operates those same buildings and counts 10 megawatts. Data Center Dynamics published a formal correction clarifying that Quantum Switch leases the buildings from Q Data and that the change of ownership does not change its operations. The 10 megawatts is almost certainly the same physical plant as part of Q Data’s 12.5 megawatts. The owner-level figure should be preferred. The two should never be summed.
In the Emirates the same error is available at larger scale. Hyperscaler capacity sits inside colocation operators’ numbers, with more than 85 per cent of hyperscaler information-technology power deployed via colocation, so adding a hyperscaler’s stated regional footprint to a colocation operator’s portfolio double counts the same halls. Within the G42 group specifically, Khazna’s stated 673 megawatt portfolio and Core42’s worldwide 150 megawatts of managed deployments must not be summed: both are group-linked and the overlap risk is real.
The cloud-region case is counterintuitive and needs stating plainly. Microsoft’s Azure Qatar Central region has been live since August 2022 with MEEZA across three availability zones in Doha. Qatar’s Ministry of Communications and Information Technology signed a cooperation agreement with MEEZA to expand that region by adding capacities of 4 and 6 megawatts. That implies both that MEEZA’s unnamed global hyperscaler is very probably Microsoft, and that Azure Qatar capacity must not be added on top of MEEZA’s published figure. The publication date of that
ministry item could not be verified, because the page is rendered in client-side script.
A related class of error is the multi-country portfolio figure attributed to a single state. Qareeb’s stated rollout of more than 50 megawatts spans seven countries. It is not Bahrain. O-Green’s target of more than 200 megawatts of artificial-intelligence-ready capacity by 2028 spans Oman, Spain and other markets and is not Oman. Syntys’ target of more than 120 megawatts of installed capacity by 2030 spans Qatar, Kuwait, Oman, Iraq and Tunisia and is not a Qatar figure. Gulf Data Hub’s claimed 90 megawatts under construction is group-wide, and no part of it should be attributed to Kuwait.
One entity error should be retired from the regional ledger entirely. Zankore, the Ooredoo, Indosat, NVIDIA and Nokia neocloud, targets Southeast Asia and principally Indonesia, with about 200 megawatts of initial capacity to be sourced from Indonesian providers for the first half of 2027. Counting it as Qatari sovereign compute is a category error.
The Kingdom publishes the region’s only government capacity series, on an undefined basis, and four official 2030 targets that no authority has reconciled.
Saudi Arabia reported 467 megawatts of operational capacity in the first quarter of 2026, up from 440 megawatts in 2025, a rise the Ministry of Communications and Information Technology framed as more than 6 per cent. The published series runs 68 megawatts in 2021, 440 megawatts in 2025 and 467 megawatts in the first quarter of 2026. It covers more than 60 data centres built by over 20 companies, with cumulative investment above 16 billion riyals, about 4.3 billion dollars (Ministry statement, 5 May 2026). The ratio of 467 to 440 reproduces the Ministry’s own growth figure, so the pair is internally consistent. Any source presenting 440 megawatts as current is one print behind.
The basis of that series is undefined, its scope is inferred rather than stated, and the gap matters. No definition, no load-versus-gross statement and no methodology note has been published. The scope appears to cover all owners, including colocation, enterprise and government self-build. Independent measurements of the same market show the resulting spread. Mordor Intelligence gives 410 megawatts of
information-technology load for 2025, S&P Global above 300 megawatts for 2025, and a tracker 290.5 megawatts for 2024. Arizton gives 222 megawatts of third-party colocation power for the first quarter of 2025, and the International Data Center Authority 440 megawatts in July 2026.
One independent primary measurement exists and should be used where an information-technology-load basis is required. KAPSARC gives 290.5 megawatts of installed information-technology load across 58 facilities in 2024, up from 122.4 megawatts in 2023. Its 2024 geographic split is Dammam 41 per cent, Riyadh 38 per cent, Jeddah 7 per cent and other locations 14 per cent (discussion paper DP00503, December 2025). Mordor at 410 megawatts of information-technology load against the
Ministry at 467 megawatts undefined is consistent with the Ministry counting gross power, or a wider owner set, or both.
The target set is unreconciled, and the contradiction is not minor. Alvarez and Marsal’s base case gives 1.0 gigawatt by 2030, KAPSARC describes 1.3 gigawatts as the national target, and the Ministry and SDAIA National Data Center Strategy gives about 1.5 gigawatts by 2030. The International Energy Agency attributes 1.9 gigawatts by 2030 and 6.6 gigawatts by 2034 to SDAIA, while the Fund and Aramco give 3.0 gigawatts by 2030. The HUMAIN, AMD and Cisco joint venture gives 1.9 gigawatts by 2030 and 6.0 gigawatts by 2034. Arizton gives 3.81 gigawatts of upcoming information-technology capacity with no date gate at all.
The unaddressed policy contradiction is that HUMAIN’s 2030 interim figure of 1.9 gigawatts exceeds the entire national 2030 target of 1.5
gigawatts by 27 per cent, from a single company. Either the national strategy was overtaken within a year of publication and has not been restated, or the company figure will not be met. No Saudi authority has addressed the gap. Both cannot be current policy. The official national number also moved by 46 per cent inside twelve months, from 1.3 to 1.9 gigawatts, which is itself a reason to show all vintages rather than a single target.
The hardest signed numbers in the Kingdom are a contractor’s stock-exchange disclosures. The softest is a gigawatt of telecommunications capacity that is not contracted.
Only a short list of Saudi rows carries evidence of energisation. The list is led by stc and center3 in Riyadh at 60 megawatts on a facility basis, with Mobily above 100 megawatts company-stated across eight sites, of which 39 megawatts is being built for artificial intelligence. HUMAIN’s GroqCloud Dammam inference region is at about 50 megawatts with service live. Google Cloud’s Dammam region has been generally available since 15 November 2023, and Oracle’s Jeddah and Riyadh regions date from 2020 and 2024. Aramco’s Dammam-7
high-performance computer at Dhahran is at 12 megawatts, and KAUST’s Shaheen III at Thuwal at 8 megawatts.
Two tracker figures must be discarded from that ledger. Figures of about 350 megawatts for Google Dammam and about 400 megawatts for Oracle come from a third-party tracker rather than from either company. Neither company publishes a Saudi capacity figure, and if both figures were real they would alone exceed the Ministry’s entire 467 megawatt national total. This is the clearest example in the Gulf dataset of tracker figures that cannot be reconciled with a national aggregate, and it is a reason to treat tracker totals for every state in this read with caution.
The hardest signed numbers in the Saudi set belong to a contractor. Al Moammar Information Systems disclosed to Tadawul on 23 December 2025 an award from HUMAIN to design and build a private artificial-intelligence data centre. It was valued at above 155 per cent of its 2024 total revenue including value-added tax, which on 2024 revenue of about 1.21 billion riyals implies roughly 1.88 billion riyals, about 501 million dollars. That figure is derived from a percentage-of-revenue disclosure rather than stated by the company. The award was expanded on 20 September 2026 from 50 to 250 megawatts, at above 689 per cent of 2025 total revenue including tax, about 8.76 billion riyals. It is executed through successive work orders issued by HUMAIN, with the third and fourth disclosed in October 2026.
The structural point about that contract is that 689 per cent of revenue is a ceiling rather than booked backlog, and the work-order structure is itself evidence of phasing. Other rows sit in the committed column. SDAIA’s Hexagon in Riyadh is at 480 megawatts, foundation stone laid in January 2026 to a Tier IV certified design with direct liquid and hybrid cooling, basis not stated. DataVolt at NEOM Oxagon has a 1,500 megawatt campus and a first phase reported variously at 300 and 360 megawatts for 2028. Sahayeb is at 24 megawatts initial and 120 megawatts expandable, one of only two Saudi rows in the ledger that states an information-technology-load basis. Equinix is at 100 megawatts in Jeddah, also explicitly information-technology load, in pre-development.
The most over-stated item in Saudi coverage is the stc arrangement. A memorandum of understanding signed on 18 December 2025 via center3 contemplates up to 1,000 megawatts, structured as HUMAIN 51 per cent and stc 49 per cent, with an initial phase of up to 250 megawatts depending on customer commitments. A Tadawul filing of 18 June 2026 disclosed a six-month extension rather than an executed agreement, citing outstanding regulatory and operational requirements, with stc stating no significant financial impact and expiry falling on or about 18 December 2026. One gigawatt of stc capacity is not contracted. It is also a single extension, so the characterisation of repeated delay is not supported by the filings.
Hyperscaler status divides cleanly between live and announced. Oracle in Jeddah and Riyadh, Google Cloud in Dammam and Alibaba Cloud in Riyadh through the Saudi Cloud Computing Company joint venture are live. The Amazon Web Services Saudi region is announced only, at 5.3 billion dollars announced on 13 May 2025 and stated as on track to launch in December 2026. It is the only hyperscaler Saudi capital figure verified to a primary vendor release. Microsoft’s Azure Saudi Arabia East was structurally complete in December 2024 with customer workloads targeted for the fourth quarter of 2026, a two-year gap between concrete and endpoint in one of the most capable operators in the market. The associated AWS artificial-intelligence zone, at up to 50 megawatts by 2028, is about 1.7 per cent of the three-gigawatt target and should not be presented as a meaningful capacity block.
The Emirates has the Gulf’s cleanest measured base, its largest announced nameplate, and the widest gap between the two.
The United Arab Emirates publishes, through Knight Frank, the only Gulf capacity figure accompanied by a published basis glossary. Live capacity surpassed 376 megawatts with 19.2 megawatts of new capacity coming online during 2025, against a national vacancy rate of 2.4 per cent, one of the lowest in Europe, the Middle East and Africa (Knight Frank, 14 October 2025). The basis is information-technology load, explicitly, and the scope is live colocation only, since Knight Frank defines aggregate supply separately as live plus planned. The 19.2 megawatt figure should be carried forward. It is the organic colocation growth rate.
Later prints move the base upward without changing its shape. ValuStrat reported more than 400 megawatts of operational colocation in July 2026, the largest base in the Gulf, alongside more than 30 billion dollars of announced artificial-intelligence and data centre investment. Arizton via Research and Markets corroborated more than 400 megawatts and over 45 per cent of total Gulf operational information-technology power on 27 October 2026. Knight Frank’s Data Centre Atlas 2026 gives existing capacity above 400 megawatts and a pipeline of about 1.8 gigawatts. The Knight Frank series should be preferred. It applies a single consistent methodology across two vintages.
Against that base sits the largest nameplate in the region. The five-gigawatt UAE-US AI Campus in Abu Dhabi is officially sourced but its basis is not stated. Only the first one-gigawatt cluster is specified. Of that, only 200 megawatts is dated to 2026. Stripping the campus out narrows the Emirati 2030 range to roughly 0.5 to 1.8 gigawatts, and that residual gap is purely live-only against all-phases measurement. The entire Emirati divergence is therefore attributable to one project, whose remaining four gigawatts has no phasing, no financing detail and no commissioning dates.
Three different 200 megawatt figures circulate for Abu Dhabi and are routinely conflated. The first is Stargate UAE phase one, developed by Khazna for OpenAI and Oracle, with chips not yet delivered. The second is a separate Microsoft 200 megawatt expansion delivered through Khazna
as part of Microsoft’s 15.2 billion dollar Emirati investment. The third is loose market commentary. Microsoft’s 1.5 billion dollar minority stake in G42 is a fourth and separate item. The campus first 200 megawatts and Stargate’s phase-one 200 megawatts appear to be the same tranche, but no source distinguishes them, so they must not be added.
Two features of the campus bear holding in view. The Emirates will finance or build equivalent facilities in the United States, which makes the arrangement reciprocal rather than straightforward inward investment, and the project draws on nuclear, solar and gas generation. The footprint claim of ten square miles appears only in secondary reporting and is unverified. Taken together with the masterplan revision described in chapter seven, the campus should be treated as a programme under active redesign rather than as a fixed quantity of capacity on a known date.
The largest Gulf operator declines to publish information-technology load, and the leading sovereign cloud publishes two incompatible chip counts.
Khazna provides the only real occupancy proxy in the Gulf, and it must be used carefully. Its first standalone environmental, social and governance report, published 6 October 2026, states that Khazna leads the Gulf market with more than 450 megawatts of contracted capacity. That capacity sits across 28 owned and operated data centres and two leased, 30 facilities in total, of which 24 are certified by the Uptime Institute. Its public website states a 673 megawatt portfolio with 30 live data centres and six ongoing projects. Four hundred and fifty of 673 is about 67 per cent of portfolio megawatts contracted, leaving roughly 223 megawatts uncontracted or under construction.
That ratio carries heavy caveats and should be read as a contracted-to-portfolio ratio rather than an occupancy rate. The two figures are different, undefined metrics from two documents, and the 673 may include capacity under construction. Khazna’s live portfolio is itself contested. S&P Global Market Intelligence reported 25 facilities and over 360 megawatts operating in October 2025, while secondary reporting in December 2025 claimed 30 live data centres totalling almost 650 megawatts. The
company’s own documents give the two figures above. The S&P 360 megawatts should be preferred as a sourced operating floor, 650 megawatts should be marked unverified, and no single live figure should be published as fact.
Khazna refuses to disclose information-technology load, and does so explicitly. Its report states that the company reports the metric for mechanical cooling load, and that information-technology loads are tracked and accounted for in clients’ consumption reports. The report adds that the mechanical load figure should not be read as total electricity consumption or total operational emissions. The hypothesis that power-usage-effectiveness reporting forces information-technology-load disclosure therefore does not hold in this market. Its pipeline is more than one gigawatt of
operational expansion over five years, announced on 14 October 2025. More than 400 megawatts of that is in international markets including Saudi Arabia and Italy, leaving under 600 megawatts of the increment domestic.
Core42 does not disclose how much compute is installed and energised as distinct from ordered. Its published metrics are all labelled worldwide rather than Emirati. They run to more than 150 megawatts of managed deployments, 76,000 graphics processing units described as deployed or planned, nine global data centres, 310 exaflops, 20 milliseconds to first token and more than ten billion tokens weekly. The same page, read on 4 November 2026, states in feature copy that customers may access more than 86,000 units across sovereign data centres globally. Two published figures from one page on one day differ by 10,000 units. Neither should be used as an installed count.
No per-operator Emirati megawatt figure could be obtained for Equinix, Moro Hub or G42 and Core42 individually. What is known at facility level is partial. Khazna’s QAJ1 in Ajman is at 100 megawatts across twenty five-megawatt halls with air-assisted direct-to-chip liquid cooling from day one. The owner itself moved initial-phase completion from a third-quarter 2025 target to December 2026, a slip of roughly fifteen months. Khazna AUH4 at Mafraq and AUH8 run to 60 megawatts combined, and Moro Hub’s solar-powered Warsan facility is above 100 megawatts planned, with phase two approved by DEWA in January 2026. The Microsoft and
du facility at 544 million dollars carries no published megawatt figure, and the capital figure itself appears only in a research table at study stage.
Qatar has the smallest credible installed base of the three large programmes and the only Gulf pipeline backed by executed debt and an awarded construction contract.
Defensible live Qatari capacity in 2026 is approximately 40 to 45 megawatts, built as MEEZA at 18 megawatts plus Syntys at 26 megawatts of live information-technology capacity. Both are operator-published 2026 figures, but on mixed bases: Syntys states information-technology load, while MEEZA’s basis is unstated and carries an at-full-occupancy qualifier, which means design or saleable capacity rather than energised and utilised load. The figure should be presented as a range rather than a point, and neither Quantum Switch’s 10 megawatts nor Azure Qatar should be added to it.
The cross-checks on each operator are good by regional standards. MEEZA’s 18 megawatts follows a four-megawatt expansion handed over on 21 June 2026 on signature of the ready-for-service certificate, nine months ahead of the contractual deadline, across more than 871,000 working hours with no lost-time incident. Its prior baseline of 14
megawatts across M-Vault 1 to 5 was stated explicitly as information-technology capacity. The sum of 14 and 4 reconciles. Syntys’ 26 megawatts follows the acquisition of Q Data QFZ LLC, which brought 12.5 megawatts of hyperscale capacity, 5 live and 7.5 under development, across two Tier III certified carrier-neutral facilities in the Qatar Free Zones. Group-level confirmation on the third-quarter call of 2 November 2026 gives Syntys 34 megawatts of active capacity across Qatar, Kuwait and Tunisia, with Qatar about 88 per cent of that.
Qatar’s pipeline is better evidenced than any other small Gulf state’s because it is financed. MEEZA’s M-Vault 6, 7 and 8 add about 44 megawatts, taking the company to more than 60 megawatts over three years from a 2026 base. The Um Garn campus is at 24 megawatts total, with its first 6
megawatts targeted operational by end-2027, and M-Vault 7 at 16 megawatts is in final design. Funding is a 1.6 billion riyal Commodity Murabaha facility from Dukhan Bank, about 439 million dollars, long-term and Sharia-compliant, agreed in August 2026 and additional to an existing facility. A construction contract above 500 million riyals, about 137 million dollars, was awarded to Elegancia Contracting and Industries at Qatar Science and Technology Park, announced via the Qatar Stock Exchange on 24 September 2026. No megawatt figure was stated in that award. None should be attached to it.
Forward contracting is the other distinguishing feature. MEEZA signed an eight-megawatt hyperscaler lease on 7 September 2026 above one billion riyals, about 270 to 275 million dollars, its largest contract to date, at a new campus. That is contracted capacity at a facility not yet built, a forward commitment rather than installed plant. Tenant identity and duration are undisclosed, as are pricing, the delivery schedule and the site. Earlier phases ran at four megawatts contracted in 2025 and delivered in June 2026, and six megawatts worth above 750 million riyals signed in October 2025. It is not disclosed whether the four, six and eight megawatt contracts are with the same counterparty; Microsoft is strongly implied but unconfirmed.
Forecasts for Qatar should come from the operators rather than from research houses. The research-house figures contradict each other. Arizton gives about 120 megawatts of utilised information-technology power by
2030, explicitly a demand-side utilisation measure implying a larger installed base, while a different Arizton report gives 27 megawatts of power capacity for 2031. The 27 megawatts is irreconcilable with MEEZA’s financed 60 megawatts alone. A separate 120 megawatt figure, Syntys’ target by 2030, is a MENA-wide number across Qatar, Kuwait, Oman, Iraq and Tunisia and is not a Qatar figure. No national Qatari megawatt target appears to exist in the public record at all.
That one facility is also the country’s only artificial-intelligence site. Ooredoo Kuwait announced the nation’s first artificial-intelligence-enabled data centre with NVIDIA on H200 accelerators on 9 November 2025 in Kuwait City. An important qualifier appears in the company’s own release. The H200 units had only just begun to be installed at announcement.
Competing figures for the same site give one megawatt on one directory and ten megawatts on a tracker. The investor deck figure should be preferred and the ten-megawatt figure should not be used.
Kuwait has two announced hyperscaler regions, one issuer-sourced megawatt figure, and three transformer stations that were not confirmed as contracted at the time of writing.
Kuwait is materially worse evidenced than any other Gulf state, and should be reported with explicit uncertainty or not at all. No operator-published megawatt figure could be verified for any Kuwaiti facility except one. The most solid data point is a facility count of five data centres at the start of March 2026. A separate directory-based account lists four, operated by Zajil Telecom, stc and Ooredoo. The only issuer-sourced megawatt figure is two megawatts, from Ooredoo’s own investor deck of November 2025. It is now a year old.
Everything else in Kuwait is unverified on megawatts. One tracker gives five megawatts for a Zain facility in Kuwait City and marks it verified. A secondary overview dated October 2026 states plainly that no reliable public source exists for the capacity of Zain’s Shuwaikh facility, and that Zain’s sustainability reporting contains no Kuwait-specific facility sheet. One
tracker against an explicit statement that no public source exists is not sufficient, and five megawatts should not be published. No figure was located for stc Kuwait and Qualitynet, Zajil Telecom, Kuwait Finance House or the National Bank of Kuwait. The regulator is not a capacity source. No licence register and no published capacity statistics exist.
The gating item is electrical. Kuwait’s Ministry of Electricity and Water is procuring three 132/11 kilovolt main transformer stations at Al-Mutlaa, Sulaibiya and north of South Saad Al-Abdullah City, at an estimated cost above 22 million dinars, about 71.5 million dollars. The tender deadline was set for 4 January 2026, postponed to 20 January and then to 10 February 2026, with a September 2026 report expecting contract award in the fourth quarter. As at 4 November 2026 the construction contract was not
confirmed as awarded. Kuwait’s hyperscaler regions cannot energise until those substations are built.
The hardest available Kuwaiti number illustrates the basis problem exactly. A government official stated that the three stations will have a total capacity of 72 megawatts and that Google requested 60 megawatts, 20 megawatts for each centre. That is grid or utility supply, explicitly, since these are transformer stations. The 60 megawatts is a requested supply figure, not information-technology load and not installed capacity, and the load behind it would be substantially lower. Google’s region remains under development, with land secured under a lease signed with the regulator in December 2024 at 12.6 million dinars over twenty years and approved by the State Audit Bureau. Microsoft’s announced Kuwaiti region, from a partnership of 6 March 2025, carries no timeline and no investment commitment. The 50 megawatt figure attached to it on one tracker is not a Microsoft figure.
Two further Kuwaiti items should be handled with care. Digitus DC, a subsidiary of Hayat Communications, provides colocation hosting of several unquantified megawatts for Google at Al-Mutlaa, potentially worth up to 50 million dollars. It is a bridge arrangement rather than the region itself, and the site is not operational. The one-gigawatt Omniva claim, originating about 2023 and funded privately, should be excluded from every pipeline total. Satellite imagery from 2026 shows work ongoing with most of the site undeveloped, and the company’s own website lists no active data
centres. Gulf Data Hub plans about three Kuwaiti facilities at around 80,000 square feet each with no megawatt figure, and no part of its claimed 90 megawatts group-wide should be attributed to Kuwait.
Bahrain’s installed capacity went backwards in 2026. Oman has the region’s most developed policy architecture and almost no operating compute.
Bahrain is the one Gulf market whose installed capacity fell during 2026. Its verifiable operator-disclosed colocation base on 4 November 2026 is in the low tens of megawatts. Batelco by Beyon at Hamala is at 2.75 megawatts across 250 racks, Tier III, live since 2021 and powered entirely by renewable energy following phase three of the Beyon Solar Park. A reported 3.0 megawatts is given for stc Bahrain in Manama, on a third-party aggregator listing of low confidence, and the Beyon White Space facility, Kalaam and Gulf Data Hub are at undisclosed megawatts. The commissioned White Space facility at Beyon Data Oasis, 6,000 square
metres, is leased long-term to and operated by Qareeb Data Centers and publishes no load figure.
The reason capacity went backwards is the Amazon Web Services write-off described in chapter seven. The megawatt size of me-south-1 was never published by Amazon. A 60 megawatt figure and a 70 megawatt national total appear on one tracker whose page returned empty on direct fetch, with no vendor, regulator or research-house confirmation. On third-party trackers Amazon was the majority of Bahrain’s tracked capacity, which is why the write-off matters. Even if 60 megawatts was once correct, it is not installed operational capacity in November 2026, and the 70 megawatt national figure should not be used.
Bahrain’s forward colocation pipeline is about 68 megawatts of information-technology load added across 2026 to 2031, on a third-party colocation scope that excludes hyperscaler, enterprise and government self-build and unfinanced artificial-intelligence campuses. The same publisher, on the same page and in the same edition, also gives power capacity of 14 megawatts for 2031 alongside a market size of 226 million dollars in 2025 rising to 393 million in 2031. The resolution adopted here is that 14 megawatts is an incremental addition in the 2031 reporting year and 68 megawatts is cumulative across six years. The 14 megawatt figure is internally consistent with the dollar and floor-area figures in the same table, and 53,000 square feet cannot host 68 megawatts of load. The publisher labels neither. Neither should be cited as a standalone national capacity figure.
Bahrain has no published national artificial-intelligence strategy. The Minister of Transportation and Telecommunications confirmed that his ministry was developing one at MEET ICT and BITEX 2026 on 27 to 28 October 2026. No official Bahraini capacity figure exists from the regulator, the Economic Development Board or the ministry, so the installed figure rests entirely on operator disclosures and commercial research. Oracle appears through Beyon Solutions’ Oracle Alloy sovereign arrangement with the Information and eGovernment Authority, announced in early November 2025 with capital, capacity, hosting location and go-live all undisclosed: that is a cloud-residency contract, not artificial-intelligence compute. There is no Microsoft or Google region in the country.
Oman’s installed base is roughly 10 to 20 megawatts across about eleven third-party colocation facilities, mostly Tier III, across seven locations. The only operator-attributed figure is Oman Data Park’s total capacity of over 10 megawatts, with a stated plan to double to 20 megawatts in five to ten years. That figure is sourced from a trade-body republication of a company interview rather than an audited disclosure. The operating entity has changed name and ownership. Oman Data Park was rebranded Otech in early 2026 with Omantel taking full ownership, and the Omani government holds 51 per cent of Omantel. Equinix launched SN1 in Salalah with Omantel in November 2024 with no megawatt figure published. Claims that Omani cloud campuses have been upgraded with H200 accelerators rest on a single stale third-party source and are marked unverified.
Oman’s announced pipeline exceeds its installed capacity by roughly thirty to fifty times and none of it is financed. A Mubbadrah-led consortium signed a framework agreement for up to 400 megawatts at Salalah at three billion dollars during the Sultan’s United Kingdom state visit in late October 2026. The consortium’s own account discloses no committed financing and no electricity supply arrangements. It gives no construction timetable, no first-phase capacity and no load basis. A separate 400 megawatts of solar and storage in the same project is generation and must not be double counted. Yamna signed a strategic cooperation agreement with the ministry for a first phase stated above one billion pounds with no capacity figure of any basis, no tenants and no timeline. A
150 megawatt project was announced in May 2026 by an Italian-Omani group with no site, financing or commissioning date. O-Green’s target of more than 200 megawatts by 2028 spans Oman, Spain and other markets and must not be attributed to Oman. There is no Duqm artificial-intelligence or data centre project. Omani activity is in Salalah and in the Seeb AI Special Zone under Royal Decree 50/2026.
The Omani policy architecture advanced again during 2026, and still with no megawatt attached to it. OPAZ confirmed through its quarterly Duqm Economist that the artificial-intelligence zone will occupy about 104,000 square metres in the Wilayat of Seeb, beside the Civil Aviation Authority building near Muscat International Airport (Oman Observer, 2026). That was the first official word on the site. The authority's board has been told to appoint a developer in coordination with the Ministry of Transport, Communications and Information Technology. The implementation roadmap was still being drafted at the time of writing. Data centres appear on the target-sector list alongside semiconductors and robotics. A target list is not a load.
A private operator has now been licensed, which is further than any Omani capacity figure has travelled. Afouq Investment and Development United said in early May 2026 that it held a licence to develop and operate a plot of roughly 105,000 square metres inside the zone. It had been entrusted at Comex 2025 with a first phase of 100,000 square metres carrying about 100 million rial of envisaged inflows alongside Egypt's Prime Group (Oman
Observer, 2026). The licence names semiconductors and quantum computing as the sectors it will court. It names no compute at all. The five-year plan expects the digital economy to reach about 4.2 per cent of gross domestic product, which is a share of output rather than a plant.
Bahrain's forward case is better evidenced in dollars and square metres than in load, which is the same problem in another currency. The kingdom's Economic Development Board describes stc Bahrain building a park above 55,000 square metres within a 300 million dollar investment. Batelco is working with Qareeb Data Centres under a 250 million dollar commitment on a Manama site above 140,000 square metres (Arabian Gulf Business Insight, 2025). Neither of those is a megawatt. The same account carries a third-party forecast of about 66 megawatts of information-technology power added across 2024 to 2029, and an official expecting tens of megawatts for data centres within eight months of October 2025. That window has since closed with no energisation figure published.
The low family should be preferred: about 3.3 gigawatts across the Gulf by 2030, stated as a range, on an information-technology-load basis, with every component dated.
The preferred estimate is a Gulf total of roughly 3.3 gigawatts by 2030. It is composed of Saudi Arabia at about 1.5 gigawatts, the United Arab Emirates at about 1.2 to 1.8 gigawatts, and the other four states at about 200 to 300 megawatts combined. It should be stated as a range, on an information-technology-load basis, with each component date-stamped to its publisher. That figure is consistent with FTI Consulting via Arab News on 21 August 2025, and with ValuStrat’s expectation that capacity triples by 2030. It also matches Knight Frank’s regional trajectory of 1.2 gigawatts at end-2025, 2.2 gigawatts in 2026 and 3.2 gigawatts in 2027.
Five reasons support that preference, in order of weight. First, it is the only family anchored on a measured base rather than on announcements. Saudi Arabia has a dated government series and the Emirates has a basis-defined live print, while the high family has no base at all. Second, the run-rate test disciplines it. The national Saudi target already requires 218 megawatts a year sustained for five years against roughly 94 delivered, and the 2034 pipeline requires about 700. Third, the independent audit of
thirteen Saudi and Emirati cases found no project at the firm-connection or energised gate. Fourth, the grid-connection record is absent in exactly the places the high family depends upon. Fifth, two independent constructions of the low family agree, while the high family reconciles with nothing, including itself.
The per-state spreads show where the divergence lives. Saudi Arabia runs from 1.0 to about 4.3 gigawatts, a spread of 4.3 times, which represents three or four different objects rather than three estimates of one. The Emirates runs from about 0.47 gigawatts to above 5 gigawatts, a spread above ten times, almost entirely attributable to one campus. Qatar runs 27 to 120 megawatts, a spread of 4.4 times created by one publisher contradicting itself. Kuwait runs 19 to 90 megawatts on a basis difference
between grid supply and load. Bahrain runs 14 to 70 megawatts, where the high figure is simply dead. Oman runs 18 to above 550 megawatts, a spread above thirty times, which is the announced-against-financed problem in its purest form.
Secondary preferences should be applied by use case. For installed capacity, prefer Knight Frank for the Emirates as the only source with a published basis glossary, the Saudi ministry for the Kingdom with an explicit undefined-basis caveat, and operator disclosures for the four small states. For financed pipeline, prefer MEEZA’s Qatari figures, which are the only Gulf numbers backed by executed debt and an awarded construction contract. For anything labelled 2030 artificial-intelligence capacity,
prefer nothing: report the announcement with its date, its publisher and the words announced, not financed.
If a single conversion key is unavoidable, the live-to-total-tracked multiple of about 3.2 times should be used. Applied in reverse to a twelve-gigawatt announced pipeline it implies about 3.7 gigawatts eventually live, which lands almost exactly on the 3.3 gigawatt estimate. The two families are reconcilable and the reconciliation factor is about three.
The high family is poorly evidenced, not impossible. Four arguments can be made for it, and each is testable.
The case for the high estimates deserves to be stated at its strongest. The first argument concerns disclosure. The audit evidence is about disclosure rather than delivery, and the author of the thirteen-case audit states explicitly that it does not establish that any project is cancelled, delayed, unfunded or unable to secure power. A state-owned developer with sovereign-fund backing has little commercial reason to publish connection terms. Absence of published evidence is therefore weaker evidence of absence here than it would be in a listed-utility market.
The second argument is that capital is genuinely present and in motion. Aramco contributed about 9.9 billion riyals in cash in 2026 alone, and
MEEZA has executed a 1.6 billion riyal facility and awarded construction. HUMAIN has issued successive work orders to a listed contractor against a ceiling of about 8.76 billion riyals. The Emirates has more than 30 billion dollars of announced artificial-intelligence and data centre investment, and Khazna has committed to more than one gigawatt of expansion over five years. Capital of that order does not guarantee delivery, but it is not a press release either.
The third argument is that the region is adding capacity in increments far larger than those reflected in historical run rates. Khazna’s Ajman facility is a single 100 megawatt site. Al Moammar’s award moved from 50 to 250 megawatts in nine months, and SDAIA’s Hexagon is 480 megawatts in one building. If even a minority of such items energise on schedule, the delivered rate steps up discontinuously rather than continuing a four-year trend line. A fourth argument follows. Demand appears to be real. Emirati vacancy was 2.4 per cent in 2025, among the lowest in Europe, the Middle East and Africa, and Khazna reports about 67 per cent of portfolio megawatts contracted.
The widest comparative study in the market does not expect oversupply either. Cushman and Wakefield, across 107 markets, judges that an artificial-intelligence infrastructure bubble appears unlikely because structural frictions constrain overbuilding, while simultaneously finding that the proportion of realistically deliverable pipeline in Europe, the Middle East and Africa is declining. Those two findings together describe a market that
under-delivers against its pipeline without collapsing, which is the central expectation of this read.
Each of those arguments is testable, and the tests should be run on a known calendar. Certain items would move this read towards the high family. They are a partner commissioning release for the Abu Dhabi campus phase one, a HUMAIN announcement of energisation at Riyadh or Dammam, and award of the Kuwaiti substation contract. The others are a published Saudi connection queue or a defined national capacity basis, and an executed stc and HUMAIN joint venture before the memorandum expires on or about 18 December 2026. None had occurred as at 4 November 2026.
Five risks are documented rather than speculative: power, policy reversibility, physical security, circular demand and the expiry of entity authorisations.
The first risk is power. It is already binding in two states. Kuwait cannot energise its hyperscaler regions until three transformer stations are built. That contract was unconfirmed as awarded at the time of writing, with no major new Kuwaiti generation expected before 2028. Saudi Arabia publishes no connection queue and its official statistics do not
disaggregate data centre consumption at all, so the delivery schedule of its pipeline cannot be externally verified. The largest item in that pipeline is evidenced by a bulk supply point tender.
The second risk is policy reversibility. The AI Diffusion Rule sits in the Code of Federal Regulations, unenforced by executive discretion and reactivatable without new rulemaking, and no statutory constraint was enacted to replace it. The third is the expiry built into the one country framework that exists. The authorisation for G42 and Core42 expires automatically on 6 April 2027 unless those entities become United States companies, in which case they must apply through the standard licensing process. Both risks are political rather than physical, and neither is within the control of any Gulf operator.
The fourth risk is physical security, which entered the column in 2026 and has already destroyed capacity. Three facilities were damaged and one regional footprint was written off with permanent data loss. The operator advised clients to consider migrating workloads out of the region. The largest Emirati masterplan is being reconsidered, with air defences and partly underground construction under discussion. The cost and
timing consequences of that redesign are not determinable from the public record.
The fifth risk is the circularity of demand. Where the investor, the landlord and the anchor tenant are the same institution or its affiliates, a signed commitment carries less information about independent demand. HUMAIN invested three billion dollars in xAI while xAI anchors HUMAIN’s flagship; HUMAIN anchors more than 80 per cent of a third party’s financed facility; Microsoft holds equity in G42 and leases from its affiliate. None of this is improper and all of it is disclosed, but a market whose demand is substantially intra-group is more exposed to a single sponsor’s change of plan than its contracted-capacity figures suggest.
A sixth item is not a risk but a disclosure gap with the same practical effect. The largest operator in the region declines to publish
information-technology load. The leading sovereign cloud publishes two chip counts differing by 10,000 units on the same page on the same day. The flagship Saudi vehicle publishes nothing machine-verifiable and has twice declined to disclose locations or deliveries. A market that cannot be measured cannot be underwritten on its own figures, whatever the underlying physical position.
The practical instruction for any buyer, lender or landlord is to replace masterplan language with energisation dates and stated measurement bases.
The single most useful operating rule in this market is to contract against energisation dates rather than campus masterplans. That formulation belongs to Data Center World Middle East, and the reasoning behind it is the whole of this read. Usable capacity is what has been built, energised and cooled, and then certified and connected. New capacity arrives in increments of 100 to 200 megawatts. Those increments are gated by grid interconnection and generation rather than by land or capital.
A second rule follows for anyone quoting a figure. Require the basis in writing. Of seven audited headline projects, one defined information-technology load. Two Saudi rows in the entire national ledger state their basis, being Sahayeb and Equinix Jeddah. Knight Frank and Arizton publish definitions and should be preferred for that reason alone. Where a basis is not stated, the figure should be carried with the words basis not stated, which is what this read does throughout.
A third rule concerns counterparty concentration. Khazna, stc and Gulf Data Hub together represent more than half of total Gulf information-technology power capacity, and close to 95 per cent of Emirati capacity is
delivered through colocation. A buyer contracting for Gulf capacity is therefore usually contracting with a landlord whose own position depends on a utility connection and, in one case, on an entity authorisation expiring in April 2027. Those dependencies should appear in the contract as conditions, not as assumptions.
A fourth rule concerns resilience, and it is new. Single-region deployment in this market now carries a documented precedent of permanent data loss. Amazon Web Services was unable to restore resources and data hosted exclusively in one Gulf region, and damage exceeded what regional and multi-zone services were designed to withstand. Multi-region architecture, with at least one region outside the Gulf, has moved from best practice to a minimum standard on the evidence of 2026.
The second absence is the operational status of HUMAIN’s Riyadh and Dammam artificial-intelligence factories. Both broke ground in August 2025. A second-quarter 2026 go-live was stated. The quarter closed with no launch announcement, and independent trackers still class both as under construction. No public evidence confirms that they opened, and the company has twice declined to disclose locations or chip deliveries. The third absence is any partner commissioning release for Stargate UAE
or the Abu Dhabi campus phase one, confirmed by reading the lead partner’s newsroom index current to 2 November 2026.
The largest gaps in this market are absences rather than disagreements, and several of them sit at the centre of the largest programmes.
What could not be established matters here as much as what could. The largest absence is the basis of Saudi Arabia’s national capacity series. No definition, no load-versus-gross statement and no methodology note accompanies the 68, 440 and 467 megawatt prints. The choice between information-technology load and gross facility power changes every derived Saudi figure by 20 to 50 per cent. That gap sits beneath the region’s largest programme. It sits beneath every target in circulation.
The fourth group of absences is national. No official capacity figure exists for Bahrain or Oman from any regulator, investment authority or ministry, so both national figures rest entirely on operator disclosure and commercial research. No Qatari national megawatt target appears in the public record. Kuwait’s regulator maintains no licence register and publishes no capacity statistics, so far as could be established. Four of six Gulf states therefore cannot be measured from official sources at all.
The fifth group concerns figures that exist but should not be used. The 500,000-chips-a-year Emirati figure was never formalised. The one-gigawatt Omniva claim in Kuwait has no operating plant. The 70 megawatt Bahrain national total is dominated by a written-off region. A 10 megawatt tracker figure for Kuwait’s artificial-intelligence facility contradicts the only issuer disclosure. A 350 megawatt Google Dammam figure and a 400 megawatt Oracle Saudi figure, if both were real, would exceed the entire Saudi national total. A reported Iranian strike on the Bahrain site in July 2026 could not be confirmed and is not used.
The final observation is methodological. Several of the conflicts in this dataset are internal to a single publisher. The same research house gives
27 megawatts and 120 megawatts for Qatar, 14 megawatts and 68 megawatts for Bahrain, and 18 megawatts and about 80 megawatts for Oman, with two different Emirati figures published on the same day. Where a publisher contradicts itself, operator and government disclosures should be preferred. Where those are absent the honest answer is a range with its reason stated.
The Gulf will build a substantial compute base by 2030 and will not build the one it has announced. The gap is about three times and it is mostly power.
The Gulf energises of the order of 3.3 gigawatts of information-technology load by 2030 on the central expectation here, against about 880 megawatts operational in 2026 and about twelve gigawatts announced. That is roughly a quadrupling of the installed base in four years, which would be a substantial industrial achievement by any measure, and
it is also roughly a quarter of what has been announced. Both statements are true at once, and most commentary on this market picks one of them.
The distribution is expected to remain highly concentrated. Saudi Arabia and the Emirates held more than 85 per cent of operational capacity in 2026, and nothing in the evidence suggests that share falling. Qatar’s financed pipeline takes one operator above 60 megawatts, while Kuwait’s depends on three unawarded substations. Bahrain’s pipeline is colocation-only and its base has fallen, and Oman’s announcements are unfinanced. The four small states together are expected to hold 200 to 300 megawatts in 2030, against roughly 75 to 105 megawatts today.
Three dates will test this read. All fall before the end of 2027. The stc and HUMAIN memorandum expires on or about 18 December 2026, which will
show whether a gigawatt of telecommunications-linked capacity converts into an executed joint venture. The Kuwaiti substation award, expected in the fourth quarter of 2026, determines whether two hyperscaler regions can energise at all. The entity authorisation for G42 and Core42 expires on 6 April 2027. That date will show whether the Emirati champion relocates its domicile, obtains fresh notice, or moves to case-by-case licensing.
Two further dates matter for the installed record rather than for policy. Amazon Web Services promised a further update on the Bahrain and Emirati loss positions in early 2027, which will determine whether any of that capacity returns. And the Amazon Saudi region, stated as on track for December 2026, together with Microsoft’s Saudi region targeted for the fourth quarter of 2026, would mark the first material hyperscaler self-build capacity in the Kingdom if both land on schedule.
The structural conclusion is that the Gulf is building a colocation and inference market with sovereign branding, not yet a frontier training market. Live training capacity in Saudi Arabia is effectively nil; the operating base is colocation, two public cloud regions, two research clusters and one inference region. The Emirates delivers close to 95 per cent of its capacity through colocation and more than 85 per cent of its hyperscaler load through leased space. That is a real and investable market. It is a different market from the one described by the announcement record, and it is gated by the grid.
Nine conclusions for the people who must act on this market.
Nine conclusions survive the evidence, and they are not the nine that most Gulf commentary offers. First, the headline divergence resolves. The factor is about three. It is the live-to-total-tracked multiple derived from Emirati data, so the two families of estimates describe one market at two gates. Second, the region’s measured operational base in 2026 is of the order of 880 megawatts, with Saudi Arabia and the Emirates holding more than 85 per cent of it and three operators holding more than half. Third, every published Saudi 2030 target requires between two and five times
the delivered rate of the past four years, and the 2034 pipeline requires six and a half times.
Fourth, the only Gulf pipeline backed by executed debt and an awarded construction contract is Qatari, which is a reversal of the usual regional hierarchy and makes MEEZA’s figures the most reliable forward numbers in the dataset. Fifth, the largest single unresolved item in the region is whether HUMAIN’s Riyadh and Dammam factories opened: a stated second-quarter 2026 go-live passed without announcement and both remain classed as under construction. Sixth, the entire Emirati divergence is attributable to one campus whose masterplan is being revised for physical-security reasons, and whose phase one has no commissioning release from any of six named partners.
Seventh, four of six Gulf states cannot be measured from official sources at all, so their figures rest on operator disclosure and on research houses that in several cases contradict themselves within a single publication. Eighth, the demand signal is partly circular. The investor, the landlord and the anchor tenant are frequently the same institution or its affiliates, which means contracted capacity carries less information about independent demand than it would elsewhere. Ninth, the practical instruction that follows from all of the above is narrow and operational. Contract against energisation dates rather than campus masterplans, and require the measurement basis in writing.
The engineering work the sector asks for, and the form it takes on an engagement.
Sovereign compute is a power and property programme before it is a silicon one, so the engineering it needs runs from the substation to the model. The work below is what operators, state programmes and their anchor tenants in this market commission. Each is delivered by senior engineers working inside the client's own team, with the Global Delivery Centres running what has been built once it is in production.
An operator or anchor tenant in this market should hold two documents before committing capital. They are the signed grid connection offer, with its energisation date, and the tenancy terms that say who gets capacity when the hall is oversubscribed. Everything else can be engineered later. Capacity stated in announced megawatts is not capacity, and Volge's first
task on most of these programmes is to restate a pipeline in contracted and energised terms. That generally yields a smaller number than the one in circulation.
Staffing follows the sequence of the build rather than the shape of the organisation. Power and property questions dominate the early months, so the lead is an engineer who has read a connection agreement and understood what it does not promise. Platform and tenancy questions dominate later. Governance is a fortnightly reconciliation of the connection date against the hall handover, and of both against cluster acceptance, because the connection agreement sets the critical path and the procurement plan does not. Slippage compounds in one direction only, so reporting that hides it buys nothing.
The first ninety days produce a reconciled capacity ledger and a workload placement policy, with little else of substance, which is the right ambition for the quarter. Measurement afterwards is utilisation against contracted capacity, and cost per unit of useful compute rather than per accelerator installed. The failures are predictable. Programmes order silicon against a hall that is not yet energised, and they run a shared national asset on single-tenant habits until the second tenant arrives and the scheduling argument begins.
Every body cited in this read for a capacity, chip or contract figure, with the year the text attributes to it.
Volge is an engineering firm providing enterprise technology transformation services to clients across the Middle East, India, APAC, Europe, and North America.
Delivery stands on two pillars: Forward Deployed Engineering, with senior specialists embedded in the client’s own team, and Intelligent Command Centre Operations, run from our Global Delivery Centres, which keeps what they build running. Senior engineers stay with an engagement from the first diagnosis through to production.
Our services span Strategy, Data, Platforms, AI and Automation, Cyber Trust, Technology Operations, and Applied Functional Engineering. Engagements begin with a diagnosis. The firm is built on working autonomy: engineers decide, innovate, co-create, and own their work. Motivated engineers build the systems that sustain our clients’ growth.
This edition may be shared with organisations evaluating a Volge engagement. It may not be altered or republished without Volge’s permission, and figures carry their stated basis and date. Volge, Intelligent Command Centre, and Intelligent Command Centre Operations are marks of Volge; other names are the trademarks of their respective owners.