亞太商訊
2026-11-17 10:00:00

Abu Dhabi's 1 GW AI Campus Needs More Power Than the Country Forecast

HONG KONG, Nov 17, 2026 - (ACN Newswire) - The first 200 MW of Stargate UAE are supposed to be live this year. The site sits inside a 26 square kilometre campus in Abu Dhabi, built by G42 through its subsidiary Khazna, operated with OpenAI and Oracle, and fitted with roughly 100,000 Nvidia Grace Blackwell GB300 chips. The partners put the investment at up to $10 billion for a 1 GW cluster.

That cluster is the easy part to describe. The hard part is the meter.

A megawatt tells you how fast a facility draws power. It does not tell you how much electricity it will consume in a year, and it does not tell you whether the grid can deliver that power at three in the morning in August, when the air conditioning load across the emirate peaks and the solar farms are dark. Those two questions decide whether the promised capacity arrives on schedule, and how much new generation Abu Dhabi has to build around it.

The public numbers on both sides are already in circulation. They are rarely put in the same sentence.

The 200 megawatt number hides the annual bill

Start with the chips, because they set the floor. TrendForce estimated that a GB300 system with 72 chips draws about 140 kilowatts. At that rate, the 100,000 chips announced in May 2025 work out to roughly 1,400 systems and about 196 MW of information technology load, which is why the partners landed on 200 MW for the first phase in 2026.

Convert the IT load to what the utility sees. Modern desert builds in the UAE are designed at a power usage effectiveness of about 1.18, so a 2026 design of 200 MW of chips becomes roughly 236 MW at the fence line. Run that for a full year and the phase consumes about 2.1 TWh.

Two point one terawatt hours is not a rounding error in this market. Wood Mackenzie estimated that every data centre in the UAE consumed about 3 TWh in 2025, roughly 2% of national electricity demand of 173 TWh. The first phase of one project, running flat out, would add two thirds of the entire country's existing data centre consumption.

Scale it up and the arithmetic gets louder. The full 1 GW Stargate cluster becomes 1.18 GW at the fence and about 10.3 TWh a year. The 5 GW campus announced in 2025 becomes 5.9 GW and about 51.7 TWh a year, close to 30% of the country's total 2025 demand.

These are ceilings, not forecasts. Training clusters do not run at 100% load every hour, and no operator signs up for that. Cut the utilisation to 70% and the 1 GW cluster still needs about 7.2 TWh a year. Cut it to 50% and it needs 5.2 TWh, which is more than the 6 TWh that the whole national data centre estate is projected to consume in 2030.

Table 1: Load converted to annual energy at an assumed design PUE of 1.18 and full utilisation

Source: Wood Mackenzie via Data Center World Middle East 🔗, with conversions calculated from published capacity figures

The last row is the one to keep in view. The country's entire data centre fleet used 3 TWh last year. Its flagship AI project is designed to use several times that on its own.

Two forecasts, one grid

In March 2026, Wood Mackenzie put UAE data centre demand at just over 6 TWh by 2030, more than double the 2025 level. EWEC, the Abu Dhabi utility that procures the emirate's power and water, has used a similar figure in its own planning work.

Read that forecast next to the campus arithmetic and the two numbers do not fit. The 6 TWh estimate was built on colocation demand: the racks that banks, telcos and cloud regions rent today, plus growth in the same shape. A 5 GW AI campus is not a bigger version of that business. It is a single customer whose load arrives in 200 MW blocks and whose utilisation curve is set by GPU replacement cycles, not by office hours.

The difference matters for anyone selling into this market. If you plan against 6 TWh, you size transformers, chillers and backup generation for a country that adds roughly 3 TWh of new data centre load over five years. If you plan against the campus pipeline, you are sizing for a single site that could pass that number before 2030.

Both numbers can be right at the same time. The forecast describes the market that exists. The campus describes the demand that has been announced but not yet energised. The gap between them is the thing nobody has published: how much of Stargate's 1 GW is already inside the 2030 figure of 6 TWh, and how much sits on top of it. My reading is that the 6 TWh is a colocation forecast with an AI campus stapled to it, and that the first revision will come from the utility rather than the researchers.

Larry Ellison, Oracle's chief technology officer, described the intent at launch: "This first-in-the-world platform will enable every UAE government agency and commercial institution to connect their data to the world's most advanced AI models." That is a statement about demand, not about supply. There is a clue in the way the project is being built. Khazna, the G42 data centre arm that holds about 60% of the UAE's commercial wholesale capacity, is the builder. Its own pipeline target is 1,000 MW. Stargate's first phase, due in 2026, would absorb a fifth of that. When the same company is both the landlord for the national forecast and the contractor for the campus, the two demand curves get blended in a boardroom long before they reach a public statistic.

Firm power is the scarce item, not sunshine

Abu Dhabi has an unusual starting position for a Gulf AI hub. It is the only country in the Middle East with an operating nuclear fleet. Barakah supplies about 21% of UAE electricity output, and EWEC lists 4,170 MW of Barakah capacity in service with another 1,390 MW to come. Solar contributed about 9% of generation in 2025, with official projections taking that above 20% by 2040.

Neither source solves the problem on its own. Nuclear runs flat and cannot follow a training cluster that ramps when a batch job starts. Solar runs for the middle of the day and stops. The campus needs both energy and firm capacity, and the firm capacity is what costs money.

The procurement list shows how the emirate is filling that gap:

Taweelah C, a 2.6 GW combined cycle gas plant at the Al Taweelah complex, with Siemens Energy supplying three SGT5-9000HL gas turbines and two steam turbines. TAQA leads it with Al Jomaih and Sembcorp, China Energy Engineering is the EPC contractor, and commissioning is expected in late 2028.The round the clock renewable project developed by Masdar with EWEC, which pairs 5.2 GW of solar with 19 GWh of battery storage to deliver 1 GW of firm output. Financial close came in July 2026 at about $6.1 billion, with operations targeted for 2027.Barakah's remaining units, which add baseload without new fuel logistics.Put the gas plant next to the campus and the scale becomes concrete. A 2.6 GW combined cycle plant running at a 60% capacity factor from 2029 produces about 13.7 TWh a year, which is more than the full 1 GW Stargate cluster needs even at full load. One power station, built for grid flexibility and carbon capture readiness, covers the flagship cluster with room to spare.

The battery project is the more interesting bet, because the arithmetic is tight. Nineteen gigawatt hours of storage against 1 GW of promised firm output is 19 hours of delivery at full rate. That is enough to carry a night. It is not enough to carry a week of heavy dust or a gas outage, and it explains why the gas plant is still in the plan.

What the campus buys before it buys chips

The conventional way to read a project like this is as a chip order. The operational way is as a procurement queue, and the first items on it are not silicon.

A 200 MW phase needs high voltage substations, transformers, switchgear and chillers before a single rack is installed. Those items have lead times measured in years, not months, and they compete with every other Gulf megaproject for the same factory slots. G42 said in October 2025 that civil, structural and mechanical work was ahead of schedule and that all long lead equipment had been procured. That sentence is the most informative line in the whole announcement, because it tells you the constraint the company was managing.

Cooling is the second queue. A design PUE of 1.18 in a desert climate implies liquid cooling at the rack and closed loop heat rejection, not evaporative cooling that drinks water. That changes the equipment list: coolant distribution units, rear door heat exchangers, dry coolers sized for 45 degree ambient air. It also changes the maintenance contract, because a coolant leak in a 140 kW rack is a fire risk, not an inconvenience.

The third queue is people. A campus of this size needs mechanical and electrical technicians, network engineers and security staff on site around the clock, in a country where the data centre workforce is thin. The regional buildout is adding capacity faster than it is adding trained staff, and that gap shows up as higher operating costs and longer commissioning.

For suppliers, the sequence matters more than the headline number. In the 2026 phase, the first $1 billion goes to switchgear, transformers, cooling and construction. Chips are the last thing to arrive and the first thing to be quoted.

The licence that came before the concrete

Stargate UAE exists because Washington changed a rule. The Biden administration had restricted advanced AI chip exports to countries including the UAE. The Trump administration rescinded that rule in 2025, and the Commerce Department set up a bilateral working group to check that the project met US security standards.

That sequence explains the shape of the deal. The UAE side committed to reciprocal investment in US AI infrastructure, and the US side opened the export channel. The campus is the visible half of a bargain whose other half sits in American data centres.

The licence also sets the pace of the chip pipeline. Microsoft's experience is the clearest example: a September 2025 export licence authorised 60,400 A100-equivalent units, on top of the 21,500 it had already deployed. Every one of those units has to be ordered, shipped, racked and cooled before it earns anything.

For anyone modelling the project, the lesson is that the constraint moves. In 2025 it was the export licence. In 2026 it is long lead electrical equipment and grid connection. In 2028 it may be the gas turbine queue that is already stretching into the early 2030s in other markets.

Khazna is the bottleneck and the beneficiary

Khazna's position explains a lot about how this market actually works. It builds the campus, it holds the majority of wholesale colocation capacity, and it is the delivery vehicle for Microsoft's 200 MW expansion. When one company is the landlord, the contractor and the operator's local partner, capacity gets allocated by relationship as much as by price.

The Microsoft numbers show the pattern. The company's cumulative UAE commitment has reached $15.2 billion, including $7.9 billion for 2026 to 2029, of which $5.5 billion is AI and cloud infrastructure. A separate $544 million hyperscale site in Dubai is being built with du, and AWS has committed $1 billion with e&.

Add Stargate's 1 GW and the colocation market forecast starts to look conservative. RAX puts the UAE colocation market at $1.39 billion in 2026, rising to $3.90 billion by 2030 at a 21.7% compound annual growth rate. That forecast was built on leased racks sold by the megawatt. It does not model a single tenant taking 1 GW of dedicated capacity, because until this year there was no such tenant.

The other side of the same coin is pricing power. With 400 MW of colocation capacity recorded in 2026 and a pipeline target of 1 GW, Khazna can price scarcity. Tenants who need capacity in 2027 are not negotiating against a competitive market. They are negotiating against a queue.

Connectivity softens the picture a little. The UAE hosts 19 international submarine cables, with fibre to the home at 99.5% coverage in 2024. The country is a genuine transit point between Europe, Asia and Africa, which is why hyperscalers keep landing here rather than in a cheaper neighbour.

The free subscription that makes curtailment expensive

One detail in the Stargate package sits outside the engineering and inside the politics. The UAE became the first country to offer ChatGPT to its residents nationwide, and the project carries a commitment to provide ChatGPT Plus to citizens and residents at no charge. The service lists at $20 a month.

That commitment changes the cost of failure. A delayed commercial data centre annoys its tenants. A delayed national AI platform that every resident has been promised annoys everyone, and it does so in public. The operators have an incentive to energise whatever capacity exists, even at poor utilisation, rather than wait for the optimum.

It also gives the project a social return that is easy to defend in budget negotiations. The electricity bill for a 1 GW campus is large, but it is being spent inside a country that treats AI capacity as public infrastructure, the same way it treats desalination and the nuclear fleet. That framing is why the gas plant, the battery project and the campus are moving on parallel tracks rather than in sequence.

The risk sits in the sequencing. Taweelah C is not expected to commission until late 2028. The round the clock renewable project targets 2027. The first 200 MW is meant to be live this year. For at least one summer, the campus will be drawing on a grid whose new firm capacity has not arrived yet, backed by a nuclear fleet that is already committed and a solar fleet that stops at sunset.

Three numbers to watch before 2028

The project was announced with one test attached: if no matching procurement lands before 2028, construction is running behind the story. That test is now measurable. Watch three things.

Substation and transformer awards tied to the Abu Dhabi campus, not to the Dubai or Microsoft projects.The commissioning date of Taweelah C, because the gas plant is the firm capacity behind the 1 GW cluster.Any revision to the 6 TWh national forecast, which is the only public number that would have to move if the campus is counted on top of the colocation market.Until one of those moves, the plain reading is that Abu Dhabi has announced more AI load than its published demand forecast contains, and has arranged the generation to serve it on a schedule that runs one to two years behind the buildings. That is not a failure. It is what building infrastructure at this speed looks like.

Data basis

Data as of November 2026. Capacity figures come from G42, OpenAI and Oracle announcements as reported by Reuters and Khaleej Times, and from the project description published by Data Center World Middle East. National demand of 173 TWh and the 3 TWh to 6 TWh data centre range come from Wood Mackenzie's March 2026 analysis. Energy conversions multiply facility load by 8,760 hours and assume a constant draw, which overstates consumption for any cluster that idles; the utilisation cases above show the range, not a prediction. The PUE of 1.18 is a design average for modern UAE builds, not a metered figure for this site. Stargate capacity and investment are targets disclosed by the partners, not completed assets.

Sources

Data Center World Middle East 🔗, the strategic case for UAE data centre investment, 2026Khaleej Times 🔗, Stargate UAE to begin in 2026, May 2025EWEC 🔗, generation portfolio and Barakah capacity listing, accessed November 2026 (site home)Enerdata 🔗, TAQA selects Siemens Energy turbines for Taweelah C, June 2026RAX 🔗, AI data centre demand in the UAE, 2026About Global Infinity Labs

Global Infinity Labs is an AI innovation company specializing in MCP data services. This article was prepared by our expert team using insights and data provided through our MCP platforms.

Contact: [email protected]

Copyright 2026 ACN Newswire. All rights reserved. www.acnnewswire.com