AI's Next Bottleneck Isn't Chips. It's Power and Water

Two years ago every serious conversation in tech circled back to Nvidia chips and who could get an allocation of H100s or Blackwells. Today the question in boardrooms has shifted. Meta, Microsoft, Amazon and Google are no longer fighting only over processors. They are fighting over megawatts, over gigawatts, over twenty year power contracts with reactors that were nearly shut down a few years back. The Gulf, having spent years building a reputation as a chip and data center hub, now faces the same question in a harder version. Where does the power come from in a desert region that already struggles with water.

The short answer is that the AI race quietly turned into an energy infrastructure race. Whoever secures cheap and reliable electricity will shape who trains and serves the next generation of AI models. Owning the newest chip matters less if there is no megawatt to plug it into. This piece tracks the real numbers behind that shift, globally and specifically across the Gulf, and asks a question that has not had enough airtime yet. What does building gigawatts of data centers actually cost a region that imports much of its water.

The numbers below cover four things: how much power AI now needs worldwide, how Meta and its peers are buying that power directly from utilities and plant operators, how the UAE and Saudi Arabia are each trying to answer the same question from very different starting points, and what almost nobody in the initial announcements mentions, the water bill that comes attached to every gigawatt built in the desert.

Start with the baseline. The International Energy Agency puts global data center electricity consumption at around 415 terawatt hours in 2024, about 1.5 percent of all the electricity used on the planet that year. By 2030 the agency projects that figure will climb to roughly 945 terawatt hours, more than doubling in six years and pushing data centers to nearly 3 percent of global electricity demand. The driver is not storage or video streaming. It is what the IEA calls accelerated servers, the GPU heavy racks built for AI training and inference, growing at around 30 percent a year and accounting for almost half of the entire net increase in data center power use through the decade.

Break the growth down by geography and the picture sharpens further. The United States adds close to 240 terawatt hours of new data center demand by 2030, a 130 percent jump. China adds around 175 terawatt hours, up 170 percent. Europe grows by 45 terawatt hours and Japan by 15. Together, the US and China account for close to four fifths of the world's added demand. Gulf states barely register yet in these global tallies, which is exactly the point of this piece. They are trying to build gigawatt scale capacity from close to zero, on a compressed timeline, in a region where every additional megawatt of power tends to pull an equivalent problem of water behind it.

The IEA frames this shift as part of a broader move it calls the Age of Electricity, in which data centers, electric vehicles and industrial cooling all pull on the same grids at the same time. For utility planners who used to size new capacity in small increments, a single AI campus requesting a gigawatt of firm power is a different kind of customer altogether, closer to a small country's grid than to one large factory.

Look at where the money is actually going and the shift becomes concrete. Amazon, Google, Meta and Microsoft are on track to spend roughly 725 billion dollars combined on AI infrastructure in 2026 alone, according to trackers built from company earnings calls. Amazon leads with around 200 billion, Google follows near 185 billion, Meta sits close to 125 billion and Microsoft around 120 billion. A growing share of that spending no longer buys chips. It buys substations, transmission lines and, increasingly, entire generating plants.

Meta made the point unmistakable in January 2026 when it signed a set of long term energy agreements covering up to 6.6 gigawatts of capacity through 2035. Vistra will supply more than 2.1 gigawatts from its Perry and Davis-Besse plants in Ohio plus Beaver Valley in Pennsylvania, with 433 additional megawatts coming from plant uprates in the early 2030s, locked in under 20 year contracts. TerraPower will build up to eight advanced reactor units for as much as 2.8 gigawatts, with the first pair targeted online as early as 2032. Oklo will add up to 1.2 gigawatts from small advanced reactors in Pike County, Ohio, potentially live by 2030. A company that until recently was best known for social apps is now one of the largest corporate buyers of advanced energy capacity in American history. That sentence alone tells you where this decade of AI infrastructure spending is actually headed.

Meta's shopping spree did not appear out of nowhere. It builds directly on a 2025 agreement with Constellation Energy, an early signal that a hyperscaler would rather buy the output of an existing plant than wait years for a new one to be licensed. Expect the rest of the industry to follow a similar script wherever an underused plant sits near enough to a planned campus.

Older baseload plants used to be the industry's problem child. Long build times, cost overruns, public skepticism after past accidents. AI changed the calculation for one blunt reason: data centers need power that runs at close to full output around the clock, for decades, without the intermittency of solar or wind. A plant that already exists and is paid off looks less like a legacy asset and more like the cheapest reliable electricity a hyperscaler can lock in today. That is why Meta is paying to keep older US plants running rather than waiting for new ones, while simultaneously betting on next generation compact reactor designs from TerraPower and Oklo that will not deliver a single watt before the early 2030s.

The logic matters for the Gulf too, and for a reason that is easy to miss. Building this kind of energy program from scratch typically takes a decade or more of licensing, safety review and construction before the first watt flows. Any government or utility starting that process today is really planning for the AI demand of the mid 2030s, not the AI demand of 2026. Which is exactly why the region's one country that already has an operating plant of this kind, rather than a plan for one, currently holds a genuine structural advantage over its neighbors.

The UAE started planning its Barakah energy program back in 2007, well before anyone in Abu Dhabi was talking about AI campuses. The four unit plant reached its full 5.6 gigawatt capacity in 2024, generating roughly 40 terawatt hours a year and supplying close to a quarter of the country's total electricity. That timing turns out to matter enormously now. Talal Al Kaissi, acting chief global affairs officer at G42's Core42, put it plainly when discussing the five gigawatt AI campus conversation now underway in the emirate. Starting such a program from scratch for that conversation, he said, would have taken more than ten years. The UAE simply did not have to start from scratch.

Under the country's Energy Strategy 2050, renewable capacity is targeted to hit 14.2 gigawatts by 2030, with total clean energy capacity, renewables plus Barakah's output, reaching 19.8 gigawatts. Add current annual renewable additions of about 1.5 gigawatts a year and battery storage climbing from a few megawatts today toward more than a gigawatt by the end of the decade, and the picture is one of a country that quietly built a clean power surplus years before it needed one for AI. Few Gulf neighbors can say the same, and it explains why global AI infrastructure money keeps finding its way to Abu Dhabi specifically rather than the region broadly.

Barakah also carries symbolic weight beyond the numbers. It is the first commercial nuclear plant built in the Arab world, a fact Emirati officials invoke often when explaining why Abu Dhabi, rather than any other Gulf capital, became the natural home for the region's largest AI campus.

Barakah's surplus capacity is exactly what makes Stargate UAE possible. Announced on May 22, 2025 during Donald Trump's Gulf tour, as part of the broader US-UAE AI Acceleration Partnership, the project envisions a five gigawatt AI campus in Abu Dhabi built with G42, OpenAI, Oracle, Nvidia, SoftBank and Cisco. A first one gigawatt cluster is being built by G42 and will be operated jointly by OpenAI and Oracle, running on Nvidia's Grace Blackwell GB300 systems, with Cisco handling security and connectivity. The campus itself spans roughly ten square miles, and officials describe it as the largest such deployment outside the United States.

The project is explicit about its energy mix in a way most AI announcements are not. Stargate UAE says it will run on a combination of nuclear, solar and natural gas to keep emissions down, a direct nod to Barakah sitting a few hundred kilometers away. The first 200 megawatts of capacity are scheduled to come online in 2026, building toward the full one gigawatt cluster before the remaining four gigawatts require, in the words of people close to the deal, deeper discussions with other hyperscalers. Security concerns tied to the UAE's technology relationships with China have also slowed parts of the broader agreement, a reminder that even with power secured, geopolitics can still be the bottleneck.

The logic behind blending nuclear, solar and gas is straightforward once you see it from a planner's chair. Nuclear supplies the steady baseload a training cluster needs at three in the morning as much as at three in the afternoon, solar shaves the daytime peak and the emissions bill, and gas fills whatever gap is left while the other two scale up. It is a hedge, not a preference, and it is the same hedge Meta is running in Ohio and Pennsylvania, just assembled with different local ingredients.

Saudi Arabia's answer runs through Humain, the AI arm of the Public Investment Fund launched in May 2025. Rather than starting with power plants, Humain started with land. The company has secured 211 plots across the kingdom and is targeting 6 gigawatts of total data center capacity within a decade. One single tender illustrates the ambition. In East Riyadh's Al-Saad area, Humain is developing six plots of one gigawatt each, spanning roughly 24 square kilometers, with construction bids that closed in early May 2026.

The first facilities, 100 megawatts each in Riyadh and Dammam, are targeted to go live in the second quarter of 2026, running on an initial deployment of 1,024 Qualcomm AI accelerators described as one of the largest such rollouts globally. Around that core sit a string of partnerships: a 5 billion dollar AI Zone deal with AWS, a one gigawatt joint venture with Saudi Telecom Company, a 500 megawatt data center partnership with xAI, and 1.2 billion dollars in financing from the kingdom's National Infrastructure Fund tied to 250 megawatts of capacity. What is conspicuously absent from every public disclosure so far is any advanced baseload power source of the kind the UAE already has running. Saudi Arabia has discussed such options for years without one operating yet, which means its data center buildout today leans almost entirely on gas and grid power while Riyadh figures out how to close that gap a decade from now.

Humain is not building all of this alone. In a separate deal, Blackstone backed AirTrunk agreed to jointly develop, finance and operate Saudi data center capacity with Humain, starting with an initial campus valued at roughly 3 billion dollars. The arrangement hands Humain a partner that already runs hyperscale campuses across Asia Pacific, useful experience for a program trying to compress a decade of buildout into a few years.

Every megawatt of data center capacity carries a hidden second bill, and in the Gulf that bill is paid in liters, not dirhams or riyals. Research firm Mordor Intelligence estimates data centers across the Middle East and Africa used about 119 billion liters of water in 2025, a figure set to climb to 426 billion liters a year by 2030. The UAE alone is projected to account for around 61 billion liters of that annual draw by decade's end. Put another way, a single one megawatt data center running standard evaporative cooling consumes roughly 25.5 million liters of water a year, close to what a town of 300,000 people needs.

The timing could hardly be worse for the optics. Gulf residents already consume more than 500 liters of water per person daily, roughly three times the European average, in a region that depends heavily on energy intensive desalination to begin with. Saudi Arabia alone desalinates more than a million cubic meters of water a day just to meet existing demand. Executives are aware of the tension. Autodesk chief executive Andrew Anagnost argued last September that the desert's abundant sunlight could power cooling systems directly through solar generation rather than drawing further on scarce freshwater, and Gulf data center operators are increasingly shifting toward liquid and immersion cooling to cut water draw. None of that changes the near term math though. Building gigawatts of AI capacity in the desert means building a parallel water strategy, and right now that strategy is still being improvised.

The industry response is already visible on construction sites. Trade publications covering the region report a broad shift away from traditional air and evaporative cooling toward liquid and immersion systems specifically to blunt this water impact, even though the retrofit adds cost and complexity to campuses that are already running behind schedule.

Abu Dhabi and Riyadh dominate headlines, but the wider Gulf is not staying still. Kuwait recently joined a 100 billion dollar AI infrastructure partnership alongside MGX and Microsoft, a smaller Gulf economy buying its way into a race it could not credibly run alone. It sits inside a much bigger envelope. Roughly 2 trillion dollars in deals were pledged across the region during Trump's May 2025 Gulf tour, spanning chips, cloud capacity, and now increasingly energy infrastructure to actually run all of it.

What ties these announcements together is less obvious than the dollar figures. Every one of them assumes a power source that either already exists, like Barakah, or has to be built essentially from scratch under time pressure, like Saudi Arabia's gas heavy buildout. Qatar and Kuwait are watching both models play out before committing fully to either, and the smart money in the region is increasingly flowing not into chips, which are commoditized and can be bought from the same handful of suppliers everywhere, but into whoever controls the electrons and the water needed to keep those chips running.

For a Gulf business leader, the practical takeaway is not that AI is running out of chips. It is that AI capacity in the region will increasingly be gated by power and water permits rather than by chip allocation. Companies planning to lease compute from Humain or from Abu Dhabi providers should ask which megawatt tranche and which water source their contract actually sits on, not just which GPU generation. Firms in construction, grid engineering, desalination technology and liquid cooling are looking at a genuine multi year buildout, arguably a steadier one than the AI application layer itself, since gigawatts of physical infrastructure take years to plan and cannot be rolled back by the next model release.

For investors the signal is similarly concrete. Watch utility and grid announcements coming out of Riyadh and Abu Dhabi as closely as chip news, because they are now the leading indicator for which AI campuses can actually be built on schedule. And watch water. A region that already imports much of what it drinks is now also trying to cool the servers training the world's AI models, and that tension has not been resolved. It has only been financed for now.

There is a broader risk sitting underneath all of this worth naming plainly. A June 2026 analysis citing Sequoia's David Cahn put the annual gap between hyperscaler AI capex and the revenue it is generating at roughly 600 billion dollars, with capital spending now running 45 to 57 percent of revenue at the largest cloud companies, a divergence Allianz Research measured at about 46 percent, wider than the telecom bust of 2001. None of that makes the power and water buildout described here optional. It does mean some of the gigawatts under construction today are being financed on the assumption that AI revenue catches up fast enough to justify them, and that assumption has not been tested yet.

IEA - Energy demand from AI

Meta - Nuclear Energy Projects announcement

CNBC - Tech AI spending approaches $700 billion in 2026

Khaleej Times - How Barakah plant gave UAE a head start in powering AI infrastructure

Forbes - UAE 2026: What Powers Microsoft's $15 Billion Abu Dhabi Data Centers

G42 - Global Tech Alliance Launches Stargate UAE

Yahoo Finance - First 200 MW from UAE's Stargate AI campus to come online next year

DataCenterDynamics - Humain secures 211 plots of land in Saudi Arabia for data centers

Construction Week Saudi - PIF Unit Tenders 6GW Data Centre for Saudi AI Push

Rest of World - The Gulf's trillion-dollar AI dream faces a water crisis

Arab News - Sunlight holds key for Gulf data center cooling says tech giant CEO

Real Assets - Blackstone-backed AirTrunk partners PIF's Humain on Saudi data centre projects

Forbes - The AI Capex to Revenue Gap Is Widening