Artificial intelligence feels weightless. A question typed into a window, an answer a second later. The weight is somewhere else. It sits on the plains of west Texas, in the desert outside Abu Dhabi, beside a fjord in Arctic Norway and in the palm-oil country of southern Malaysia. It is measured in chips, transformers, cooling towers and power lines, and in a unit that used to belong to national energy plans: the gigawatt. In 2026, the largest technology companies will spend more on that machinery than the whole world invests in producing oil and gas. So what does it really cost?
APRIL'S NUMBERS On 7 April, Anthropic's annualised revenue was reported to have passed $30 billion, up from about $9 billion at the end of 2025. OpenAI had crossed $25 billion in February. Two companies that barely existed as businesses five years ago were earning at the pace of industrial groups. A fortnight later, at Google Cloud Next in Las Vegas, Google split its eighth-generation TPU into two chips. The TPU 8t, for training, links 9,600 chips in one superpod with two petabytes of shared memory and 121 ExaFlops of compute, almost three times its predecessor. The TPU 8i, for inference, triples the on-chip memory and is built for AI agents that cannot wait. Every one of those advances needs a building, a grid connection and a power contract. The revenue is real. So is the bill.
TRAINING, THEN SERVING Training a frontier model ties up vast clusters of chips for weeks or months. Then the second bill arrives. Every answer a model gives, to every user, around the clock, runs on hardware too. Training is a project. Serving is forever. That is why efficiency has become a weapon. Chip utilisation, memory bandwidth, network layout and the way workloads are scheduled decide how much each dollar of hardware and each kilowatt-hour actually delivers. A brilliant model that is too expensive to serve is a research paper, not a business. And the supply chain is narrow. A handful of chip designers, one dominant foundry, a few packaging plants. Land, grid connections and transformers. One delay anywhere slows everything behind it.
THE CAPITAL BILL Amazon told investors in February that it would invest about $200 billion in 2026, up from around $125 billion in 2025. Microsoft guided to about $190 billion for its fiscal year. Alphabet raised its range to $180–190 billion and warned that 2027 would be significantly higher. Meta lifted its range to $125–145 billion, blaming the rising price of components and data centres. Together, roughly $630–650 billion in a single year. The International Energy Agency puts the capital spending of the biggest technology companies above $400 billion in 2025, rising by a further 75% in 2026.
WHAT A GIGAWATT COSTS The industry now prices AI capacity by the gigawatt, roughly the output of a large nuclear reactor. Jensen Huang of Nvidia puts the cost of a one-gigawatt AI facility at $50 billion to $60 billion, most of it in chips, networking and computers rather than concrete. Meta's Hyperion campus in Richland Parish, Louisiana, shows the scale. In July 2026 Meta confirmed it will grow beyond five gigawatts, on a footprint Mark Zuckerberg has compared to Manhattan. Construction alone is put at about $50 billion. With the chips, Bloomberg estimates the bill could approach $250 billion. The building itself costs very different amounts depending on the city. One 2026 industry benchmark put construction at about $8.2 million per megawatt in northern Virginia, $9.1 million in Texas, $9.7 million in Stockholm, $10.3 million in Dublin, $11.8 million in Frankfurt, $12.4 million in Singapore, $14.2 million in Tokyo and $15.7 million in central London. Land and labour matter. Power matters most.
AMERICA: STARGATE AND THE RACE FOR POWER The United States uses more data centre electricity than any other country, and it is where the largest single projects are rising. Stargate, the $500 billion venture led by OpenAI with Oracle and SoftBank, is the emblem. Its first campus, at Abilene in Texas, is planned at 1.2 gigawatts. By late April about 200 megawatts were running, with completion due in the fourth quarter. Six more sites, in Texas, New Mexico, Wisconsin, Michigan and Ohio, take planned capacity above nine gigawatts by 2029. Plans are not buildings. OpenAI has already dropped a larger expansion at Abilene, and Lordstown in Ohio has moved to ban new data centres. Local opposition, grid queues and equipment shortages are now part of the American story.
ENERGY IS NO LONGER A FOOTNOTE Data centres used about 485 terawatt-hours of electricity in 2025, according to the IEA, up 17% in a year when global demand grew 3%. AI-focused data centres used half as much again as the year before. By 2030 the agency expects around 950 terawatt-hours, close to 3% of the world's electricity. The grid is struggling to keep pace. Gas turbine orders rose 70% in 2025, and the IEA expects 15 to 27 gigawatts of gas generation to be built on site at data centres by 2030, most of it in the United States. High-bandwidth memory is expected to stay short until the end of 2027. Inside the halls, AI server power density rose elevenfold between 2020 and 2025. By 2027 a single rack could draw as much power at its peak as 65 homes.
WATER AND HEAT All that power becomes heat, and heat has to go somewhere. The Lawrence Berkeley National Laboratory found that US data centres consumed about 17 billion gallons of water directly for cooling in 2023, around 64 billion litres, and roughly twelve times as much again through the power stations that supplied them. Google's thirstiest site, at Council Bluffs in Iowa, used about a billion gallons in 2024. Its air-cooled site at Pflugerville in Texas used ten thousand. In dry regions, water has become as political as power.
EUROPE: AMBITION MEETS THE PRICE OF POWER Europe wants computing of its own. Its energy bill keeps getting in the way. Brussels has set aside €20 billion for AI gigafactories, each built around more than 100,000 advanced processors, with the European Investment Bank drawing in private money. Ireland shows what happens when a small grid hosts a large industry: data centres used 23% of the country's electricity in 2025, more than every urban home in Ireland combined. In April 2026, OpenAI put Stargate UK on hold indefinitely, citing the cost of energy and regulation. Days later, Narvik in Arctic Norway, a 230-megawatt campus run by Nscale with the Norwegian investor Aker and once billed as Europe's first Stargate, signed Microsoft instead. Microsoft will lease 30,000 Nvidia Vera Rubin GPUs there, on top of a $6.2 billion commitment made in 2025. Cheap hydropower and cold air still win.
THE MIDDLE EAST: SOVEREIGN WEALTH AND A NEW KIND OF RISK The Gulf has the capital, the land and the energy, and it is spending all three. Stargate UAE, a one-gigawatt cluster built by G42 with OpenAI, Oracle, Nvidia, Cisco and SoftBank, sits inside a planned five-gigawatt UAE–US AI campus in Abu Dhabi. Its first 200 megawatts are due in 2026. In Saudi Arabia, Humain, the Public Investment Fund's AI company, is reported to be targeting around 6.6 gigawatts over the next decade, in a country whose entire data centre load was about 410 megawatts in mid-2026. AWS has committed $5.3 billion to a new Saudi region. Then came the war. In early March 2026, Iranian drones and missiles struck three Amazon data centres, two in the UAE and one in Bahrain, knocking them offline and disrupting banking, payments and delivery apps across the region. Analysts believe it was the first deliberate air strike on data centres in any conflict. A former US National Security Council official suggested Gulf facilities may now need missile defence. Security has joined power and chips on the bill.
ASIA: SCALE, SOVEREIGNTY AND THE JOHOR SQUEEZE In China, Alibaba pledged RMB 380 billion, about $52 billion, over three years to 2028, and its chief executive has told analysts it will probably spend more. Chips from Huawei and Alibaba's own designers are carrying more of the load as US export controls keep Nvidia's best hardware out. India has become a battleground for American capital. In December 2025 Amazon committed more than $35 billion by 2030, Microsoft $17.5 billion from 2026 to 2029, and Google $15 billion for an AI hub at Visakhapatnam with AdaniConneX and Airtel. South Korea secured more than 260,000 Nvidia GPUs at the APEC summit in Gyeongju, shared between the government, Samsung, SK, Hyundai and Naver. Malaysia shows the limits. Johor, across the causeway from Singapore, has about 1,110 megawatts of live capacity, third in Asia-Pacific after Tokyo and Singapore, and more than 8.5 gigawatts in the pipeline. Vacancy is under 1%. With power and water tariffs rising, a 50-megawatt facility could pay up to RM120 million a year for electricity alone.
WHO PAYS The money moves in layers. Chipmakers sell the accelerators. Cloud companies and developers build the halls. Utilities and gas producers supply the power. Model companies rent the capacity, and businesses decide whether the tools are worth the subscription. Each layer carries its own risk. Chipmakers ride the spending cycle. Cloud companies carry the cost of halls that might sit half empty. Model companies have to serve every answer for less than they charge for it. The energy needed for each AI task is falling at least tenfold a year, according to the IEA. Consumption by AI data centres still rose by half in 2025. Governments have joined in. Seoul, Riyadh, Abu Dhabi and Brussels are all paying for capacity they control. Not every job needs the biggest model. Smaller, specialised systems running close to the work will win plenty of business.
THE DISCIPLINE TEST The cost of AI is no longer hidden in a software budget. It is written into national grids, Gulf security plans, European energy policy and the balance sheets of the largest companies on earth. Some of what is being built today will look visionary. Some will look like the fibre-optic glut of 2001. The difference will come down to discipline: where it is built, how it is powered, how efficiently it runs, and whether the services on top of it earn their keep. The race is no longer about the model. It is about everything underneath it, and who pays the bill.