The chain nobody prices, but everyone pays There is a quiet accounting identity underneath the entire AI economy, and most balance sheets do not show it. When a model returns a token — a word, a fragment of code, a pixel of an image — that token is the output of a defined number of floating-point operations. Those FLOPs run on silicon. And that silicon, every transistor switch of it, is paid for in joules of electricity. Tokens resolve to FLOPs; FLOPs resolve to energy. The unit cost of intelligence is, at the bottom, a unit cost of power. For a sense of scale, Epoch AI's 2025 analysis estimates a typical ChatGPT query consumes roughly 0.3 watt-hours — about 1,080 joules — a figure it argues is far lower than the widely repeated 3 Wh estimate, thanks to model and hardware optimisation. One query is trivial. Multiply it by hundreds of millions of daily users, then by the far heavier cost of training frontier models and running agentic, multi-step reasoning, and the joules compound into a national-grid-sized appetite. This is why we argue energy is becoming the next currency. Not as a metaphor about importance, but literally: as AI becomes the substrate of economic activity, the marginal cost of producing value increasingly collapses into the marginal cost of a watt-hour. Whoever controls cheap, abundant, reliable power controls the cost basis of intelligence itself. The demand curve is bending, and credible bodies are alarmed The International Energy Agency, in its 2025 Energy and AI special report, projects that global electricity consumption from data centres will roughly double — from about 415–485 TWh in 2024–2025 to around 945 TWh by 2030 — reaching just under 3% of total global electricity demand. AI-focused capacity grows fastest of all, tripling over the period. The IEA also reported that data-centre electricity use grew 17% in 2025 alone, with AI-specific demand surging roughly 50%. In the United States — the epicentre of the buildout — Lawrence Berkeley National Laboratory's December 2024 report found data centres consumed about 4.4% of national electricity in 2023 and could reach between 6.7% and 12% by 2028, with consumption climbing from 176 TWh in 2023 toward 325–580 TWh. The IEA notes that in the US, data centres account for nearly half of all electricity demand growth between now and 2030. The capital follows the kilowatts. The IEA found that the combined capital expenditure of just five technology companies in 2025 — over USD 400 billion, expected to rise another 75% in 2026 — is now larger than global investment in oil and natural gas production. That single comparison reframes the entire debate: the world's marginal energy dollar is no longer chasing hydrocarbons to move things. It is chasing electrons to think. Data-centre electricity demand: ~945 TWh by 2030 (IEA Base Case), roughly double 2025. US data centres: 4.4% of national power in 2023, potentially 6.7–12% by 2028 (LBNL). Tech capex (5 firms) > global oil and gas supply investment in 2025 (IEA). Power, not chips, is now the binding constraint For two years the bottleneck story was GPUs. That story has changed. The scarcest input in frontier AI is increasingly the megawatt — interconnection queues, transmission, and firm baseload generation. The IEA warns that around 20% of planned data-centre projects risk delay unless grid constraints are addressed. You can buy chips faster than you can energise them. Hyperscalers have responded by behaving like utilities. In May 2024, Microsoft and Brookfield signed what they called the largest corporate power-purchase framework ever — over USD 10 billion to develop more than 10.5 GW of new renewable capacity between 2026 and 2030, roughly eight times larger than any prior corporate deal. Satya Nadella told analysts the company intended to nearly double its data-centre footprint, a scale of power procurement no corporation had previously attempted. When the largest software companies on earth start signing decade-long electricity contracts measured in gigawatts, the signal is unmistakable: the AI race has become an energy race wearing a software costume. Compute as a strategic reserve asset Once a resource determines national competitiveness, states stop treating it as a commodity and start treating it as a reserve. That is now happening with compute and the power behind it. Analysts and policy bodies increasingly describe AI infrastructure as critical national infrastructure — comparable to electricity grids, ports, or pipelines — and governments are stockpiling GPUs the way they once stockpiled strategic oil or enriched uranium. The numbers are sovereign in scale. France's President Macron announced roughly €109 billion in AI infrastructure investment in February 2025. South Korea has framed a sovereign-AI initiative in the hundreds of billions of dollars. Canada launched a Sovereign AI Compute strategy in 2026 to build domestically owned capacity. By 2026, global spending on sovereign AI systems is projected to surpass USD 100 billion. The United States' hyperscaler fleet alone runs on the order of a million high-end GPUs — a concentration of computational power that is now a measure of national strength. This is the deeper meaning of 'energy as currency.' A barrel of oil is fungible and storable; you bank it against future demand. Compute capacity backed by firm power is becoming the new barrel — a strategic reserve of latent intelligence that a nation can draw down to defend, automate, discover, and grow. The metric that will decide winners: intelligence per watt If energy is the input and tokens are the output, the governing efficiency ratio is intelligence per watt — how much useful machine cognition you extract from each unit of power. Nvidia has been explicit about this reframing. At GTC 2026, Jensen Huang recast data centres as 'AI factories' whose primary product is the token and whose defining metric is tokens per watt, arguing successive architectures push that curve sharply upward. This matters because it splits the economics of AI into two competitive fronts. The first is silicon efficiency — better chips, better cooling, better orchestration squeezing more tokens from each joule. The second, and ultimately more durable, is the cost and cleanliness of the joule itself. A region with electricity at one-fifth the global price and a chip one generation behind can still deliver cheaper intelligence per dollar than a frontier-chip operator paying peak grid rates. For businesses, the implication is strategic rather than technical. The price of an AI-delivered outcome will increasingly track the underlying cost of energy and the efficiency of the stack converting it. Procurement, location, and architecture decisions made today are, in effect, long-dated bets on the future price of a watt. The Gulf's structural advantage: turning sunlight and capital into tokens Few regions are better positioned to win an intelligence-per-watt economy than the Gulf, and the United Arab Emirates in particular. The advantage rests on three reinforcing pillars: extraordinarily cheap power, deep sovereign capital, and explicit national AI ambition — a combination almost no other geography holds simultaneously. Start with power. Abu Dhabi's 2 GW Al Dhafra plant, inaugurated in 2023 as the world's largest single-site solar facility, sealed a power-purchase tariff of roughly 1.32 US cents per kWh — at the time a world record low. Regional solar production costs run a fraction of the global average. For firm, around-the-clock baseload, the Barakah Nuclear Energy Plant reached its full four-unit capacity of 5.6 GW after Unit 4 entered commercial operation in September 2024, generating about 40 TWh a year and supplying roughly a quarter of the UAE's electricity. Cheap solar by day, carbon-free nuclear around the clock: a near-ideal energy mix for power-hungry AI factories. Then add capital and ambition. MGX, the Abu Dhabi AI investment vehicle launched in 2024 with Mubadala and G42 as founding partners, was reported to be mobilising on the order of USD 100 billion toward AI infrastructure and semiconductors, and is a named partner in the US Stargate venture. The UAE was the first country to appoint a dedicated Minister of State for Artificial Intelligence, back in 2017 — turning AI from an industrial bet into a state project. Cheap power: Al Dhafra solar PPA at ~1.32 c/kWh, a world-record-low tariff. Firm baseload: Barakah nuclear at 5.6 GW, ~40 TWh/yr, ~25% of UAE electricity. Sovereign capital: MGX mobilising on the order of $100bn for AI infrastructure. State-level ambition: AI as national strategy, not a private-sector experiment. Stargate UAE and the export of sovereign compute The clearest expression of this thesis is physical and under construction. In May 2025, G42, OpenAI, Oracle, Nvidia, SoftBank, and Cisco announced Stargate UAE — a 1 GW compute cluster in Abu Dhabi and the first node of OpenAI's 'OpenAI for Countries' programme outside the United States, set within a planned 5 GW UAE–US AI campus spanning roughly 10 square miles and powered by a mix of nuclear, solar, and gas. The first 200 MW is targeted to go live in 2026, built by G42's Khazna using Nvidia's Grace Blackwell systems. The scale is best understood against the US flagship. OpenAI's domestic Stargate programme — a USD 500 billion, 10 GW commitment unveiled in January 2025 with SoftBank, Oracle, and MGX — had reached nearly 7 GW and over USD 400 billion in planned investment by late 2025. The UAE is not a bystander to that buildout; through MGX it is a financier of it, while simultaneously hosting the largest such campus beyond American soil. This is energy-as-currency made literal. The UAE is converting an endowment of cheap electrons and sovereign savings into the most strategically valuable industrial output of the era — compute — and positioning itself as a neutral, well-powered host for the intelligence needs of allied nations and global enterprises alike. What this means for businesses If energy is becoming the currency of the AI economy, then every organisation's AI strategy is, whether it admits it or not, an energy strategy. Three implications follow for executives. First, treat compute cost as an energy derivative. The long-run price of the AI capability your business depends on will track the cost and reliability of power behind it. Diversifying toward providers and regions with cheap, firm, clean electricity is no longer an ESG footnote — it is cost-of-goods management for the intelligence you consume. Second, optimise for intelligence per watt inside your own stack. Right-sizing models to tasks, caching, and routing work to the most efficient tier matters more than chasing the largest model for every query. Efficiency at the application layer compounds directly into margin, because each unwasted token is an unspent joule. Third, read the geography. The centre of gravity for affordable, sovereign compute is shifting toward power-rich regions — and the Gulf is engineering itself to be one of them. Businesses that build where energy is cheap and abundant, on platforms designed to extract maximum output per watt, will hold a structural cost advantage over those tethered to constrained, expensive grids. In the AI era, the question is no longer only 'what can the model do?' It is 'what does the energy behind it cost — and who controls it?'