The AI industry has a resource problem it can no longer wave away with growth charts. Data centers used about 485 terawatt-hours of electricity in 2025; the International Energy Agency projects nearly double that, about 950 terawatt-hours, by 2030, roughly what the entire nation of Japan uses today. This is the ledger: what the power costs, what the water costs, who discloses what, and the efficiency counterattack that might bend both curves.
The power ledger: 485 to 950 terawatt-hours
The macro numbers come from the IEA's Energy and AI report, updated in April 2026. Global data-center electricity consumption climbed 17% in 2025; AI-focused facilities grew even faster, up 50%. By 2030 the agency's central projection is about 950 TWh, roughly 3% of global electricity. AI is the driver: electricity demand from AI-optimized data centers is projected to more than triple by 2030.
The United States carries most of the growth. A Department of Energy report updated in 2026 finds US data centers could claim between 9.5% and 15.3% of national electricity by 2030; the reference case lands at 11.8%, or 649 TWh. The IEA adds that data centers are on course to account for nearly half of all US electricity demand growth between now and 2030. By then, the agency notes, America will use more electricity processing data than manufacturing all energy-intensive goods combined: aluminum, steel, cement, and chemicals.
Capital is moving faster than the grid. The five largest technology companies poured more than $400 billion into data centers and related infrastructure in 2025, with plans calling for a 75% jump in 2026, pushing annual spending above $700 billion, as reported by WebProNews. Supply has not kept pace: nearly 2 TW of clean energy, 1.6 times current grid capacity, sits delayed in interconnection queues, according to ERM's analysis of the power crunch.
The gigawatt campus ledger
The buildout is no longer measured in megawatts. Here is what is actually announced, with disclosed figures and reported plans clearly separated:
| Campus | Power (status) | Water (status) |
|---|---|---|
| Stargate Abilene, Texas (Crusoe, Oracle, OpenAI) | 1.2 GW at full buildout, disclosed; 618 MW delivered to Oracle as of Sept 2026; adjacent 900 MW Microsoft campus announced | Not disclosed (estimated) |
| Meta Hyperion, Richland Parish, Louisiana | Reported at up to 5 GW at full build in industry mapping (not company-confirmed at that figure) | Not disclosed (estimated) |
| xAI Colossus, Memphis, Tennessee | About 1.7 GW running of 2 GW planned, per industry mapping | Disclosed: ~794,000 gal/day via the local utility (Sept 2026); $80M recycled-water plant announced |
| Crusoe (all projects) | 4.9 GW contracted, 40+ GW pipeline, company-disclosed (Sept 2026) | Not disclosed |
The Abilene story shows the volatility underneath the announcements. Oracle and OpenAI stepped back from expanding the campus to 2 GW earlier this year, citing financing challenges and shifting demand forecasts, and Microsoft has since agreed to lease around 700 MW of the capacity instead, Bloomberg reported via Data Center Dynamics. Most of the gigawatt era is still a construction site: of the 18 biggest US AI campuses mapped this month, about 6 GW of a planned 30 GW is actually running, and 10 have nothing switched on yet, per Epoch AI satellite estimates cited in the mapping.
The capex behind these campuses is exactly what the industry's spending warnings are about. We tracked the arithmetic when Gartner warned that $3.5 trillion of AI spending may never pay off: $8.5 trillion needed versus $3.6 trillion forecast to be earned. Every gigawatt on this ledger is a bet that the revenue shows up.
The water ledger
Power gets the headlines; water gets the lawsuits. Cooling is the quiet constraint, and disclosure is finally improving:
| Operator | What is disclosed | The number |
|---|---|---|
| AWS | First-ever disclosure, June 2026 | 2.5 billion gallons withdrawn in 2025; efficiency 0.12 L/kWh; withdrawals fell 2% while capacity grew; 75% toward water-positive 2030 |
| Meta | Most complete accounting of the majors | 813M gallons globally in 2023, 95% (776M) from data centers; indirect intensity 3.92 L/kWh from purchased electricity, the only disclosure of its kind; 1.6B gallons restored to stressed regions in 2024 |
| Microsoft | Efficiency trend + water-positive claim | Water positive in FY2025, five years early; efficiency 2.3 to 0.27 L/kWh since the early 2000s; Arizona sites worst in fleet |
| xAI, Memphis | Utility-reported usage | ~794,000 gallons/day in late Sept 2026, 62% of the 1.28M allowed; among the utility's biggest customers; full buildout peak projected up to 13M gal/day |
| Meta, Gallatin TN | Utility records (Tennessean) | ~2% of the city's water sold in FY2025-26; ranks 3rd to 5th among customers |
| Aligned Data Centers, Utah | Reported usage | 80M gallons (West Valley) + 47.4M (West Jordan) in 12 months, in a Colorado River Basin state |
The Memphis case shows why water is local and political. The city relies entirely on the Memphis Sand Aquifer for every drop of residential, industrial and agricultural water, and the aquifer is losing water faster than it recharges. After public pressure, xAI committed to an $80 million recycled-wastewater plant with German firm Cerafiltec, designed to reuse 13 million gallons a day, enough for Colossus's cooling with water left over for local businesses. Construction on the plant is currently paused, with xAI committing to restart in the first quarter of 2027. In Arizona, Microsoft's own disclosures show its Arizona sites running the worst water efficiency in its fleet, because desert heat forces heavier evaporative cooling.
The transparency scorecard
Put together, the disclosure picture is uneven but moving in the right direction. AWS went from the least transparent of the four major cloud providers to publishing withdrawal data for the first time in June 2026. Meta discloses the most complete accounting, including indirect water embedded in electricity, which no other major publishes. Microsoft reports fleet efficiency and claims water-positive status five years early. xAI's usage is known only because the local utility said so. The pattern: where regulators, utilities, or journalists ask, numbers appear; where nobody asks, they do not.
The efficiency counterattack
There is a second story running against the ledger, and it is about doing more per watt and per drop:
Microsoft's Maia 300, announced October 12, 2026, is the clearest corporate bet on it. The company expects the chip to deliver 40% more tokens per megawatt than leading GPU-based systems, with up to 7x the rack-scale token throughput of its Maia 200 predecessor, and plans multi-gigawatt deployments starting in 2027, with Maia becoming the primary inference system for its own products. The caveats: the 40% figure is a company expectation against unnamed competitors, the chip does inference only (no training), and independent testing will be needed before treating it as a general result, as IEEE's ComSoc analysis notes. Maia 200 is already in production across the US, Europe and Australia, delivering up to 1.7x more tokens per watt on Teams and Copilot workloads, which gives the claim a track record.
The market is already voting for efficiency. We documented the 75% collapse in token prices this year: cheaper inference is the market expression of the same efficiency race. DeepSeek's V4.1-Flash topped LMArena's WebDev board at roughly a tenth of its nearest rival's cost (LMArena snapshot, October 18). And the World Bank argued this month that "small AI" could transform developing countries (The Guardian, October 18), the counter-narrative that efficiency, not scale, is the prize for most of the world.
What the ledger says
Four conclusions survive scrutiny. First, power is the binding constraint on AI growth, not chips: the grid queue is longer than the GPU queue. Second, water is local and political in a way power is not: one aquifer city hosting one of the largest training clusters on Earth is a governance problem, not an engineering one. Third, disclosure is improving but remains voluntary and uneven; the numbers we have exist because someone asked. Fourth, efficiency is the only lever that bends both curves at once, and it is accelerating: custom silicon, cheaper inference, and smaller models all point the same way. The ledger will keep growing. Whether it grows sustainably depends on the fourth line, not the first three.
By ByteHerald Staff. Macro figures: IEA Energy and AI (April 2026 update), US DOE 2026; campus figures via Data Center Dynamics and American Buildout; water figures via company sustainability disclosures and utility records as cited; Maia 300 details via Microsoft's October 12 announcement as reported by IEEE ComSoc and Recul.
Comments
Post a Comment