## INSIDE THE AI-LED RESOURCE RACE

## Source details

**Canonical URL:** [INSIDE THE AI-LED RESOURCE RACE](https://www.imf.org/-/media/files/publications/fandd/article/2025/12/van-de-graaf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/fandd/article/2025/12/van-de-graaf.pdf.md)
- [Structured JSON version](/-/media/files/publications/fandd/article/2025/12/van-de-graaf.pdf.json)

---

### A hungry industry
- Data centers already use about 1.5 percent of global electricity supply.
- The IEA expects data center demand to more than double by 2030, with AI responsible for much of the increase.
- AI accounts for less than a tenth of added power demand this decade, far below that of electric vehicles or air-conditioning.
- In the United States and Japan, data centers could account for nearly half of new demand by 2030.
- In Ireland, data centers already use more than a fifth of the country’s electricity.
- Northern Virginia: data centers already consume about one-quarter of the state’s power, forcing utilities to delay or cancel other connections.
- Dublin’s grid operator froze new projects in 2022, approving only those that could generate their own power.
- Singapore halted approvals altogether in 2019 and now allows facilities only under strict efficiency rules.

### Big Tech turns to power
- Microsoft, Amazon, and Google have each signed multibillion-dollar power purchase agreements that rival those of traditional utilities.
- Some firms are adding on-site generation or investing in new technologies: Microsoft exploring nuclear (including small modular reactors and mothballed plants), Google backing advanced geothermal, Amazon testing hydrogen for backup power.
- With President Donald Trump rolling back many of President Joe Biden’s climate policies, Big Tech has become a significant driver of clean-energy investment in the US.
- Despite renewables growth in Europe, demand in the US—home to more than 40 percent of the world’s data centers—still leans heavily on natural gas, adding to emissions.

### Smarter machines
- AI can help manage electricity: balancing power grids, forecasting renewable output, and optimizing energy use in buildings and industry; some cities pipe waste heat from server farms into district heating networks.
- Hardware improvements: new generations of chips (e.g., Nvidia’s Blackwell processors and Google’s tensor processing units (TPUs)) deliver more operations per watt.
- Software efficiency: China’s DeepSeek, released in January 2025, was trained at a fraction of the cost and energy of comparable models from OpenAI and Google.
- Efficiency paradox: the Jevons paradox suggests cheaper computing power may spur even greater usage, potentially offsetting efficiency gains.

### Semiconductors and minerals
- Training state-of-the-art models requires thousands of specialized chips, most designed by Nvidia and manufactured almost exclusively in Taiwan Province of China by TSMC.
- Chip fabrication is resource-intensive: a cutting-edge plant can consume as much electricity as a small city and requires vast amounts of ultrapure water.
- Upstream mineral needs include gallium, germanium, silicon, rare earths, copper.
- By 2030, data centers could be consuming more than half a million metric tons of copper and 75,000 tons of silicon each year—enough to lift their share of global demand to 2 percent, according to the IEA.
- For gallium, data centers could account for more than a tenth of total demand.
- China controls 80–90 percent of global refining of silicon, gallium, and rare earths.
- In 2023 China restricted exports of gallium and germanium; since late 2024 new curbs have followed on tungsten, tellurium, bismuth, indium, and molybdenum.
- Prices for many of these metals have spiked; governments (Washington, Brussels, Tokyo, Seoul) have launched critical-mineral strategies including recycling programs and alliances with resource-rich countries.

### Land and water
- Hyperscale data centers prefer locations with cheap power, abundant water, and fast fiber-optic links.
- Two-thirds of new US data centers since 2022 have been built in water-stressed regions.
- Water disputes have arisen in Arizona, Spain, and Singapore over competing claims between households and Big Tech.
- Most of AI’s water footprint is indirect: power plants that supply data centers consume far more water than the centers themselves.
- Climate and network latency influence siting: Ireland’s cluster reflects transatlantic cable hubs; Abu Dhabi’s planned 5-gigawatt campus chosen partly to minimize delays with Asia and Europe; colder countries (Norway, Iceland) tout lower cooling needs.

### Policy challenges and implications
- Planning uncertainty: forecasts of data center demand for 2030 diverge widely; the highest published estimate is nearly seven times the lowest.
- Speed of construction leaves little time for certainty; governments must expand electricity systems quickly without overbuilding or locking in fossil fuels.
- Transparency gap: “There is little public reporting from the industry on data center use of electricity, water, or minerals.”
- Sustainability and equity risks: expanding grids and supply chains without safeguards could repeat boom-and-bust cycles; benefits of AI may concentrate in the rich world if developing economies remain raw-material suppliers and face higher implied costs for energy and capital.
- Two possible pathways:
  - Managed well: AI boom could accelerate clean energy and foster more resilient supply chains.
  - Mismanaged: risk of locking in new emissions and deepening resource dependence.
- Core contention: The AI surge is a material contest—over electrons, gallons, wafers, and ores—and outcomes will shape both leadership in AI and the sustainability and distribution of its gains.

*Source: Thijs Van de Graaf, “Inside the AI-Led Resource Race,” F&D, DECEMBER 2025.*

---


_Source: https://www.imf.org/-/media/files/publications/fandd/article/2025/12/van-de-graaf.pdf_