Geographic Inequality of Compute
Geographic Inequality of Compute
Data centers go where electricity is cheap. This is rational economics. But it creates a geographic distribution where:
- The environmental burden (land use, water consumption, local pollution, grid strain) is borne by one community
- The capability (AI services, cloud computing, digital infrastructure) benefits a different community
- The profits flow to a third location (corporate headquarters, shareholders)
This is a classic environmental justice pattern, applied to AI infrastructure.
Where Data Centers Go
Data centers cluster in locations with:
- Cheap electricity (often fossil fuels, sometimes hydro or nuclear)
- Cool climates (reducing cooling costs)
- Available land (large facilities, often rural)
- Permissive regulations (fast permitting, low environmental standards)
- Tax incentives (localities competing for jobs and investment)
This means: rural areas, developing regions, places where electricity is cheap because it’s dirty or because other users are subsidizing infrastructure.
Who Bears the Costs
Local impacts include:
- Water consumption for cooling (often in water-stressed areas)
- Strain on electrical grids
- Land use (large facilities, often on previously open land)
- Noise and light pollution
- Heat discharge
- Local air quality impacts from backup generators
These costs are borne by people who may or may not use the AI services the data centers provide.
Who Receives the Benefits
AI capabilities disproportionately benefit:
- Wealthy individuals who can afford subscriptions
- Large corporations that can integrate AI into workflows
- Urban knowledge workers
- English-speaking populations (current AI capability bias)
The people living near the data center may have limited access to the services it provides.
The Colonial Echo
This pattern has historical precedent:
- Resource extraction that benefits distant populations
- Infrastructure built to serve external interests
- Local costs, external benefits
- Decision-making power held elsewhere
Calling it “colonial” may be too strong, but the structural similarity exists.
The Renewable Energy Question
Providers claim data centers use renewable energy. This claim is complicated:
- Additionality: Is the renewable energy truly additional, or does the data center consume renewable capacity that would otherwise serve other users?
- Grid effects: A data center using “100% renewable” may force other grid users onto fossil fuels
- Location matters: A data center in a coal-heavy grid, buying renewable credits from a hydro-heavy grid, isn’t actually reducing emissions
- Verification: Renewable claims are often unverifiable by users
The geography of clean energy and the geography of cheap energy don’t fully overlap.
Implications
- AI’s environmental footprint is not evenly distributed
- Users benefit from impacts they don’t see or bear
- Environmental justice frameworks apply to AI infrastructure
- “Renewable energy” claims require scrutiny
Open Questions
- Should AI pricing reflect true environmental costs, including geographic inequality?
- What obligations do AI providers have to host communities?
- Can AI infrastructure development be done equitably?
- How should users weigh benefits against distributed harms?
See Also
- Training vs Inference Footprint — the scale of energy consumption
- Embodied Carbon — other distributed environmental costs
- The Irony of AI for Climate — whether net impact is positive
- Dependency Lock-in — once committed, addressing these issues is harder
- The Access Gradient — the economic parallel: who can afford access vs. who gets locked out mirrors who bears physical costs vs. who benefits
- The Nuclear Renaissance Question — nuclear siting replicates the geographic inequality pattern: local risk, distant benefit
- Stranded Assets Risk — host communities bear both the operational burden and the stranding risk if demand shifts
- The One More Query Problem — the aggregate demand that justifies these geographically unequal facilities