The standing collection · Energy & Infrastructure

AI is rewriting the rules of the power grid.

540 TWh

global data center electricity consumption in 2024 — roughly the UK's entire annual grid

The IEA projects this could nearly double to 1,050 TWh by 2030. The concern isn't just total consumption — it's when that power is drawn, and whether the grid can keep up without locking in decades of fossil fuel dependency.

Global DC power 2024

540 TWh

Projected 2030

1,050 TWh

Growth 2019–2024

+80%

US share of global DC

~33%

Global electricity consumption · TWh/yr

ConfirmedProjectedHover dots for events

Global DC power (2024)540TWh

Projected (2030)1,050TWh

Growth 2019–2024+80%

Share of world electricity~2%

Projected share by 20303–4%

Scale: 540 TWh ≈ the UK's entire annual electricity consumption. The 2030 projection approaches Japan's national grid.

The AI factor: Training GPT-4 consumed ~1,300 MWh. But inference — running models for users — now accounts for ~60% of all AI electricity and never stops.

Trajectory: If this pace continues, new capacity additions must outpace all other demand combined in many US states.

When you use power matters as much as how much

The timing problem: A data center running at full power overnight uses cheap, low-carbon baseload. The same data center ramping up at 5 PM forces gas peaker plants to spin up — raising costs and emissions for everyone.

Total gridData centers

Peak hours (4 PM – 9 PM)

Grid carbon intensity~420 g CO₂/kWh

Electricity cost~14¢/kWh

Residential AC, EV charging, and cooking all compete simultaneously. Every additional MW of data center demand can force a peaker plant online.

Off-peak (midnight – 7 AM)

Grid carbon intensity~250 g CO₂/kWh

Electricity cost~7.8¢/kWh

Excess renewable generation is often curtailed at night — data centers can absorb it. Most training workloads are flexible and don't need to run at 5 PM.

Carbon-aware computing: Google, Microsoft, and others automatically shift non-urgent workloads to off-peak hours. If all hyperscalers adopted this, carbon impact could drop 30–40% without adding a single megawatt of new generation.

New generating capacity needed 2024–2030

The IEA estimates 200+ GW of new global capacity is needed — more than France and Germany combined. CO₂ badge shows lifecycle emissions per MWh. Hover bars for pros/cons.

Climate impact & source trade-offs

Climate context

DC emissions (2023)180 Mt CO₂e/yr

Global aviation915 Mt CO₂e/yr

DC growth rate~15%/yr

Aviation growth rate~3%/yr

Data centers emit ~20% of what aviation emits today. The concern is trajectory: aviation grows slowly, data centers are accelerating.

Solar + storage: Cheapest new source in most of the US. Tech giants signing record PPAs. Challenge: intermittency requires battery backup for cloudy/windless periods.

Nuclear (SMR): Microsoft restarted Three Mile Island. Google contracted Kairos SMRs. 24/7 zero-carbon baseload — but years from commercial scale.

Natural gas — the bridge risk: Builds in 2–3 years vs. 10+ for nuclear. But locks in decades of emissions. The "bridge fuel" risk is that the bridge never ends.

Efficiency — Power Usage Effectiveness (PUE)

Industry average · lower is better

PUE = Power Usage Effectiveness. 1.0 = perfect. The industry halved overhead losses since 2010, but AI servers run hotter — putting upward pressure on newer facilities.

Hyperscaler PUE vs. industry average

Google

1.1

Microsoft

1.18

Meta

1.1

Amazon

1.15

Industry Avg

1.58

Google and Meta achieve 1.10 — 44% more efficient than the industry average of 1.58. Many older colocation facilities still operate above 1.8. AI accelerator clusters (NVIDIA H100/H200) often require direct liquid cooling, which changes the equation further.

Big tech renewable commitments · 2023 actuals vs. stated targets

Google

24/7 clean by 2030

64%

Microsoft

Carbon negative by 2030

58%

Amazon

100% renewable by 2025

90%

Meta

Net zero by 2030

100%

Apple

Carbon neutral by 2030

100%

Annual match ≠ 24/7 clean power. Many companies report "100% renewable" by buying Renewable Energy Certificates (RECs) — a financial instrument that doesn't guarantee the electricity at your data center is clean at the moment it's used. Google's 24/7 carbon-free energy standard requires matching hour-by-hour — a much harder target most haven't yet achieved.

Water — the hidden constraint

Global DC water use (2023)

620Bgal

Enough for a city of 2M for a year

Air-cooled (typical)

1.8gal/kWh

Per kilowatt-hour consumed

Liquid-cooled (modern)

0.2gal/kWh

10× more water-efficient

Why water is becoming a flashpoint: Large data centers are being denied permits in drought-stressed regions — Phoenix, Amsterdam, parts of Ireland — because they strain local water supplies. AI accelerator clusters run significantly hotter and require direct liquid cooling. As AI scales, water rights may become as contested as land and power access.

The standing collection · Energy & Infrastructure

AI is rewriting the rules of the power grid.

540 TWh

global data center electricity consumption in 2024 — roughly the UK's entire annual grid

Global DC power 2024

540 TWh

Projected 2030

1,050 TWh

Growth 2019–2024

+80%

US share of global DC

~33%

Global electricity consumption · TWh/yr

ConfirmedProjectedHover dots for events

Global DC power (2024)540TWh

Projected (2030)1,050TWh

Growth 2019–2024+80%

Share of world electricity~2%

Projected share by 20303–4%

Scale: 540 TWh ≈ the UK's entire annual electricity consumption. The 2030 projection approaches Japan's national grid.

The AI factor: Training GPT-4 consumed ~1,300 MWh. But inference — running models for users — now accounts for ~60% of all AI electricity and never stops.

Trajectory: If this pace continues, new capacity additions must outpace all other demand combined in many US states.

When you use power matters as much as how much

Total gridData centers

Peak hours (4 PM – 9 PM)

Grid carbon intensity~420 g CO₂/kWh

Electricity cost~14¢/kWh

Residential AC, EV charging, and cooking all compete simultaneously. Every additional MW of data center demand can force a peaker plant online.

Off-peak (midnight – 7 AM)

Grid carbon intensity~250 g CO₂/kWh

Electricity cost~7.8¢/kWh

Excess renewable generation is often curtailed at night — data centers can absorb it. Most training workloads are flexible and don't need to run at 5 PM.

New generating capacity needed 2024–2030

The IEA estimates 200+ GW of new global capacity is needed — more than France and Germany combined. CO₂ badge shows lifecycle emissions per MWh. Hover bars for pros/cons.

Climate impact & source trade-offs

Climate context

DC emissions (2023)180 Mt CO₂e/yr

Global aviation915 Mt CO₂e/yr

DC growth rate~15%/yr

Aviation growth rate~3%/yr

Data centers emit ~20% of what aviation emits today. The concern is trajectory: aviation grows slowly, data centers are accelerating.

Solar + storage: Cheapest new source in most of the US. Tech giants signing record PPAs. Challenge: intermittency requires battery backup for cloudy/windless periods.

Nuclear (SMR): Microsoft restarted Three Mile Island. Google contracted Kairos SMRs. 24/7 zero-carbon baseload — but years from commercial scale.

Natural gas — the bridge risk: Builds in 2–3 years vs. 10+ for nuclear. But locks in decades of emissions. The "bridge fuel" risk is that the bridge never ends.

Efficiency — Power Usage Effectiveness (PUE)

Industry average · lower is better

PUE = Power Usage Effectiveness. 1.0 = perfect. The industry halved overhead losses since 2010, but AI servers run hotter — putting upward pressure on newer facilities.

Hyperscaler PUE vs. industry average

Google

1.1

Microsoft

1.18

Meta

1.1

Amazon

1.15

Industry Avg

1.58

Big tech renewable commitments · 2023 actuals vs. stated targets

Google

24/7 clean by 2030

64%

Microsoft

Carbon negative by 2030

58%

Amazon

100% renewable by 2025

90%

Meta

Net zero by 2030

100%

Apple

Carbon neutral by 2030

100%

Water — the hidden constraint

Global DC water use (2023)

620Bgal

Enough for a city of 2M for a year

Air-cooled (typical)

1.8gal/kWh

Per kilowatt-hour consumed

Liquid-cooled (modern)

0.2gal/kWh

10× more water-efficient