Answer: Location is critical for AI carbon footprint

The carbon footprint of an AI model depends on three primary factors: the energy consumption of training, the energy efficiency of the datacenter (measured by PUE), and the renewable energy mix of the geographic region where the datacenter is located. A model trained in a datacenter with PUE=1.2 and 80% renewables in Reykjavik (Iceland) has a drastically lower carbon footprint than the same model in a datacenter with PUE=1.8 and 30% renewables in a fossil-heavy region.

Choosing the right datacenter is not purely a technical decision: it is a sustainability and brand alignment decision that impacts the perceived environmental responsibility of any organization doing ML training.

What is PUE (Power Usage Effectiveness)

PUE is the ratio of total power consumed by a datacenter to the power consumed by IT infrastructure alone (servers, storage, networking):

PUE = Total Facility Power / IT Equipment Power

Example:
- Total Facility Power = 1 MW (including cooling, power distribution, lighting)
- IT Equipment Power = 800 kW (servers, GPU, CPU)
- PUE = 1 MW / 0.8 MW = 1.25

A PUE of 1.0 is theoretically perfect: it means 100% of power is dedicated to actual compute, and 0% is used for cooling and overhead.

In the real world:

• PUE < 1.2: state-of-the-art datacenter (hyperscalers like Google, Meta, Apple. Rare.)
• PUE 1.2–1.5: modern and efficient datacenter (best practice)
• PUE 1.5–2.0: legacy or poorly managed datacenter (mediocre)
• PUE > 2.0: very inefficient datacenter (rare in 2026, legacy)

Google achieved an average PUE of 1.10 globally in 2023 — the best result among major hyperscalers. This means that for every 1 kWh of GPU compute, only 0.10 kWh of cooling/overhead is required.

What PUE is NOT

PUE does not directly speak to sustainability: it measures only the mechanical efficiency of the datacenter. A datacenter with PUE=1.1 powered 100% by coal is technically efficient but environmentally disastrous. For a complete assessment, you need both PUE and the renewable energy mix of the region.

Geographic location: the deciding factor

The region where a datacenter is located determines access to renewable energy and ambient temperature, which in turn impacts cooling costs.

Optimal datacenters for AI training (2026):

• Iceland (Reykjavik) — 99% hydroelectric, PUE 1.1–1.2, natural cooling from cold climate. Google, Meta, Microsoft have significant presences.
• Scandinavia (Sweden, Norway, Finland) — 70–100% hydroelectric, PUE 1.15–1.3, excellent network infrastructure.
• Canada (Montreal, Vancouver) — 60–90% hydroelectric in suitable regions, PUE 1.2–1.4, natural cooling.
• Portugal — 50–60% renewables (wind + solar), PUE 1.3–1.5, geographically close to central Europe.
• New Zealand, Tasmania (Australia) — 80%+ renewables, PUE 1.2–1.4, temperate climate.

Regions to avoid for "green" AI training:

• France — 70% nuclear (low-carbon but not renewable, political/reputational risks), average PUE 1.3–1.5
• Germany — 50% renewables (but volatile in winter), PUE 1.4–1.6, water-based cooling (emerging climate risks).
• USA (Texas, Virginia) — 20–40% renewables, PUE 1.2–1.4, fossil grid still relevant.
• China, India — 30–50% renewables (growing), PUE 1.5–2.2 (legacy), poor transparency.

Measuring carbon footprint: the complete formula

A realistic estimate of carbon footprint for an AI model training:

# Carbon Footprint of an AI Training Job

compute_energy_kwh = 1000  # kWh of GPU/CPU during training (e.g.: 100h x 10 kW)
datacenter_pue = 1.25       # Power Usage Effectiveness
total_energy_kwh = compute_energy_kwh * datacenter_pue  # 1250 kWh

# Carbon intensity of the region (kg CO2 / kWh)
# Examples 2026:
# Iceland: 0.08 kg CO2/kWh
# Sweden: 0.15 kg CO2/kWh
# Portugal: 0.25 kg CO2/kWh
# USA avg: 0.35 kg CO2/kWh
# Germany: 0.30 kg CO2/kWh
# India: 0.65 kg CO2/kWh

carbon_intensity_kg_per_kwh = 0.15  # Sweden

total_carbon_kg = total_energy_kwh * carbon_intensity_kg_per_kwh
total_carbon_metric_tons = total_carbon_kg / 1000

print(f"Training energy: {total_energy_kwh} kWh")
print(f"Carbon footprint: {total_carbon_metric_tons:.2f} metric tons CO2")
print(f"Comparison: {total_carbon_metric_tons:.2f} tons = "
      f"{total_carbon_metric_tons * 1000 / 4.6:.0f} km of car driving")
    

For the same model in different regions:

• Iceland — 1.25 × 0.08 = 0.10 tons CO2
• Sweden — 1.25 × 0.15 = 0.19 tons CO2
• USA — 1.25 × 0.35 = 0.44 tons CO2
• India — 1.25 × 0.65 = 0.81 tons CO2

The same training is 8x dirtier in India than in Iceland. For a large model like GPT-scale, the carbon delta can be hundreds of tons.

Metrics beyond PUE: ERE, WUE, CUE

PUE alone is insufficient. The Green Software Foundation recommends additional metrics:

A datacenter with PUE=1.1 but 100% fossil is worse than a datacenter with PUE=1.4 but 90% renewables from a total environmental perspective.

Cooling strategies: natural vs mechanical

Cooling represents 30–50% of a datacenter's energy overhead (the rest is power distribution, uninterruptible power supply, lighting).

Cooling strategies (ordered by efficiency):

1. FREE COOLING (Ambient air)
   - Applicable only in cold climates (Iceland, Norway, Canada)
   - PUE potential: 1.05–1.15
   - No water, no climate risks

2. EVAPORATIVE COOLING (Adiabatic)
   - Requires dry climate (relative humidity < 50%)
   - PUE: 1.15–1.25
   - Minimal water use

3. WATER COOLING (Chilled water loops)
   - Global standard
   - PUE: 1.2–1.4
   - High water use (drought risk)

4. IMMERSION COOLING (Special fluid)
   - Emerging, high cost
   - PUE: 1.1–1.2
   - Very low water use
   - GPU/CPU submerged in dielectric fluid
    

Google recently experimented with immersion cooling to further reduce PUE in datacenters with high-density GPU.

How to choose a sustainable cloud/datacenter provider

If you are an organization doing AI model training, provider choice is critical. Here is a checklist:

CHECKLIST FOR CHOOSING A SUSTAINABLE DATACENTER

Tier 1 (Essential):
☐ PUE publicly reported (if < 1.25, good; < 1.15, excellent)
☐ Renewable energy commitment > 60%
☐ Carbon footprint publicly reported for your workload
☐ Geographic location with predominantly green grid

Tier 2 (Differentiator):
☐ ERE, WUE, CUE metrics publicly available
☐ Water usage reporting + drought resilience
☐ 3rd-party audit (Science-Based Targets initiative, Google Committed)
☐ Clear Net-Zero roadmap with deadline (e.g., 2030, 2025)

Tier 3 (Nice to have):
☐ Carbon offset program
☐ On-site renewable (solar panels on roof)
☐ Partnership with local renewable utilities
☐ Datacenter heat reuse (e.g., district heating)
    

Major cloud provider sustainability snapshot (2026):

• Google Cloud — Global PUE 1.10, 100% carbon-free energy commitment (24/7), maximum transparency. Best-in-class.
• Microsoft Azure — PUE 1.12–1.25 (varies by region), 100% renewable energy target by 2025 (stated), good transparency.
• Amazon AWS — PUE 1.2–1.3 (old datasheets), vague renewable energy commitment, lower transparency than Google/Microsoft.
• Oracle Cloud — PUE not publicly reported, low transparency, not recommended for green AI.

Implications for machine learning training

For a moderately-sized AI model (e.g., fine-tuning a 7B model on custom dataset):

# Scenario: Fine-tuning a 7B LLM

# Training parameters
training_hours = 48  # GPU hours on A100
gpu_power_draw = 310  # Watts per A100 at full compute
compute_energy = (training_hours * gpu_power_draw) / 1000  # kWh

datacenter_efficiency = {
    'Iceland': {'pue': 1.15, 'carbon_intensity': 0.08},
    'Sweden': {'pue': 1.25, 'carbon_intensity': 0.15},
    'USA': {'pue': 1.30, 'carbon_intensity': 0.35},
}

print(f"Compute energy required: {compute_energy:.1f} kWh\n")

for region, params in datacenter_efficiency.items():
    total_energy = compute_energy * params['pue']
    carbon = total_energy * params['carbon_intensity']
    print(f"{region}:")
    print(f"  Total energy (with overhead): {total_energy:.1f} kWh")
    print(f"  Carbon footprint: {carbon:.2f} tons CO2")
    print(f"  Cost in €: {total_energy * 0.12:.2f}")
    print()
    

Result:

• Iceland: 16.8 kWh, 1.34 tons CO2, €2.02
• Sweden: 16.8 kWh, 2.52 tons CO2, €2.02
• USA: 16.8 kWh, 5.88 tons CO2, €2.02

Energy cost is the same, but carbon footprint varies by 4.4x simply based on geographic choice.

Trends and investments in 2026–2027

What's changing in the datacenter market:

• GPU cooling specialists — Companies like CoolIT, Asetek are developing liquid cooling for dense GPU. Expected PUE below 1.1 by 2027.
• On-site renewable mandate — Microsoft, Google are investing in solar/wind directly on datacenter grounds to decouple from grid.
• Compute relocation patterns — Organizations are moving AI training to regions with better carbon intensity, even with latency trade-offs.
• Carbon-aware compute — Tools like Kubernetes' carbon-aware scheduler change job execution times to take advantage of low-carbon hours.
• Skepticism about carbon offsets — Growing demand for "real" reduction (true green) rather than low-quality offsets.

FAQ