Geography is a Systems Variable

Same cluster, same model, same duration — but does location change the cost?

ops

intermediate

Compare identical training runs across four grid regions to discover whether geography matters more than hardware choice or training duration for carbon footprint.

The Question

You have a 256-GPU cluster training a model for 30 days. Does it matter where that cluster is located? Not for latency or throughput — those are fixed by the hardware. But for carbon emissions, water usage, and total cost of ownership, does geography matter — and if so, by how much?

Prerequisites

Complete Tutorial 1: The Memory Wall. No other prerequisites are required — this tutorial can be completed independently.

What You Will Learn

Calculate the carbon footprint of identical training runs in different regions
Quantify the gap between the cleanest and dirtiest electricity grids
Compare geography vs. training duration as levers for sustainability
Apply the EconomicsModel to show how carbon pricing changes the cheapest option

Background: Grid Carbon Intensity

Every kilowatt-hour of electricity has a carbon cost, measured in grams of CO2 per kWh (gCO2/kWh). This number depends entirely on how the electricity is generated:

Region	Primary Source	Carbon Intensity
Quebec	Hydroelectric	~20 gCO2/kWh
Norway	Hydroelectric	~29 gCO2/kWh
US Average	Mixed (gas, coal, renewables)	~390 gCO2/kWh
Poland	Coal-dominated	~820 gCO2/kWh

The range is wide. How wide — and whether it matters more than other levers like training duration or hardware choice — is what this tutorial quantifies.

1. Setup

import mlsysim
from mlsysim import Engine

2. Two-Region Comparison

Let’s run the same training job in two locations: Quebec (hydroelectric) and Poland (coal-dominated). Same fleet, same model, same 30-day duration. The only variable is where the electricity comes from.

from mlsysim import SustainabilityModel, Systems
from mlsysim.systems.types import Fleet, Node, NetworkFabric
from mlsysim.core.constants import Q_
from mlsysim.show import table, info

# 256-GPU cluster: 32 DGX H100 nodes
fleet = Fleet(
    name="256-GPU Training Cluster",
    node=Systems.Nodes.DGX_H100,
    count=32,
    fabric=Systems.Fabrics.InfiniBand_NDR
)

solver = SustainabilityModel()

# Quebec: hydroelectric grid
res_quebec = solver.solve(
    fleet=fleet, duration_days=30,
    datacenter=mlsysim.Infra.Grids.Quebec
)

# Poland: coal-heavy grid
res_poland = solver.solve(
    fleet=fleet, duration_days=30,
    datacenter=mlsysim.Infra.Grids.Poland
)

carbon_q = res_quebec.carbon_footprint_kg / 1000  # tonnes
carbon_p = res_poland.carbon_footprint_kg / 1000
ratio = carbon_p / carbon_q if carbon_q > 0 else 0

table(
    ["Region", "Carbon (tonnes CO2)"],
    [
        ["Quebec (Hydro)", f"{carbon_q:.1f}"],
        ["Poland (Coal)", f"{carbon_p:.1f}"],
    ]
)
info(Ratio=f"{ratio:.0f}x")

Region          Carbon (tonnes CO2)
───────────────────────────────────
Quebec (Hydro)                  2.7
Poland (Coal)                 167.2
Ratio:  61x

Same cluster. Same model. Same duration. The carbon footprint differs by roughly 40x depending on the electricity grid. This is not an optimization — it is a location decision.

3. All-Region Sweep

Let’s expand the comparison to all four grid regions in the Infrastructure Zoo, adding energy consumption, water usage, and PUE to the picture.

grids = [
    mlsysim.Infra.Grids.Quebec,
    mlsysim.Infra.Grids.Norway,
    mlsysim.Infra.Grids.US_Avg,
    mlsysim.Infra.Grids.Poland,
]

region_results = {}
rows = []
for grid in grids:
    r = solver.solve(fleet=fleet, duration_days=30, datacenter=grid)
    energy_mwh = r.total_energy_kwh.magnitude / 1000
    carbon_t = r.carbon_footprint_kg / 1000
    water_kl = r.water_usage_liters / 1000
    region_results[r.region_name] = r
    rows.append([r.region_name, f"{energy_mwh:,.1f}", f"{carbon_t:,.1f}", f"{water_kl:,.1f}", f"{r.pue:.2f}"])

table(["Region", "Energy (MWh)", "Carbon (t)", "Water (kL)", "PUE"], rows)

Region          Energy (MWh)  Carbon (t)  Water (kL)   PUE
──────────────────────────────────────────────────────────
Quebec (Hydro)         136.8         2.7         0.0  1.06
Norway (Hydro)         136.8         1.4         0.0  1.06
US Average             144.5        62.0       260.1  1.12
Poland (Coal)          203.9       167.2       366.9  1.58

Notice that energy consumption also varies between regions because of different PUE values. A modern liquid-cooled facility (PUE 1.1) wastes less energy on cooling than a legacy air-cooled datacenter (PUE 1.6). But the dominant factor is carbon intensity — it creates the 40x gap.

4. Geography vs. Training Duration

Is it better to train longer in a clean region or shorter in a dirty region? Let’s compare 30 days in Quebec against just 10 days in Poland.

# 30 days in Quebec
res_30d_quebec = solver.solve(
    fleet=fleet, duration_days=30,
    datacenter=mlsysim.Infra.Grids.Quebec
)

# 10 days in Poland (1/3 the training time)
res_10d_poland = solver.solve(
    fleet=fleet, duration_days=10,
    datacenter=mlsysim.Infra.Grids.Poland
)

c_q = res_30d_quebec.carbon_footprint_kg / 1000
c_p = res_10d_poland.carbon_footprint_kg / 1000

table(
    ["Scenario", "Carbon (tonnes CO2)"],
    [
        ["30 days in Quebec", f"{c_q:.1f}"],
        ["10 days in Poland", f"{c_p:.1f}"],
    ]
)
info(Ratio=f"{c_p/c_q:.1f}x")

Scenario           Carbon (tonnes CO2)
──────────────────────────────────────
30 days in Quebec                  2.7
10 days in Poland                 55.7
Ratio:  20.4x

Key Insight

Geography is a larger lever than training duration for carbon footprint. Even training for one-third the time in Poland produces more carbon than the full 30-day run in Quebec. The carbon intensity gap between hydro and coal grids is so large that no reasonable reduction in training time can compensate. For any organization serious about sustainable AI, datacenter location is not a logistics detail — it is a first-order systems design decision with 40x impact.

5. Economic Angle: When Carbon Has a Price

What happens when carbon emissions carry a financial cost? Carbon pricing (through taxes or cap-and-trade) changes the economics of datacenter location. Let’s compute TCO with a carbon price of $50/tonne.

from mlsysim import EconomicsModel

econ = EconomicsModel()
carbon_price = 50  # USD per tonne CO2

rows = []
for grid in grids:
    tco = econ.solve(fleet=fleet, duration_days=30, grid=grid)
    carbon_cost = (tco.carbon_footprint_kg / 1000) * carbon_price
    total = tco.tco_usd + carbon_cost
    rows.append([tco.region_name, f"${tco.tco_usd:,.0f}", f"${carbon_cost:,.0f}", f"${total:,.0f}"])

table(["Region", "TCO ($)", "Carbon Cost ($)", "Total ($)"], rows)

Region             TCO ($)  Carbon Cost ($)   Total ($)
───────────────────────────────────────────────────────
Quebec (Hydro)  $7,727,973             $137  $7,728,110
Norway (Hydro)  $7,727,973              $68  $7,728,042
US Average      $7,728,902           $3,100  $7,732,002
Poland (Coal)   $7,736,025           $8,358  $7,744,383

At $50/tonne, carbon pricing adds a visible cost differential between regions. At higher carbon prices (some jurisdictions already charge $100+/tonne), the difference becomes even more pronounced, potentially shifting which region offers the lowest TCO.

Your Turn

Exercises

Exercise 1: Predict before you compute. Training for 30 days in Quebec vs. 10 days in Poland — which produces more carbon? Write your prediction, then run both scenarios. Were you right? What does this tell you about the relative magnitude of grid carbon intensity vs. training duration?

Exercise 2: At what carbon price does geography change the cheapest option? Sweep carbon price from $0 to $500/tonne in steps of $50. For each price, calculate the total cost (TCO + carbon cost) for all four regions. At what price does a region other than the default cheapest become the best option? Print a table showing the crossover.

Exercise 3: Sweep PUE from 1.0 to 2.0. Create custom grid profiles using from mlsysim.infra.types import GridProfile with US Average carbon intensity but varying PUE. Sweep PUE from 1.0 to 2.0 in steps of 0.1. How much does total energy increase? At what PUE does facility overhead exceed the IT energy itself?

Self-check: If you train for 30 days in Quebec (20 gCO2/kWh) vs. 15 days in Poland (820 gCO2/kWh), and both use the same fleet and power, which produces more total carbon? Show the mental calculation: the ratio of carbon intensities is 41x, and the ratio of durations is 2x, so Poland is still 41/2 = ~20x worse.

Key Takeaways

Summary

Grid carbon intensity creates a 40x gap between the cleanest (Quebec, ~20 gCO2/kWh) and dirtiest (Poland, ~820 gCO2/kWh) regions
Geography dominates training duration as a sustainability lever: 10 days in Poland emits more than 30 days in Quebec
PUE amplifies energy use but carbon intensity is the dominant factor in emissions
Carbon pricing changes the economics: at $50-100/tonne, location becomes a financial variable, not just an environmental one
Datacenter location is a systems design decision with first-order impact on sustainability and, increasingly, on cost

Next Steps

The $9M Question – Quantify the infrastructure cost of chain-of-thought reasoning
Scaling to 1000 GPUs – Discover the hidden reliability cost at scale
Sensitivity Analysis – Use sensitivity sweeps to find which parameter matters most
Infrastructure Zoo – Browse all regional grid profiles and datacenter configurations