Region selection for carbon

SAMPLE REGION INTENSITIES (gCO2/kWh, approximate):

VERY CLEAN (mostly renewable):
  eu-north-1 (Sweden: hydro):              ~50
  ca-central-1 (Quebec: hydro):             ~50
  europe-north1 (Finland, low-carbon):       ~30

MEDIUM:
  us-west-2 (Oregon: mixed hydro/renew):   ~120
  us-west-1 (California: solar-heavy):      ~250
  eu-west-1 (Ireland: wind):                ~180

HIGH CARBON:
  us-east-1 (Virginia: gas + some coal):    ~350
  us-east-2 (Ohio: coal-heavy):              ~400
  ap-southeast-2 (Sydney: coal-heavy):       ~700

EXTREMELY HIGH:
  ap-southeast-1 (Singapore: coal/gas):     ~500-600
  ap-east-1 (Hong Kong):                      ~550

LOW-CARBON CHARACTERISTICS:
  Hydroelectric grids                    Norway, Sweden, Quebec
                                        baseline ~30-50 gCO2/kWh

  Solar-heavy + low coal                 California, parts of Asia
                                        baseline ~150-250 gCO2/kWh

  Wind-heavy                            Texas, Northern Europe
                                        baseline ~100-200 gCO2/kWh

  Nuclear-heavy                         France, parts of E. Europe
                                        baseline ~50-100 gCO2/kWh

HIGH-CARBON CHARACTERISTICS:
  Coal-heavy grids                       Parts of US, Australia,
                                        Southeast Asia
                                        baseline 400-700 gCO2/kWh

  Gas-heavy grids                         Some US East regions
                                        baseline 300-450 gCO2/kWh

TRADE-OFF                                  IMPACT
──────────────────────────────────────────────────────────────────
LATENCY                                    Distance from users
                                          increases latency
                                          (acceptable for batch;
                                          not for real-time)

PRICING                                     Cleaner regions sometimes
                                          slightly pricier
                                          (variance: ±5-15%)

SERVICE AVAILABILITY                       Newer / smaller regions
                                          have fewer services
                                          (Lambda, Bedrock, etc.
                                          may not be everywhere)

DATA RESIDENCY                             Compliance requirements
                                          may mandate region
                                          (GDPR, HIPAA, etc.)

DATA TRANSFER                              One-time migration cost
                                          + ongoing inter-region
                                          transfer if hybrid

COMPLIANCE                                 SOC 2, ISO require
                                          documented data handling

WORKLOAD TYPE                  PRIORITY OF FACTORS
──────────────────────────────────────────────────────────────────
Real-time customer-facing       Latency >> Carbon >> Cost
                                Region near users wins

Batch processing                Cost ≈ Carbon (latency low priority)
                                Clean-grid region usually wins

Backup / archive                Cost ≈ Carbon
                                Cheapest clean region wins

ML training                     Cost > Carbon (some flex)
                                Mixed; depends on org's ESG goals

Development                     Cost > Carbon
                                Closer to engineers usually wins

Internal tools                   Cost > Carbon
                                Closer to office wins for latency

BEFORE MIGRATING regions for carbon reduction:

  □ Latency check: target region's latency still acceptable for
    customer experience?

  □ Cost check: comparable cost in target region?
    (allow ±10% acceptable; >15% may not justify)

  □ Service availability: all needed services present in target
    region? Check each cloud service in your stack.

  □ Data transfer one-time cost: how much to move existing data?

  □ Data transfer ongoing cost: will hybrid setup create new
    inter-region transfer charges?

  □ Compliance: data residency rules permit target region?

  □ Team familiarity: any operational concerns with new region?

  IF ALL GREEN: proceed with migration plan
  IF ANY YELLOW: investigate; may need different approach
  IF ANY RED: this workload not a migration candidate

PATTERN: primary + secondary in different regions
  Primary:    closer to users (cost + latency priority)
  Secondary:  cleaner grid (carbon priority)

WHEN PRIMARY HEALTHY: traffic stays primary
WHEN PRIMARY DEGRADED: traffic shifts to secondary
                       (now running at lower carbon during
                       a less-than-optimal period)

The secondary serves both DR and a carbon backup function.

TYPICAL ML training workload comparison:

  In us-east-1 (350 gCO2/kWh):
    100 GPU-hours × 1.5 kW × 0.35 kg/kWh = 52.5 kg CO2 per run

  In eu-north-1 (50 gCO2/kWh):
    Same 100 GPU-hours × 1.5 kW × 0.05 kg/kWh = 7.5 kg CO2 per run

  REDUCTION: 86% for the same compute

  Annual (if run daily):
    us-east-1: 52.5 kg × 365 = 19.2 tons CO2/year
    eu-north-1: 7.5 kg × 365 = 2.7 tons CO2/year
    Annual savings: 16.4 tons (86%)

MIGRATION PLAN typical steps:

  WEEK 1-2: Assessment
    Calculate carbon impact of current and target regions
    Verify latency / cost / compliance per checklist
    Estimate migration effort

  WEEK 3-4: Pilot
    Migrate one non-critical workload first
    Measure: latency, cost, carbon
    Validate against expectations

  WEEK 5-8: Production migration
    Move main workload in waves
    Maintain rollback capability

  WEEK 9-10: Verification + decommission
    Confirm target region performance matches
    Decommission source-region resources
    Document the migration in audit log + carbon report

EFFORT: 2-4 weeks per workload for clean migrations
ANNUAL CARBON IMPACT: 70-90% reduction typical

MISTAKE                                   FIX
──────────────────────────────────────────────────────────────────
Don't check latency for the new region    Test from user-representative
                                          locations before committing

Forget inter-region data transfer costs   Include in TCO; can be
                                          significant for data-heavy
                                          workloads

Migrate first; check compliance later     Compliance check is the
                                          first gate; some workloads
                                          can't move at all

Migrate everything to "save the world"     Some workloads have hard
                                          constraints; respect them

Underestimate operational change           New region means new ops;
                                          team familiarity matters

TEAM:    ML platform team
WORKLOAD: nightly ML training cluster
COST:    $40K/month
COMPUTE: 100 GPU-hours/day average

CURRENT STATE:
  Region: us-east-1
  Annual carbon: 18.4 tons CO2/year (350 gCO2/kWh)

CANDIDATE REGIONS:
  Option A: eu-north-1 (50 gCO2/kWh):
    Annual carbon: 2.6 tons CO2/year (86% reduction)
    Latency to ML pipeline: +50ms (acceptable; async batch)
    Cost: comparable ±2%
    Services available: all needed (verified)
    Compliance: no concerns (training data not subject to residency)
    Data transfer cost: ~$3K one-time migration

  Option B: us-west-2 (120 gCO2/kWh):
    Annual carbon: 6.3 tons CO2/year (66% reduction)
    Latency: +30ms
    Cost: comparable
    Services: all available
    Data transfer: ~$1K

DECISION ANALYSIS:
  Option A delivers 86% reduction; effort is higher (overseas)
  Option B delivers 66% reduction; effort lower
  Annual carbon saved:
    A: 15.8 tons
    B: 12.1 tons

DECISION: Option A (eu-north-1)
  Rationale: largest carbon reduction; cost neutral

EXECUTION:
  Week 1-2: detailed migration plan, pilot batch ML run
  Week 3-4: full production migration in 3 waves
  Week 5: decommission us-east-1 resources
  Total effort: 1 engineer × 5 weeks = ~$25K loaded

ROI on the migration effort:
  Carbon: 15.8 tons/year saved (significant for ESG report)
  Cost: roughly neutral (small data transfer cost amortized)
  Verdict: clear win for sustainability narrative

DOCUMENTED in carbon report; presented to leadership as a
positive ESG signal.

WORKLOAD candidate:    __________
  Class: batch / async / customer-facing / latency-sensitive

CURRENT region:    __________
  Carbon intensity: __________ gCO2/kWh (look up)

CANDIDATE cleaner region:    __________
  Carbon intensity: __________ gCO2/kWh

CARBON REDUCTION estimate:
  Current annual: __________ tons CO2
  Target annual: __________ tons CO2
  Reduction: __________ tons (____ %)

CHECKLIST:
  □ Latency acceptable for this workload?
  □ Cost variance within 10%?
  □ All services available in target?
  □ Compliance / residency permits target?
  □ Team familiar with operating in target region?

EFFORT estimate:    __________ weeks

DECISION:
  □ Migrate (carbon impact justifies)
  □ Defer (constraints unresolved)
  □ Not migratable (compliance / latency)

If migrate: schedule pilot for: __________