Global Banking Electricity Outlook
2026–2038 · The transition to machine intelligence
In the base scenario, annual electricity demand for a stylized portfolio of 100 large banking groups rises from 60 TWh in 2026 to 177 TWh in 2038, a 2.95× increase. Average demand rises from 6.85 GW to 20.22 GW. The accelerated scenario reaches 491 TWh and 56.05 GW.
Scope and starting point
The model represents 100 large banking groups across the world. It is not an enumerated ranking of the 100 largest banks and does not cover the entire banking sector. It includes offices, branches, ATMs, bank-operated computing and the electricity attributable to banking workloads at external providers, including cooling and power losses. Customer and borrower energy use, financed emissions and unrelated provider workloads are excluded.
The 60 TWh starting point consists of 50 TWh on bank-operated premises plus 10 TWh allocated to external computing. The 50 TWh estimate is constructed as 20 groups × 1.5 TWh + 30 × 0.5 TWh + 50 × 0.1 TWh. These are assumed size classes, not measured cohorts. A 40–80 TWh starting range is an expert sensitivity range, not a confidence interval.
Observed reference points
The following disclosures anchor the order of magnitude. They cover different years and operational boundaries; their 11.26 TWh sum is not a harmonized sector statistic. None validates the 60 TWh portfolio baseline or the assumed AI share.
| Bank | Reporting period | Electricity, TWh |
|---|---|---|
| Agricultural Bank of ChinaReported electricity | 2024 | 3.084 |
| JPMorgan ChasePurchased electricity; on-site solar reported separately | 2024 | 2.041 |
| Bank of AmericaReported operational electricity | 2024 | 1.827 |
| Wells FargoPurchased electricity plus own solar | 2025 | 1.479 |
| State Bank of IndiaRenewable and non-renewable electricity, converted from GJ | 2025/26 | 0.858 |
| SantanderGroup electricity | 2025 | 0.775 |
| HSBCReported coverage 98.2% of FTE; before scaling small offices | 2025 | 0.634 |
| MUFGGroup and consolidated entities within report scope | FY2024 | 0.561 |
Scenario results for 2038
| Indicator | Efficient | Base | Accelerated |
|---|---|---|---|
| Annual electricity, TWh | 83.5 | 177.1 | 491.0 |
| Multiple of 2026 | 1.39× | 2.95× | 8.18× |
| Average demand, GW | 9.53 | 20.22 | 56.05 |
| Indicative capacity provision, GW | 16.2 | 31.6 | 83.1 |
Twelve-year trajectory
| Year | Efficient Annual electricity, TWh | Base Annual electricity, TWh | Accelerated Annual electricity, TWh | Base average, GW | Base provision, GW |
|---|---|---|---|---|---|
| 2026 | 60.0 | 60.0 | 60.0 | 6.85 | 13.3 |
| 2027 | 60.7 | 61.6 | 62.5 | 7.03 | 13.5 |
| 2028 | 61.7 | 64.0 | 66.5 | 7.30 | 13.8 |
| 2029 | 63.0 | 67.4 | 72.9 | 7.69 | 14.2 |
| 2030 | 64.6 | 72.1 | 82.6 | 8.23 | 14.9 |
| 2031 | 66.1 | 77.7 | 95.9 | 8.87 | 15.8 |
| 2032 | 67.9 | 85.1 | 115.3 | 9.72 | 16.9 |
| 2033 | 70.1 | 94.9 | 143.7 | 10.84 | 18.5 |
| 2034 | 72.6 | 107.8 | 185.1 | 12.30 | 20.5 |
| 2035 | 75.6 | 124.6 | 245.3 | 14.22 | 23.2 |
| 2036 | 79.1 | 146.5 | 332.6 | 16.73 | 26.7 |
| 2037 | 81.2 | 160.7 | 403.0 | 18.35 | 29.0 |
| 2038 | 83.5 | 177.1 | 491.0 | 20.22 | 31.6 |
Model and assumptions
Let t be years after 2026. Building and other non-computing demand B starts at 35 TWh and falls 2.5% per year. Conventional computing L starts at 19 TWh and grows 2% per year. AI computing A starts at 6 TWh. All three starting allocations are assumptions. Total annual electricity is E = B + L + A.
L(t) = 19 × 1.02ᵗ
A(t) = A(t−1) × (1 + g)
E(t) = B(t) + L(t) + A(t)
Net annual growth in AI electricity
| Period | Efficient | Base | Accelerated |
|---|---|---|---|
| 2027–2030 | 20% | 35% | 50% |
| 2031–2036 | 15% | 30% | 45% |
| 2037–2038 | 8% | 15% | 25% |
These rates already include efficiency gains. Assuming a 25% annual improvement in useful AI work per kWh, the implied gross work growth is (1 + net energy growth) × 1.25 − 1. For example, base AI electricity growth of 35% corresponds to 68.75% more useful AI work each year. This is a modeling convention, not a measured universal efficiency law.
Average power is E / 8.76, using a standardized 8,760-hour year. Indicative capacity provision uses assumed load factors of 0.50 for buildings and 0.80 for computing, plus 15% headroom. It is total provision including headroom, not spare reserve, measured connected capacity or an engineering specification for N+1/2N resilience.
P = 1.15 × [B / (8.76 × 0.50) + (L + A) / (8.76 × 0.80)]
The external-provider share of computing rises linearly from 40% in 2026 to 60% in 2038. In the base case, 2038 demand divides into approximately 86.4 TWh on bank-operated premises and 90.8 TWh at external providers. The base computing multiplier is 6.05×, while the AI component alone rises 21.2×. These should not be confused with the whole-portfolio multiplier of 2.95×.
Interpretation and uncertainty
Adoption of advanced machine intelligence over the next decade is a scenario premise. The accelerated case explores unusually large computational demand; it is not a quantified forecast of superintelligence itself. Agent activity, training allocation, model architecture, hardware efficiency, regulation, energy availability and outsourcing could all materially change the path. No probabilities are assigned to the three cases.
Keeping base-case proportions and growth rates, a 40–80 TWh starting point produces approximately 118–236 TWh in 2038. This only tests baseline sensitivity; it is not the full uncertainty range. In particular, the accelerated case is highly sensitive to sustained compound growth and should be read as a stress scenario.
Planning implication: the base scenario suggests roughly three times the electricity and average power, and 2.38 times the indicative capacity provision. Rising demand alone does not establish a need for orbital energy or the commercial viability of any specific generation technology.
Review annually and when material bank disclosures, provider energy allocations or measured AI workload and efficiency data become available. Version changes should identify revisions to the baseline, population and growth assumptions.
Sources and calculation files
Bank disclosures provide scale references. Scenario construction, extrapolation and conclusions are the Echelon Group’s own assumptions and calculations. Sources do not endorse this outlook.
The workbook’s Russian label «Резерв мощности» refers to the model’s total indicative capacity provision including 15% headroom, as defined above.
- Agricultural Bank of China · Sustainability / green finance report 2024; environmental indicators
- JPMorgan Chase · 2024 Sustainability Report, p. 20. On-site solar separately 57,420 MWh.
- Bank of America · 2025 Sustainability Report, 2024 environmental data
- Wells Fargo · 2025 Operational Sustainability Performance Data, p. 2, Aug 2026
- State Bank of India · Sustainability Report FY 2025–26, energy indicators, p. 148
- Santander · Annual Report 2025, p. 130
- HSBC · ESG Datapack 2025, p. 11; coverage 98.2% FTE.
- MUFG · ESG Data Book 2025, pp. 4–6
- IEA · Key Questions on Energy and AI (2026) · Context: worldwide data centres, not a bank-specific forecast.