AI for Public Good Open evidence & tools · ai4publicgood.com

AI and climate change · Who pays for big AI's energy demand?

The kind of AI that adds load — and the kind that removes it.

The AI conversation is dominated by the electricity that hyperscale data centres will consume. But there is a second, documented story: AI already deployed inside the energy system is measurably cutting demand, waste and emissions. This briefing sets both sides of the ledger, then derives Fair Energy and Public-Benefit Principles for AI Infrastructure from the evidence.

The load AI adds
≈ Japan

IEA analysis projects global data-centre electricity use could double by 2030 — by then consuming as much electricity in a year as Japan does today.

The load AI can remove
> 13 EJ / yr

Widespread adoption of already-commercial AI optimisations could save over 13 exajoules a year by 2035 — more than Indonesia's entire annual energy consumption (IEA, Energy and AI, 2025).

scroll ▾

The first open tool that turns AI's energy demand into a fair, auditable deal — not a debate.

Built for Regulators — stress-test an interconnection request Utilities — size the negawatt program Communities — quantify who pays AI operators — arrive with commitments, not promises
01 · Framing

Two entries in the same ledger

Both columns are real. The policy question is not whether AI uses energy — it is whether the balance of deployment, and of who finances each column, is fair.

▲ What big AI adds≈ a Japanof data-centre electricity demand by 2030
▼ What deployed AI removes≈ an Indonesiaof annual energy, savable by 2035 with commercial AI optimisation
◂▸

move the cursor · drag on touch · arrow keys

▲ AI that adds load
  • Hyperscale training & inference Data-centre electricity use on track to double by 2030 — the equivalent of adding a Japan to the world's grids. IEA, Energy and AI (2025)
  • Publicly financed foundations Grids, transmission upgrades, water systems and permitting that serve private AI facilities are often recovered from all ratepayers. Session premise · under review
  • Capacity competition Large facilities compete for scarce connection capacity with housing, manufacturing and hospitals; some utilities have delayed other connections. Session premise · under review
  • Scale-rewarding incentives The industry rewards larger models and more compute, even where smaller, domain-specific models would serve the task at a fraction of the energy. Session premise · under review
▼ AI that removes load
  • Buildings become optimisers Deployed systems deliver 15–30% energy savings and ~20% peak-demand cuts across offices, hospitals, campuses and homes. Casebook cases 10–12 · Observatory
  • Grids waste less AI forecasting, mapping and dispatch cut distribution losses, curtailment and penalty-driven balancing — 95% forecast reliability, ~50% lower deviation penalties. Casebook cases 2–4
  • Transport avoids fuel Fuel-efficient routing alone avoided an estimated 2.7 Mt CO₂e in 2024; smart depot charging cut peak charges 22%. Casebook cases 8–9
  • Industry burns less Steel, cement and refining deployments cut fuel per tonne, rework and flaring — thousands of tonnes of CO₂ avoided per site, per year. Casebook cases 13–14 · Observatory

Documented demand reductions from deployed AI

Energy or peak-demand reduction per deployment, midpoint of reported range · self-reported figures, IEA Casebook on AI in Energy & IEA Energy and AI Observatory (2025–26)

Energy consumptionPeak demand / peak charges
02 · Evidence base

27 deployments, measured outcomes

Case studies from the IEA × IndiaAI Mission Casebook on AI in Energy (15 peer-screened deployments, published Feb 2026) and the IEA Energy and AI Observatory. Figures are as reported by the deploying organisations. Filter by sector; expand any card for the underlying numbers.

03 · Practice under review

Who pays, today

Four cost-shift mechanisms the board is asked to review. These are the session's premises — the evidence base above shows the counterfactual: load-removal is deliverable now, so the current allocation is a choice, not a necessity.

Mechanism A

Ratepayers finance private compute

Utilities recover the cost of grid, transmission, water and permitting upgrades from all customers — households and local businesses indirectly finance infrastructure primarily serving private AI companies.

Mechanism B

Capacity is reserved away from the public

Large AI facilities compete with housing, manufacturing and hospitals for scarce electricity; in some regions new industrial or residential connections have been delayed because capacity is held for data centres.

Mechanism C

Costs are local, returns are global

Carbon, water consumption and land use are borne by host communities, while the economic returns accrue globally. Corporate emissions are measured; local impact and community resilience rarely are.

Mechanism D

Scale is rewarded over efficiency

Bigger models and more compute are the industry's default success metric — yet many applications are served effectively by smaller, domain-specific or locally deployed models needing far less energy.

Can we deploy AI that removes load rather than adding it — and find a financial model that supports it?

The evidence section answers the first half: yes, at commercial scale, today. The principles below propose the second half — the financial and governance model.

Out of scope for this board: comparison of specific technologies or AI models · data sovereignty.

04 · The India lens

If the principles work anywhere, they must work in India

One of the world's largest energy markets and its fastest-growing major AI adopter — where the trade-off between AI infrastructure investment and energy access is not theoretical. 12 of the casebook's 15 deployments are Indian: India is simultaneously where the pressure is highest and where load-removing AI is already proven.

Demand pressure
+50%

Growth in India's electricity consumption 2020–2025 — from cooling, appliances, industry, EV fleets and data centres, all competing for the same buildout.

Recoverable waste
Mid-teens

India's aggregate technical & commercial (AT&C) distribution losses, vs 5–8% in peer economies. The gap is a multi-gigawatt resource that AI grid tools can recover — no new generation required.

Digital rails already laid
20 crore

Smart meters sanctioned under RDSS (≈5 crore installed by Dec 2025) — the data layer that AI loss-mapping, indexing and demand-response tools plug into.

▲ The trade-off, sharpened
  • Same capex, two claimants Every crore of grid, transmission and water buildout that serves hyperscale compute is a crore not serving last-mile connections, cooling-driven residential demand or industrial electrification.
  • Distribution utilities are already stressed AT&C losses raise procurement costs, weaken utility finances and delay payments to generators — socialising data-centre infrastructure costs onto this system compounds an existing fragility. Casebook, case study 3
  • Cross-subsidy politics Indian tariffs deliberately shift costs between consumer classes to protect households and agriculture. Where large AI loads enter as ordinary industrial consumers, the "who pays" question lands directly on that settlement.
▼ Why load-removal pays more in India
  • Loss reduction beats new build Closing even part of the mid-teens → 5–8% loss gap frees more capacity per rupee than new generation — and AI mapping/forecasting is how utilities in Maharashtra, Tamil Nadu and Delhi are doing it. Casebook, cases 3 & 6
  • Renewables need AI to be bankable With ~135 GW solar heading toward 280 GW by 2030 and non-fossil at 50% of installed capacity, forecasting accuracy is revenue: halved deviation penalties across 500 MW, ~95% dispatch reliability. Casebook, cases 2 & 4
  • Efficiency serves access Buildings, depots and microgrid deployments (15–30% savings, 22% peak cuts) defer capacity buildout — leaving headroom for connections, not just compute. Casebook, cases 5, 8, 10–12

India already has the policy rails the principles need.

Additionality and queue protection → connection and open-access rules under State ERCs. Cost causation → dedicated HT tariff categories. Net-load responsibility → RDSS and the Energy Efficiency Financing Platform. Flexibility duty → time-of-use tariffs and formal demand response (the casebook's own ask for e-bus depots, via an India Energy Stack). Right-sizing → IndiaAI Mission compute procurement standards. Metering the bargain → smart-meter data access. The principles are not new institutions for India — they are a fairness test applied to programmes already running.

05 · Digital twin

Model the deal, not just the debate

Three linked simulations of one AI-infrastructure deal. Configure a hypothetical data centre for your region; the twins compute the energy balance, the household bill impact, and the facility's behaviour across a day on the grid — using per-unit coefficients from the deployments in the evidence base. All three feed the generated term sheet.

Twin 01 · The energy balance

The deal on the table

data centres draw near-constant power

The load-removal portfolio

≈0.26 GWh saved per building·yr — casebook cases 10–12
AI grid mapping & forecasting — cases 3, 4, 6
depots, HVAC fleets, demand response — case 8, Voltalis

What the twin computes

Facility demand
Household equivalents
Portfolio removes
Emissions avoided

Principles scorecard

Net-load responsibilitycomputed
Flexibility at scalecomputed

Twin 02 · Who pays the bill

The infrastructure behind the facility

grid, transmission & water works — adjust to your jurisdiction

Who absorbs it

Total network cost
Socialised · per household
Dedicated tariff · operator pays
Households relieved

Twin 03 · A day on the grid — flexibility

workload moved from the evening peak into the solar window

Model assumptions (illustrative — replace with your utility's data)
  • Facility energy = MW × 8 760 h × load factor. Peak ≈ nameplate MW.
  • India profile: ~1 200 kWh/household·yr, grid factor 0.71 tCO₂/MWh, utility throughput 20 TWh/yr, system peak ~3.8 GW, AT&C losses mid-teens, ₹83/$. Mature profile: ~9 500 kWh/household·yr, 0.35 tCO₂/MWh, 30 TWh/yr, peak ~6 GW, losses ~6%.
  • Building lever: 1.2 GWh/building·yr average use × 22% AI savings (casebook cases 10–12, SISAB, Infosys). Peak: 0.05 MW shed per building.
  • Loss lever: 1 percentage point recovered = 1% of utility throughput (cases 3 & 6).
  • Flexibility levers count against peak, not energy (case 8, Voltalis).
  • Twin 02: network upgrade costs for large loads span roughly $0.5–3M per MW depending on jurisdiction and grid headroom; straight-line recovery, no financing cost. Socialised recovery spreads cost across all customers; a dedicated tariff assigns it to the causer.
  • Twin 03: stylised demand and solar curves (evening-peak system); flexible mode moves the selected share of the facility's evening-peak (18:00–22:00) draw into the 09:00–16:00 solar window, energy-neutral across the day.
  • This is a policy-illustration model, not an engineering study. Deployment coefficients are self-reported figures from the IEA Casebook (CC BY 4.0).
06 · Expected output

Fair Energy and Public-Benefit Principles for AI Infrastructure

A draft for the board's consideration. Each principle is anchored either in the deployment evidence (teal) or in the practices under review (amber).

Additionality of supply

New hyperscale AI load is matched by new clean generation and grid capacity procured by the operator — it does not draw down capacity needed for housing, hospitals, manufacturing or existing industry. Connection queues protect public-interest load.

AnchorsMechanism B

Cost causation, not cost socialisation

Whoever causes infrastructure cost pays it. Grid, transmission and water upgrades built primarily to serve AI facilities are financed through dedicated tariff classes and long-term take-or-pay commitments — not recovered from all ratepayers.

AnchorsMechanism A

Net-load responsibility

Operators of load-adding AI co-finance load-removing AI in the host grid — a "megawatts-for-negawatts" obligation. The evidence shows removal is bankable now: 15–30% building savings, ~20% peak cuts, halved deviation penalties. A demand-reduction obligation on large AI loads is therefore a financeable instrument, not a tax on innovation.

Anchors

Flexibility is a duty of scale

Any multi-megawatt load — data centre or depot — operates as grid-responsive demand: shifting, shedding and storing at times of system stress, under time-of-use tariffs and formal demand-response mechanisms, rather than taking firm 24/7 power.

Anchors

Local costs count, locally

Carbon, water, land and grid stress are measured and compensated where they occur. Permits carry benefit-sharing: waste-heat reuse, water stewardship, community energy funds and resilience investment in the host region — accounted alongside corporate emissions.

AnchorsMechanism C

Right-size the model

Public procurement and policy reward energy per outcome, not parameter count. The casebook's own selection finding: high-impact AI in energy is use-case driven, explainable and resource-efficient — running on edge devices, local controllers and modest cloud footprints, not hyperscale clusters.

Anchors

Meter the bargain, publicly

Facility-level public reporting of electricity, water and grid impact — and independently verifiable accounting of load removed — so regulators and communities can audit whether an operator's AI portfolio is net-adding or net-removing. What is metered can be governed.

AnchorsMechanism C