AI Consumes Electricity, but Also Designs the Grid
The Duality of AI and Energy Through the Lens of the IEA Report and an Execution Framework
(1) Introduction — Why I Read This Report to the End
When I read technical reports, I often pay more attention to the “structure” than to individual “sentences.” The IEA’s “Energy and AI” report was no exception. AI appears as a power-consuming entity, but also as a tool for optimizing the grid. A dual role.
To me, this was not just a technical issue but a sign that the system itself demands both collision and redesign.
This article aims to
— Structurally summarize the core content of the report, and
— Lay out a step-by-step framework to translate it into an executable system.
(2) Key Summary of the Report — The Structure of Conflict and Coexistence Between AI and Energy
1. AI cannot exist without electricity, as data centers consume power at an industrial scale.
As of 2024, data centers consume approximately 415 terawatt-hours (TWh), which accounts for 1.5% of the global electricity usage. By 2030, this is projected to rise to 945 TWh — roughly equivalent to Japan’s entire electricity consumption. AI-specific data centers now consume more electricity than aluminum smelters. For instance, training GPT-4 required tens of thousands of NVIDIA A100 or H100 GPUs operating 24/7 for several weeks, demanding a load of several to tens of megawatts, which is comparable to the electricity consumption of over 100,000 households.
2. The structure of AI-related power consumption consists of training, inference, and cooling.
Training involves optimizing billions of parameters through repeated large-scale matrix multiplications, requiring parallel GPU processing. Inference generates outputs based on inputs, with energy loads increasing as tokens accumulate during continuous processing. Cooling is necessary to prevent server overheating, using fans, coolants, or insulating oils, and accounts for 30–40% of total energy use — highlighting the need for improved Power Usage Effectiveness (PUE). Cooling systems have evolved from air to water to immersion cooling, the latter achieving PUE values below 1.1, as seen in facilities operated by Meta and Microsoft.
3. Big Tech companies are moving away from reliance on national power grids by building proprietary energy systems.
Google operates solar and wind power plants with energy storage systems and is advancing a 24/7 carbon-free energy initiative. Microsoft is investing in Small Modular Reactors (SMRs) and collaborating with TerraPower. AWS secures wind energy via global power purchase agreements (PPAs) and runs its own wind farms in selected regions. Meta installs dedicated solar facilities near its data centers and incorporates immersion cooling systems. These initiatives aim to achieve energy independence, fulfill ESG (environmental, social, and governance) goals, and mitigate supply chain risks.
4. AI can strengthen the existing power grid, potentially unlocking 175 GW of capacity without requiring new infrastructure.
AI-based grid control allows real-time forecasting of renewable energy generation, which reduces surplus electricity. It also enables load shifting to alleviate congestion and uses dynamic line rating to adjust power transmission based on weather and wind conditions. For example, the U.S. operator TenneT implemented AI-based load shifting and reduced grid congestion by 15%. Through such innovations, AI allows smarter utilization of existing infrastructure without the need for large physical upgrades.
5. AI accelerates innovation in energy technology by a factor of ten.
In CO₂ capture, AI models such as graph neural networks and XGBoost predict high-performance adsorbent materials like MOFs and zeolites. In battery research, AI forecasts optimal chemical compositions and charging/discharging pathways. For synthetic fuels, large language models analyze scientific papers and patents to identify promising catalyst candidates. In heavy industries like cement and steel, reinforcement learning models optimize CO₂-reduction processes. MIT has shown that reinforcement learning can reduce energy use in CO₂ capture by 20%.
6. Despite its benefits, AI also emits carbon and requires a balanced approach.
Under the base-case scenario, data center emissions are projected to increase from 180 megatonnes (Mt) in 2024 to 300 Mt by 2035. In a worst-case “Lift-Off” scenario, emissions could rise to 500 Mt. At the same time, AI-driven optimization of energy systems could reduce total energy-related emissions by 5–7%. This reveals a duality: AI can both reduce and contribute significantly to greenhouse gas (GHG) emissions.
7. Explainable AI is essential for deployment in the energy sector.
In contexts like grid control, CO₂ processing, and energy policy-making, it is critical to explain the reasoning behind AI decisions. Models must present both predictions and the underlying logic using numerical data, clear reasoning, and visualizations to satisfy regulatory and operational requirements. While models like GPT, DNNs, and CNNs offer high performance, they are often black boxes. In contrast, tools like XGBoost with SHAP, autoencoders, and rule-based systems offer slightly lower performance but provide much higher trustworthiness and transparency.
8. Emerging markets hold high leapfrog potential in the AI-energy convergence.
Although many of these countries have a large base of internet users, they lack extensive data center infrastructure — creating an opportunity for leapfrogging. Realizing this potential requires stable electrical grids, widespread renewable energy, and a skilled digital workforce. AI can rapidly scale forecasting capabilities, improve power plant efficiency, and support local energy optimization. Countries like India, Indonesia, and Kenya are leading the way with AI-powered microgrids and digital energy storage systems.
Summary: AI is Both a Heavy Energy Consumer and the Strongest Catalyst for Energy Innovation. The key to future energy strategy lies in who controls and utilizes this duality.
(3) Execution Framework — Turning This Flow into a Structured System
1.Define Preconditions — What Needs to Be in Place? The report outlines an inspiring flow, but actual implementation requires more than a vision — it demands clear conditions across technology, infrastructure, and institutions.
① Technological Preconditions: AI Must Be Interpretable and Controllable. AI must move beyond simple prediction to explainable control. Especially in grid operations, carbon capture, and automated data centers, operators need to understand why AI made a decision.
Examples:
— LSTM + SHAP explanations
— GNN with feature attribution
— Rule-based corrective logic.
Accuracy alone is no longer enough — trust and explainability take precedence.
② Infrastructure Preconditions: AI Must Be Sensor-Connected and Bidirectional. AI cannot function without real-time data. Sensor-based infrastructure is essential.
Use cases:
— Grid systems: sensing temperature, load, wind, and solar radiation
— Facility management: vibration, heat, leak detection sensors.
AI must not be a passive calculator but an active controller in a digital twin framework. AI is the “brain”, but sensors and actuators are the “nervous system” — they must come first.
③ Institutional Preconditions: AI Must Be a Co-Pilot, Not a Commander. Public and safety concerns prevent full AI autonomy in energy systems. Therefore, AI should make interpretable suggestions, with final decisions approved by human operators or policymakers.
Institutional conditions:
— Transparent, explainable reports — Manual override protocols for anomaly detection
— Clear operator-AI interface standards. Executable AI must be designed as a “reviewable assistant,” not an omnipotent engine.
2.Assess Executability — Where and How to Start? Even with sound theory, execution depends on context. We must evaluate priority areas, practical bottlenecks, and geographic feasibility to avoid turning a strategy into wishful thinking.
① Deployment Priorities: Not All Systems Are Ready. We prioritize areas with high demand, clear structure, and human interpretability.
Get Deokjin Choi’s stories in your inbox
Join Medium for free to get updates from this writer.
Priority
— 1st: Renewable forecasting → Strong demand, clear prediction structure, direct link to decisions
— 2nd: Grid load shifting → Sensor-based, semi-controllable, integrable with current systems
— 3rd: CO₂ capture control → Technically complex, high cost, sensitive to regulation.
▶ Adoption depends more on operational trust than technical feasibility.
② Bottlenecks: People and Policy More Than Tech.
Bottleneck Description:
— Unexplainable AI → Black-box models fail to provide rationale → operator rejection
— Data access → Lack of high-resolution, real-time sensing and poor inter-system connectivity
— Policy rigidity → No protocols to validate/approve AI recommendations.
▶ The tech exists — but institutions and people are not fully prepared.
③ Geography Matters: Malaysia Leapfrogs, Korea Integrates.
— Malaysia / Emerging → AI-led microgrids + dashboard for operators → Flexible regulation allows bottom-up design
— Korea / Advanced → Integrated AI into legacy systems → Assistive control within rigid legacy structure
— Big Tech → Full-stack AI optimization of private energy systems → Own infrastructure and full data sovereignty.
▶ Execution hinges less on national identity and more on infrastructure control and data access.
3.System Structuring — What Would an Actual Implementation Look Like? The abstract “role of AI” must be translated into a connected system that links prediction, judgment, execution, and feedback in a continuous loop.
Example 1: AI Grid Optimization Loop (For Fixed Infrastructure).
— Step 1: Real-time sensing → Environment data: wind, temperature, load, power flow.
— Step 2: AI forecasting → LSTM for renewable energy prediction / GNN for power congestion forecasting.
— Step 3: AI judgment → Load shifting decisions / ESS (Energy Storage System) dispatching / SHAP-based model explanation.
— Step 4: Operator review → Dashboard interface for approval or manual intervention.
— Step 5: Execution & feedback → Action execution with data returned to retrain models.
Example 2: Leapfrog Microgrid Model (e.g., Malaysia).
— Step 1: Local energy setup → Deploy local renewables (solar, wind) and battery storage.
— Step 2: Edge AI forecasting → Use lightweight models (e.g., XGBoost) for real-time demand prediction.
— Step 3: Autonomous scheduling → Optimize energy use with explainable AI decision logic.
— Step 4: Operator approval → One-click dashboard confirmation or rule-based automation.
— Step 5: Grid integration → Connect to national grid or enable peer-to-peer (P2P) energy sharing.
▶ In infrastructure-limited regions, combine local optimization with explainable AI to reduce regulatory friction and accelerate adoption.
Design Principles — 3 Rules for Systemization:
1.Design “explanation” before prediction — transparency is more important than accuracy.
2. AI should assist, not dominate — assign AI the role of a co-pilot.
3.Establish a closed loop — Predict → Judge → Execute → Feedback → Retrain.
(4) Conclusion — How Should We Understand the Clash Between AI and Energy?
This report is not merely about technological progress. It reveals how AI and energy collide and cooperate at the same time, and why this tension demands a new kind of system design. AI is a dual entity — it consumes electricity and optimizes it. This duality is not just a technical challenge, but a structural one. Only through deliberate and thoughtful system design can we transition toward a sustainable energy future. Therefore, our goal should not be limited to deploying AI technologies. We must go further — to design executable, structured systems that make integration real. For AI to be truly embedded into the power grid, the system must be built with Explainability, Institutional trust, and Feedback loops at its core. The clash between AI and energy is a symptom of poorly designed systems. Our task now is not just to observe that conflict — but to resolve it through the creation of new, integrated architectures.
