Our Blog

Blog Index 

IEA Warns Data Centre Electricity Use Could Double by 2030 — Power Grids Are the New AI Bottleneck

Posted on 19th Nov 2026 14:06:57 in Artificial Intelligence, Machine Learning

Tagged as: AI energy demand, data centres, IEA, power grid, India data centres

Every conversation about artificial intelligence eventually arrives at chips. The International Energy Agency's World Energy Outlook 2026, published on 19 November, argues that the more urgent shortage may be somewhere far less glamorous: the power line outside the data centre fence. In the agency's Stated Policies Scenario — built on measures governments have announced but not yet fully adopted — global electricity consumption by data centres could reach about 1,000 terawatt-hours (TWh) a year by 2030, and the grids that must carry that load are not being built fast enough.

The warning lands at an awkward moment. AI capital expenditure is at a record high, Anthropic has just priced the largest IPO on record, and every hyperscaler is racing to add capacity. The IEA is not predicting that the buildout stops. It is saying the buildout now depends on physical systems — transformers, copper, cables, connection queues and local consent — that move on a completely different clock from software.

What the IEA actually said

The headline number is contained in the Stated Policies Scenario: data centre electricity use could roughly double to around 1,000 TWh a year by 2030, with tight supply chains, local opposition and limited grid capacity potentially delaying new projects. The agency's related update, Key Questions on Energy and AI, puts the central projection slightly lower at about 950 TWh in 2030, up from 485 TWh in 2025 — roughly 3% of global electricity demand by the end of the decade.

The growth rate matters more than the level. Global data centre electricity demand grew 17% in 2025, in line with IEA projections, while consumption from AI-focused data centres surged 50% in the same year. The IEA's satellite tracking shows that "AI factories" — data centres built specifically for AI workloads — have more than tripled in capacity in the past 18 months. In its central projection, electricity use from AI-focused facilities triples between 2025 and 2030.

The report also identifies where the strain is already visible. The IEA cites large project pipelines in Texas, Virginia, Finland and the United Kingdom, and notes that in several regions the combined capacity of data centres waiting to connect to the grid now exceeds current peak demand. That is a queue problem measured in years, not months.

Why the grid, not the chip, is now the constraint

Three numbers from the outlook explain the bottleneck.

First, 26 million kilometres. That is how much new transmission line the IEA says grids need by 2035 under the Stated Policies Scenario, alongside the replacement or refurbishment of about 16 million kilometres of existing lines. For scale, the Earth's circumference is roughly 40,000 kilometres, so the required build-out is equivalent to running new cable more than 600 times around the planet.

Second, 40% to 50%. That is the projected rise in peak electricity demand by 2035, a figure that describes the maximum load a system must be able to serve on its worst hour — the most expensive hour to engineer for. Data centres are unusually demanding on this measure because their load is dense, arrives in multi-hundred-megawatt steps, and is uncertain: operators routinely oversize their initial grid connections as halls fill up gradually.

Third, 75%. Under the Current Policies Scenario, which reflects only rules already in force, grid spending is projected to rise by about 75% over the next decade. The IEA calls power networks "one of the biggest energy investment stories of the next decade" — a polite way of saying the money has not been spent yet.

The supply chain backs that up. Copper, cables and transformers are already tight, and the agency flags that some electricity technologies depend on critical inputs from a small number of producers, notably China. Gas turbine orders jumped 70% in 2025. A shortage of high-bandwidth memory, essential for AI accelerators, has developed over the past six months and is expected to persist through at least the end of 2027.

Developers are responding by going off-grid where they can. In the United States, around one-fifth of projects relying on onsite natural gas generation have started land clearing or construction. The IEA's analysis finds that reliably serving variable AI load with onsite gas requires overbuilding generation capacity by 30% to 70% relative to demand, and that 15 to 27 GW of onsite gas may power data centres by 2030, mostly in America. Even so, most operators still prefer a grid connection — which means the queue remains the main event.

Efficiency is improving fast — and demand is rising faster

It would be easy to read the IEA's numbers as evidence that AI is uniquely wasteful. The report is more interesting than that. Measured per task, AI is becoming dramatically more efficient: software and hardware advances have cut energy use per AI task by at least an order of magnitude annually in recent years. Simple text queries now typically consume less electricity than running a television over the same period. If every conventional internet search were performed as a simple AI text query, total consumption would be under 4 TWh a year — less than 1% of data centre electricity use today.

The problem is what comes next. Video generation, extended reasoning and agentic workflows can consume hundreds or thousands of times more energy per query than simple text generation. Demand is therefore the product of three moving parts: falling energy per task, surging usage, and models that keep unlocking heavier use cases.

Density compounds the challenge. An individual server rack inside an advanced data centre is about the size of a large refrigerator, but by 2027 it could carry peak power demand equivalent to 65 households. The IEA calculates that AI server power density rose 11-fold between 2020 and 2025 and is set to increase a further fourfold by 2027. These racks also swing by more than 50% of rated capacity within a second, need to evacuate heat equivalent to 30 natural gas boilers per rack, and can weigh more than a pickup truck. That is why battery storage — potentially 20 to 25 GW installed at data centres globally by 2030 — is becoming part of the AI infrastructure stack rather than an optional extra.

India's data centre power bill: 13.56 GW by 2031-32

For Indian readers the global projection is not abstract. In a March 2026 reply in the Rajya Sabha, the Ministry of Electronics and IT reported that national data centre capacity rose from about 375 MW in 2020 to around 1,500 MW by 2025, and that the Ministry of Power estimates electricity demand from data centres will reach 13.56 GW by 2031-32. The same reply noted that 38,231 GPUs have been onboarded through 14 empanelled providers at a subsidised average of ?65 per hour — roughly one-third of the global average cost — and that the SHANTI Act, passed to support small modular and micro nuclear reactors, is intended to give AI and data centre loads a firmer power option.

Independent research points in the same direction. Wood Mackenzie forecasts India's operational data centre capacity rising from 2.2 GW in 2025 to 12 GW by 2030, a compound annual growth rate near 40%, with AI-dedicated capacity expanding almost 24-fold from 275 MW to 6,546 MW. Its electricity forecast is blunter still: data centre power demand growing 20-fold by 2040 from about 10 TWh in 2025, reaching roughly 7% of total electricity demand.

Note that these are different metrics. The Ministry of Power figure describes electricity demand; Wood Mackenzie's 12 GW describes installed IT capacity. Both point at the same constraint.

The most useful line in the Wood Mackenzie release is about what now decides where projects get built. Reliable, cost-competitive power has overtaken land and capital as the primary limiting factor. "Land and capital are no longer the limiting factors for data centre developers in India. What determines site selection and delivery timelines now is access to firm, round-the-clock power at the node level," said Dr Rashika Gupta, Vice President of Research. Maharashtra and Tamil Nadu account for around 65% of installed IT load today, with Andhra Pradesh, Telangana, Uttar Pradesh and Karnataka expected to absorb the next wave.

IEEFA's analysis adds the local dimension: an average data centre consumes as much power as an aluminium smelter or roughly 100,000 homes, and India's data centre capacity could grow from 1.4 GW to 9 GW by 2030, consuming about 3% of national electricity, up from less than 1% today. Because these facilities cluster in a handful of hubs, the strain is concentrated on local networks rather than spread evenly across the country. Water is the second underappreciated constraint, particularly in Tamil Nadu and Karnataka.

Five things Indian businesses should watch

Most Kambalaweb readers will not build a data centre. But every business buying cloud, colocation or AI inference is about to feel these constraints in its bills and contracts.

  • Power is now a procurement question. States with liberalised open-access frameworks and competitive intra-state transmission charges are becoming the preferred locations for new capacity. Where your provider builds determines its power cost — and eventually your price.
  • Read the energy clause in colocation and cloud contracts. Tariff pass-through, minimum committed load and connection-capacity charges are where power inflation shows up. A contract signed on today's tariff without an indexation clause is a bet on electricity prices staying flat.
  • Expect price divergence between regions. If grids are the constraint, then the same GPU-hour will cost different amounts in different states and countries. Workload placement — latency-tolerant training versus latency-sensitive inference — becomes a finance decision, not just an engineering one.
  • Efficiency is a direct cost lever. Model routing, caching, smaller models for simple tasks and batching all reduce both compute and energy per task. The IEA's own finding — that energy per AI task is falling by an order of magnitude a year — is available to your team, not just to frontier labs.
  • Watch the policy signals. Connection-queue reform, non-firm grid connections with demand-response obligations, time-of-use tariffs and storage incentives such as viability gap funding for standalone battery projects all change the economics of where AI workloads run. The IEA's recommended package is exactly this: manage project pipelines, reward flexibility, and allocate grid costs transparently.

The bottom line

The AI trade of the last two years was priced on silicon. The next phase will be priced on electrons, transformers and interconnection rights. The IEA is explicit that with the right conditions — better demand disclosure, faster permitting, flexible connections and fair cost allocation — data centres need not raise electricity prices for everyone else. But that is a policy outcome, not a default. For Indian businesses planning AI spending through 2030, the practical assumption is that compute supply will be shaped as much by state power policy and grid capex as by the next chip generation.

Sources

whatsapp me