> ## Content Index > Fetch the complete content index at: https://www.sotkn.com/llms.txt > Use this file to discover other available public pages before exploring further. # How Much Power Does an AI Data Center Use? Read the Meter - URL: https://www.sotkn.com/explained/how-much-power-does-an-ai-data-center-use/ - Published: 2026-10-31T13:27:39.000Z - Updated: 2026-10-31T13:27:39.000Z - Description: A large AI data center is built to draw from about 100 MW to more than 1 GW, and the biggest running campus could use about as much electricity in a year as a million American homes. Most data centers draw far less than their grid hookup, and that gap decides who gets paid. - Author: Sean - Tags: Mechanism, #desk-explained EXPLAINED · Mechanism · Last reviewed: 31 Oct 2026 · Next review: after TNB’s third-quarter briefing and the IEA’s next data centre electricity update A large AI data center is built to draw anywhere from about 100 MW to more than 1 GW of power. The biggest one running today, Amazon’s New Carlisle campus in Indiana, is estimated at about 1,244 MW for the whole site, and if it ran flat out for a year it would use about 10.9 TWh, about as much electricity as a million American homes (our arithmetic). In practice most data centers use far less than the grid hookup on their gate: Australia’s grid operator found its operating data centers used about 27% of their maximum connection capacity in early 2026. **In short.** A large AI data center is built for 100 MW to more than 1 GW. The biggest running campus, New Carlisle in Indiana, could use about 10.9 TWh a year flat out, about a million American homes’ worth (our arithmetic). But there are two numbers on every site: the grid hookup on the gate and the energy on the meter. Data centers run steadily, at load factors of 75–90%, yet use only a fraction of their hookup while they fill: about 27% in Australia in early 2026 and about 22% in Malaysia in June 2026 (our arithmetic). Worldwide, data centers used about 485 TWh in 2025, slightly more than 1.5% of all electricity, but 20–30% in the places they cluster, and 23% of Ireland’s metered power. Who gets paid: utilities and wire builders increasingly on the hookup, generators on the meter, landlords on the kilowatts reserved. The stuck step is the hookup itself. Why an investor cares: power is the input every AI data center is short of, and the bill for it is split between the hookup and the meter. Whoever is paid on the hookup gets paid early; whoever is paid on the meter waits for the halls to fill. This page builds on [what an AI data center is](https://www.sotkn.com/explained/what-is-an-ai-data-center/) and answers one question: how much power does an AI data center really use? ## How much power does an AI data center use? A large AI data center is designed for somewhere between about 100 MW and several gigawatts. The International Energy Agency studied 21 large AI-focused data centers in the United States and found they range from 600 MW to 5 GW in nameplate capacity, each using as much electricity as 600,000 to 5 million households once complete. The largest one actually running is Amazon’s campus in New Carlisle, Indiana, built for Anthropic. Epoch AI, a research group that tracks big AI sites from satellite images and permits, estimates its computers can draw about 1,037 MW and the whole site about 1,244 MW once cooling and other overhead are added. If it ran at that level for a whole year it would use about 10.9 TWh, about what a million American homes buy, at the US Energy Information Administration’s average of 10,791 kWh per home (our arithmetic). It got there quickly. Epoch’s timeline puts the site at about 398 MW of IT power in June 2025, 626 MW in December 2025, 910 MW in March 2026 and 1,037 MW in October 2026, with about 1,925 MW projected by March 2028\. Our [ranking of the biggest AI data centers](https://www.sotkn.com/explained/biggest-ai-data-center-in-the-world/) shows how the other giants compare. Two things make that headline number misleading. A megawatt is a rate, not an amount, and most data centers do not use all the power they are hooked up for. Both are worth a minute. ## Megawatts on the gate, megawatt-hours on the meter A megawatt (MW) measures how fast power can be drawn. A megawatt-hour (MWh) measures how much energy is actually used: one megawatt for one hour. A year has 8,760 hours, so every 100 MW drawn around the clock is 876 GWh a year, about what 81,000 American homes buy (our arithmetic). A data center quotes several different megawatts. IT power is what the computers can draw. Facility power adds cooling and other overhead; the ratio between the two is [PUE](https://www.sotkn.com/explained/pue-explained/), about 1.2 at New Carlisle on Epoch’s figures (our arithmetic). The grid hookup is the maximum the utility has agreed to deliver. And the energy used is what the meter reads at the end of the month. Most headlines quote the first or third; the electricity system feels the fourth. Here is how those numbers look for the world, three countries in our region and one campus. | Where | Grid hookup or rating (the sign) | Energy used (the meter) | What it means | | ------------------------------------- | --------------------------------------------------------------- | ------------------------------------------------------------ | ---------------------------------------------------------------------------------------------- | | **World**, 2025 | n/a | About 485 TWh, slightly more than 1.5% of world electricity | Like about 55 GW running all year (our arithmetic); about 950 TWh by 2030 in the IEA base case | | **United States**, 2023 | n/a | 176 TWh, about 4.4% of US electricity | Like about 20 GW nonstop (our arithmetic); 6.7–12% by 2028 | | **Ireland**, 2025 | n/a | 7,663 GWh, 23% of metered electricity | An average draw of about 875 MW for the whole country’s fleet (our arithmetic) | | **Australia** (main grid), FY2026 | 2.2 GW maximum connection for operating sites | About 5 TWh, about 3% of the grid | About 27% of the hookup used in Q1 2026 | | **Malaysia**, first half 2026 | 5.65 GW in system across 42 projects | 4.0 TWh, about 6% of TNB’s units sold (our arithmetic) | 1,256 MW load in June 2026, about 22% of the hookup (our arithmetic) | | **Singapore** | 200 MW newly allocated in August 2026 | About 7% of national electricity | Capacity is rationed by allocation rounds | | **One campus:** New Carlisle, Indiana | About 1,037 MW IT, 1,244 MW total (Epoch AI estimate, Oct 2026) | Up to about 10.9 TWh a year if run flat out (our arithmetic) | More than all of Ireland’s data centers used in 2025 | *Sources by row: IEA (April 2026), US DOE and Lawrence Berkeley National Laboratory (December 2024), CSO Ireland (July 2026), AEMO (August 2026), TNB (August 2026), Singapore’s MDDI (May 2024) and EDB/IMDA (August 2026), Epoch AI (October 2026). Full list under References. Years and definitions differ by row, so compare scale, not decimals.* ## Why does a data center use less power than its grid hookup? Because it fills up slowly and keeps spare room. Australia’s grid operator AEMO gives the clearest public numbers. Its operating data centers had about 2.2 GW of maximum grid connection capacity in the first quarter of 2026, but actual demand was about 27% of that. At the same time they ran at a load factor near 90%, meaning their demand barely moved through the day. Those two numbers sound contradictory but they measure different things. The IEA explains the difference: load factor is average load divided by peak load, and data centers typically score 75–90%. Capacity utilisation is average load divided by the nameplate capacity, and that is often much lower. A data center is flat but far from full. - **It is modular.** Halls and racks are installed in stages, so demand ramps over years. AEMO says operators often expect five to 10 years to reach full capacity. - **It keeps redundancy.** Spare capacity is held back so a failure does not take the site down. AEMO expects even a mature fleet to average only about 45% to 51% of its connection. - **It over-reserves.** Grid connections are scarce and future demand is uncertain, so operators contract more than they need. The IEA warns this can oversize networks and block capacity other users could have had. ![Share of grid hookup data centers use: Malaysia 22 percent in June 2026, Australia 27 percent in Q1 2026, AEMO mature-fleet estimate 45 to 51 percent, full hookup 100 percent](https://storage.ghost.io/c/cd/10/cd10e222-131f-4ba4-b4a9-ca4214e8c12d/content/images/2026/10/data-center-hookup-used-share.png) Flat, but far from full. TNB and AEMO; Second Order arithmetic. Malaysia shows the same gap. TNB told analysts on 28 August 2026 that 42 data center projects with 5.65 GW of maximum demand were in its system by June 2026, while actual data center load was 1,256 MW. That is about 22% of the hookup (our arithmetic). Our [Malaysia Pulse on plugged-in versus used power](https://www.sotkn.com/markets/malaysia/malaysia-data-centre-power-use-october-2026/) covers that gap in detail, and our [contracted power explainer](https://www.sotkn.com/explained/contracted-power-explained/) shows why the same gap matters for revenue. ## How much electricity do data centers use worldwide? About 485 TWh in 2025, according to the IEA, slightly more than 1.5% of the world’s electricity. Use grew by more than 15% from 2024, adding about 70 TWh in a year. Spread evenly over the year, 485 TWh is like about 55 GW running nonstop (our arithmetic). The four largest hyperscalers alone, Amazon, Microsoft, Alphabet and Meta, used about 110 TWh in 2024, up 20%. The IEA’s base case has data center use roughly doubling to about 950 TWh by 2030, around 3% of world electricity. The faster part is AI: consumption by AI-focused data centers more than triples to about 465 TWh by 2030, driven by shipments of AI chips that rise 2.2 times from 2025\. The IEA expects around 20 GW of AI servers to be added each year by 2030\. See [AI training vs inference](https://www.sotkn.com/explained/ai-training-vs-inference/) for what those chips do. In the United States, a Lawrence Berkeley National Laboratory report released by the Department of Energy in December 2024 put data center use at 176 TWh in 2023, about 4.4% of US electricity, up from 58 TWh in 2014\. It projected 325 to 580 TWh by 2028, or 6.7% to 12% of the country’s power. The global share hides the local one. The IEA says that while data centers used about 1.5% of world electricity in 2025, they accounted for 20–30% of demand in the states, metro areas or regions where they cluster. Ireland is the national extreme: its Central Statistics Office reported on 7 July 2026 that data centers used 7,663 GWh in 2025, up 10% in a year and 23% of all metered electricity, from 5% in 2015\. That is an average draw of about 875 MW for the whole country’s fleet (our arithmetic), less than one New Carlisle campus at full tilt. ## How much power do data centers use in Singapore, Malaysia and Australia? **Singapore.** Singapore’s government said in May 2024 that data centers use about 7% of the country’s electricity and account for 82% of the information and communications sector’s emissions. New capacity is rationed: in August 2026 its Economic Development Board and IMDA provisionally allocated 200 MW to four operators, 50 MW each. If all 200 MW ran flat out for a year it would be about 1.75 TWh (our arithmetic), a sixth of New Carlisle’s full-tilt figure. See our [Singapore Pulse on its power plan for data centres](https://www.sotkn.com/markets/singapore/singapore-power-plan-data-centres-green-fuel-october-2026/) and [what Keppel DC REIT does](https://www.sotkn.com/stack/ai-data-centers/what-does-keppel-dc-reit-do/). **Malaysia.** Malaysia is where the region’s growth shows up on the meter. TNB reported data center energy use of 4.0 TWh in the first half of 2026, 2.1 times the 1.9 TWh a year earlier, and load rising from 603 MW in June 2025 to 1,256 MW in June 2026\. That is about 6% of the 70,328 GWh TNB sold in the half (our arithmetic). Behind it sit 61 secured projects with 8.35 GW of maximum demand, 67% of it in Johor. See [where AI data centers are located](https://www.sotkn.com/explained/where-are-ai-data-centers-located/). **Australia.** AEMO counts around 165 data centers on the main eastern grid using about 5 TWh in the year to June 2026, about 3% of the grid’s consumption, with average monthly demand up from nearly 200 MW in early 2018 to more than 700 MW by mid-2026\. New South Wales takes 66% and Victoria 30%. In AEMO’s central case, use triples to about 15 TWh by 2030 and reaches about 34 TWh, or 13% of the grid, by 2036\. See [what NEXTDC does](https://www.sotkn.com/stack/ai-data-centers/what-does-nextdc-do/) and our [Australia Pulse on Sydney sites waiting for power](https://www.sotkn.com/markets/australia/australia-nsw-data-centre-power-condition-october-2026/). ## Who gets paid: the hookup or the meter? This is where the gap between the two numbers turns into money. Different parts of the chain are paid on different meters. - **Utilities and wire builders, increasingly on the hookup.** Network costs are recovered through both per-kWh and per-kW charges, the IEA notes, and regulators are shifting more onto large loads. Virginia has approved a new rate class for loads over 25 MW with load factors over 75%, taking effect from January 2027, with minimum demand charges, exit fees and long-term commitments. Hydro-Québec has proposed a rate for data centers above 5 MW about double what large consumers pay now. Our [Wire on Duke’s 75% minimum bill](https://www.sotkn.com/wire/duke-data-center-power-tariff-75-percent-minimum-bill-wire/) and [Malaysia’s charge on idle booked power](https://www.sotkn.com/markets/malaysia/malaysia-data-centre-unused-power-charge-october-2026/) show the same idea: pay for the hookup, used or not. - **Generators and fuel sellers, on the meter.** They earn on megawatt-hours, so they are paid as the halls fill. The IEA says global gas turbine orders in 2025 were the second-highest on record and the highest in 25 years, and that data centers were an important contributor in the United States. See [how hyperscalers buy power](https://www.sotkn.com/stack/hyperscalers/how-do-hyperscalers-buy-power/). - **Landlords, on the kilowatts reserved.** Colocation rent is usually set on the IT capacity a tenant reserves, and it starts when that power is live. See [how a data center lease works](https://www.sotkn.com/explained/data-center-lease-explained/) and [hyperscale vs colocation](https://www.sotkn.com/explained/hyperscale-vs-colocation-data-center/). - **Everyone else on the grid, sometimes.** If networks are built for hookups that never fill, the IEA warns, the cost of oversized wires and reserves can spread to other customers. It also found no clear, systematic link between demand growth and retail price changes across the markets it compared for 2019 to 2024, so the effect depends on the grid. Our [Analysis on utilities billing before power](https://www.sotkn.com/analysis/utilities-bill-data-centers-before-power-where-cash-sticks/) follows that cash. - **Last: the data center owner.** It pays for the hookup from day one and earns only on what it fills. At 27% utilisation, most of the connection is a cost waiting to become revenue. ## Where is the stuck step? The hookup itself. Getting a large grid connection is slow and scarce, which is exactly why operators over-reserve. AEMO’s 2026 assessment shows the pile-up: proposed data center connection capacity on its grid rose from 38 GW to 67 GW in a year, across 225 projects, about 30 times the 2.2 GW operating sites have today (our arithmetic). At the same time, about 36% of the projects on its 2025 list have since been cancelled, including more than 30% of those classed as committed. ![Four numbers for one data center's power: grid hookup (the stuck step, utilities paid on it), facility power (New Carlisle about 1,244 MW), IT power (about 1,037 MW, landlords price on it), energy used (Ireland 7,663 GWh in 2025, generators paid on it)](https://storage.ghost.io/c/cd/10/cd10e222-131f-4ba4-b4a9-ca4214e8c12d/content/images/2026/10/data-center-four-power-numbers.png) Which power number is it? AEMO, TNB, Epoch AI, CSO Ireland; Second Order. So two clocks run at once. One is the queue for the hookup, measured in years. The other is the ramp inside the hookup, which AEMO puts at five to 10 years to maturity. Our [page on how long an AI data center takes to build](https://www.sotkn.com/explained/how-long-does-it-take-to-build-an-ai-data-center/) explains why the building is the quick part, and our [data center cost page](https://www.sotkn.com/explained/how-much-does-it-cost-to-build-an-ai-data-center/) shows what sits on the balance sheet while it waits. Our view: for an investor, the most useful power number is not the headline megawatts. It is the gap between the hookup and the meter, and how fast it is closing. A site at 27% that is filling fast is a different business from one stuck at 27%. ## What would change these numbers? - **Faster AI fill.** If AI demand fills halls faster than AEMO’s five-to-10-year ramp, utilisation rises and the meter catches up with the hookup. - **Use-it-or-lose-it rules.** The IEA lists stronger financial requirements and use-it-or-lose-it measures for reserved grid capacity. Where they bite, over-reserving gets expensive and hookups shrink toward real use. - **More efficient chips and cooling.** The IEA notes efficiency can slow, but not offset, growth. A lower PUE lowers facility power for the same IT load. - **Flexible loads.** Shifting non-urgent work in time or across sites, or using on-site storage, could let data centers connect with less firm capacity. - **Cancellations.** AEMO’s 36% cancellation rate is a reminder that the hookup queue is not a forecast. Watch which projects actually energize. ## How to read a data center power number - **MW or MWh?** A rate of draw and an amount of energy are different things. Multiply MW by 8,760 for a year at full draw. - **IT, facility or hookup?** Facility power is IT power times PUE. On AEMO’s numbers, a hookup can be about twice to nearly four times what the sites draw (our arithmetic, from 45–51% and 27% utilisation). - **Used or planned?** A pipeline of gigawatts is a queue. A meter reading is a fact. - **What share, and of what?** 1.5% of the world, 4.4% of the US and 23% of Ireland are all true at once. - **Who is paid on it?** The hookup pays the utility early; the meter pays the generator later; the owner earns last. ## FAQ **How much power does an AI data center use?** A large AI data center is built to draw from about 100 MW to more than 1 GW. The biggest running campus, Amazon’s New Carlisle site in Indiana, is estimated by Epoch AI at about 1,037 MW for its computers and 1,244 MW for the whole site in October 2026\. Run flat out for a year, 1,244 MW would use about 10.9 TWh, about what a million American homes buy (our arithmetic). Most data centers use much less than their grid hookup in practice. **How much electricity does a data center use per year?** Multiply the power it actually draws by 8,760 hours. Every 100 MW drawn around the clock for a year is 876 GWh, about what 81,000 American homes buy in a year at the US Energy Information Administration’s average of 10,791 kWh per home (our arithmetic). A data center that averages half its hookup uses half that. **How much electricity do data centers use worldwide?** About 485 TWh in 2025, according to the International Energy Agency, slightly more than 1.5% of the world’s electricity. That is like about 55 GW running nonstop all year (our arithmetic). The IEA’s base case has it roughly doubling to about 950 TWh by 2030, around 3% of world demand, with AI-focused data centers more than tripling to about 465 TWh. **What share of US electricity do data centers use?** About 4.4% in 2023, or 176 TWh, according to a Lawrence Berkeley National Laboratory report released by the US Department of Energy in December 2024\. The same report projects 6.7% to 12% of US electricity by 2028, or 325 to 580 TWh. **Why does a data center use less power than its grid connection?** Because it fills up slowly and keeps spare capacity. Data centers install computers in stages, hold redundant capacity in reserve, and often reserve more grid capacity than they need because connections are scarce. Australia’s grid operator AEMO found operating data centers used about 27% of their maximum connection capacity in early 2026, while running at a steady load factor near 90%. It expects a mature fleet to average about 45% to 51%. **How many homes can a 1 GW data center power?** About 810,000 American homes if it ran flat out all year, and fewer at lower use (our arithmetic: 1 GW for 8,760 hours is 8.76 TWh, at 10,791 kWh per home per year). The IEA puts it the same way: the 21 large AI data centers it studied in the US, at 600 MW to 5 GW each, would each use as much electricity as 600,000 to 5 million households once complete. **How much electricity do data centers use in Singapore, Malaysia and Australia?** Singapore’s government says data centers use about 7% of the country’s electricity. In Malaysia, utility TNB reported 4.0 TWh of data center use in the first half of 2026, up 2.1 times in a year, and a load of 1,256 MW in June 2026\. In Australia’s main grid, AEMO counts around 165 data centers using about 5 TWh in the year to June 2026, about 3% of the total, heading to about 34 TWh, or 13%, by 2036 in its central case. **Do AI data centers use more power than regular data centers?** Per site and per rack, yes. The IEA says AI server power density rose 11 times between 2020 and 2025, and that a rack the size of a large refrigerator could by 2027 draw as much at peak as about 65 households. AI-focused sites are also much larger: the US AI campuses the IEA tracked range from 600 MW to 5 GW, against tens of megawatts for a typical older data center. **Is this a list of stocks to buy or sell?** No. Second Order Explained pages explain how the money and the constraints move. They do not give buy or sell calls. Not investment advice. Do your own research. Related: [what is an AI data center](https://www.sotkn.com/explained/what-is-an-ai-data-center/); [AI data center water use](https://www.sotkn.com/explained/ai-data-center-water-usage/); [GPU utilization](https://www.sotkn.com/explained/gpu-utilization-explained/); [who builds AI data centers](https://www.sotkn.com/explained/who-builds-ai-data-centers/); [data center financing](https://www.sotkn.com/explained/data-center-financing-explained/); [data center tiers](https://www.sotkn.com/explained/data-center-tiers-explained/). ## Stack doors Layer: [AI Data Centers](https://www.sotkn.com/stack/ai-data-centers/). Also: [Hyperscalers](https://www.sotkn.com/stack/hyperscalers/), [Neoclouds](https://www.sotkn.com/stack/neoclouds/), [Semiconductors](https://www.sotkn.com/stack/semiconductors/). Markets: [US](https://www.sotkn.com/markets/us/), [Malaysia](https://www.sotkn.com/markets/malaysia/), [Singapore](https://www.sotkn.com/markets/singapore/), [Australia](https://www.sotkn.com/markets/australia/). [Get the Wire](https://www.sotkn.com/#/portal/signup) — short daily notes when a data center gets its hookup, its megawatts or its money. **How we check this** Figures come from the IEA’s Key Questions on Energy and AI (April 2026), the US Department of Energy’s release of Lawrence Berkeley National Laboratory’s data center energy report (December 2024), Ireland’s Central Statistics Office (July 2026), AEMO’s 2026 Electricity Statement of Opportunities and its data center forecasting overview (August 2026), TNB’s 2QFY2026 analyst briefing (August 2026), Singapore’s Ministry of Digital Development and Information (May 2024) and EDB/IMDA fact sheet (August 2026), Epoch AI’s New Carlisle directory page and timelines dataset, and the US Energy Information Administration’s household electricity FAQ, all read on 31 October 2026 and listed under References. Epoch AI’s figures are estimates from satellite images and permits, not company numbers. TNB’s first half is January to June 2026\. Countries count data centers differently, so the table compares scale, not exact shares. Every conversion we derived, such as TWh to average gigawatts, homes equivalents and shares of a hookup, is labelled “our arithmetic”; one year is taken as 8,760 hours and one US home as 10,791 kWh a year. The featured photo is a 750 kV transformer and substation from Wikimedia Commons (CC0), graded by Second Order; it is not a site named here. This page explains how the money and the constraints move. It does not give buy or sell calls. Not investment advice. Do your own research. **Last reviewed:** 31 Oct 2026 · **Next review:** after TNB’s third-quarter briefing and the IEA’s next data centre electricity update Spot an error? [Tell us](mailto:sean@thekopinotes.com) and we will correct it and note the change here. [Our editorial standards →](https://www.sotkn.com/editorial-standards/) ## References 1. [IEA — Key Questions on Energy and AI (report, 16 Apr 2026, CC BY 4.0)](https://www.iea.org/reports/key-questions-on-energy-and-ai?ref=sotkn.com) — data centres about 485 TWh in 2025, slightly more than 1.5% of world electricity, about 950 TWh by 2030; AI-focused data centres triple to about 465 TWh; 20–30% of demand in cluster regions; load factor 75–90% vs lower capacity utilisation and over-contracting; 21 large US AI sites of 600 MW–5 GW = 600,000–5 million households; Virginia large-load rate class from Jan 2027; Hydro-Québec data centre rate; gas turbine orders 2. [US Department of Energy — DOE releases new report evaluating increase in electricity demand from data centers (Lawrence Berkeley National Laboratory report, 20 Dec 2024)](https://www.energy.gov/articles/doe-releases-new-report-evaluating-increase-electricity-demand-data-centers?ref=sotkn.com) — US data centres 176 TWh in 2023, about 4.4% of US electricity; 325–580 TWh and 6.7–12% by 2028 3. [Central Statistics Office Ireland — Data Centres Metered Electricity Consumption 2025 (7 Jul 2026)](https://www.cso.ie/en/releasesandpublications/ep/p-dcmec/datacentresmeteredelectricityconsumption2025/keyfindings/?ref=sotkn.com) — 7,663 GWh in 2025, up 10% from 6,973 GWh; 23% of metered electricity, from 5% in 2015 4. [AEMO — 2026 Electricity Statement of Opportunities (Aug 2026)](https://www.aemo.com.au/-/media/files/electricity/nem/planning%5Fand%5Fforecasting/nem%5Fesoo/2026/2026-electricity-statement-of-opportunities.pdf?ref=sotkn.com) — 2.2 GW of maximum connection capacity for operating data centres; about 27% aggregate utilisation in Q1 2026 at a load factor near 90%; mature fleet 45–51%; 225 projects, 38 GW to 67 GW proposed; about 36% of 2025 projects cancelled 5. [AEMO — 2026 ESOO data centre forecasting overview (Aug 2026)](https://www.aemo.com.au/-/media/files/electricity/nem/planning%5Fand%5Fforecasting/nem%5Fesoo/2026/2026-esoo-data-centre-forecasting-overview.pdf?ref=sotkn.com) — around 165 data centres used about 5 TWh in FY2026, about 3% of NEM consumption; monthly demand from nearly 200 MW (2018) to more than 700 MW; NSW 66%, Victoria 30%; about 15 TWh by FY30 and 34 TWh (13%) by FY36 in Step Change; load maturity 5–10 years 6. [Tenaga Nasional Berhad — Analyst briefing 2QFY2026 (28 Aug 2026)](https://www.tnb.com.my/assets/quarterly%5Fresults/Analyst%5FBriefing%5F-%5F2QFY2026.pdf?ref=sotkn.com) — data centre energy use 4.0 TWh in 1HFY2026 vs 1.9 TWh (2.1x); load 603 MW (Jun 2025) to 1,256 MW (Jun 2026); 42 projects with 5.65 GW in system; 61 secured projects, 8.35 GW; 70,328 GWh total units sold 7. [Ministry of Digital Development and Information, Singapore — Opening remarks at the ATxSummit Digital Sustainability Forum (30 May 2024)](https://www.mddi.gov.sg/newsroom/opening-by-sms-janil-puthucheary-at-atxsummit-digital-sustainability-forum/?ref=sotkn.com) — data centres are 7% of Singapore’s total electricity consumption and 82% of the information and communications sector’s emissions 8. [EDB and IMDA — Four data centre proposals selected in DC-CFA2 (fact sheet, 21 Aug 2026)](https://www.edb.gov.sg/content/dam/edb-en/about-edb/media-releases/news/four-data-centre-proposals-selected-as-part-of-second-data-centre/Fact%20Sheet%20-Four%20data%20centre%20proposals%20selected%20as%20part%20of%20second%20data%20centre.pdf?ref=sotkn.com) — 200 MW provisionally allocated, 50 MW each to four operators 9. [Epoch AI — Anthropic-Amazon New Carlisle, AI data centers directory and timelines dataset (read 31 Oct 2026, CC BY 4.0)](https://epoch.ai/data/ai-data-centers/directory/anthropic-amazon-new-carlisle?ref=sotkn.com) — IT power 398 MW (Jun 2025), 626 MW (Dec 2025), 910 MW (Mar 2026), 1,037 MW (19 Oct 2026); total facility power 1,244 MW; projected 1,925 MW IT / 2,310 MW total by Mar 2028 10. [US Energy Information Administration — How much electricity does an American home use? (FAQ)](https://www.eia.gov/tools/faqs/faq.php?id=97&t=3&ref=sotkn.com) — 10,791 kWh per US residential customer in 2022, about 899 kWh a month S **[Sean](https://www.sotkn.com/author/sean/)** Writes Second Order by TKN — a plain-English map of AI infrastructure and semiconductors for non-specialist investors. Focus: who gets paid, where money sticks, and which step is stuck. Covers AI infrastructure, semiconductors, and second-order stack reads. [Author page →](https://www.sotkn.com/author/sean/) · [Editorial standards →](https://www.sotkn.com/editorial-standards/)