Compute and Electricity: The Defining Challenge of Our Times

The collision between artificial intelligence and the physical electricity system has become the central bottleneck of the modern economy. Compute is no longer limited mainly by silicon or algorithms. It is limited by how much firm power you can get, where you can get it, and how fast it can be connected. This is not a passing shortage. Energy has reasserted itself as the binding resource for intelligence, and that changes which companies, which grids and which countries can keep scaling.

Power equals progress. Countries that are building energy production will be the leaders of the future.

That is my position and I want to be blunt about what it implies. A nation can write any AI strategy it likes. If it cannot deliver firm kilowatts on a two year timeline, the strategy is a press release. Capital, chips and talent are all mobile. Grid capacity is not.

Global data centre electricity consumption rising from 415 TWh in 2024 to a projected 945 TWh in 2030, with data centre demand growing 17 percent in 2025 against 3 percent for total world demand

The Scale Of The Demand Shock

Start with the numbers rather than the rhetoric. The International Energy Agency puts global data centre consumption at around 415 TWh in 2024, about 1.5% of world electricity, rising to roughly 945 TWh by 2030 in its base case. That increment is close to Japan's entire annual electricity use, added to the world's load in six years. The IEA also notes data centre consumption has compounded at around 12% a year since 2017, more than four times the growth rate of total electricity demand (IEA, Energy and AI).

The 2026 update is sharper still. In its April 2026 report, the IEA said data centre electricity demand rose 17% in 2025 while total global electricity demand grew about 3%, and that AI focused facilities grew faster than the sector as a whole. Capital expenditure by the five largest technology companies passed $400bn in 2025 and is set to rise a further 75% in 2026 (IEA, Key Questions on Energy and AI).

Two features of that demand make it harder to serve than the headline TWh suggest.

It is concentrated. The United States accounted for about 45% of global data centre power use in 2024, China about 25% and Europe about 15%. Within those markets it clusters into a handful of grids.
It is firm. A training run needs continuous high utilisation for days or weeks. Inference at global scale is persistent base load. Rack density has moved from 5 to 15 kW a few years ago to 40 to 100 kW and beyond for AI configurations, which drags cooling and water into the same constraint.

The AI share of data centre power sat at roughly 5% to 15% in recent years. The IEA sees it reaching 35% to 50% by 2030. The load is not only growing, it is becoming less flexible as it grows.

The Grid Is Already Repricing

The clearest evidence that this is a physical constraint rather than a forecast is what has happened to capacity prices in the largest competitive market in the United States.

PJM Interconnection's latest capacity auction cleared at the $333.44/MW-day cap, the third consecutive auction to set a record and the third to hit a ceiling (Utility Dive). PJM's independent market monitor attributed $6.5bn, around 40%, of the $16.4bn total to data centre load, with roughly $6.2bn of that tied to data centres that have not been built yet. Governors across the PJM footprint negotiated a price ceiling for the next two auctions rather than let the market clear freely.

The equipment market tells the same story from the supply side. GE Vernova ended the second quarter of 2026 with a 116 GW gas turbine backlog, up from 100 GW a quarter earlier, and expects at least 125 GW under contract by year end, with reservations now being taken for 2031 deliveries (GE Vernova 8-K, SEC EDGAR). Siemens Energy is carrying the largest backlog in its history at about €136bn. Mitsubishi Power is sold out into 2028 even while doubling production.

Read those two facts together. Buyers are paying record prices for capacity that does not yet exist, and the three firms who build the fastest form of new firm generation have no delivery slots left this decade. That is the definition of a hard constraint.

Nuclear Came Back Because The Physics Fits

Nuclear power is having its moment for an unsentimental reason. High capacity factor, low carbon, always available generation matches the load shape of an AI data centre better than any other mature technology. The hyperscalers did not rediscover nuclear on principle. They ran out of alternatives that clear on their timeline.

Microsoft's 20 year agreement with Constellation Energy to restart Three Mile Island Unit 1, now the Crane Clean Energy Center, covers roughly 835 MW with first power targeted for 2027. Amazon contracted about 1,920 MW from Talen Energy's Susquehanna plant. Google signed the first corporate agreement for small modular reactor output with Kairos Power, up to 500 MW. Meta has assembled the largest disclosed portfolio at up to 6.6 GW across TerraPower, Oklo, Vistra and Constellation. Across more than a dozen agreements, hyperscalers have committed to around 9.8 GW of nuclear capacity.

The forward pipeline is moving faster than the delivered capacity. The IEA reports that conditional small modular reactor offtake agreements grew from 25 GW at the end of 2024 to about 45 GW by early 2026. Oklo broke ground on its Aurora powerhouse at Idaho National Laboratory in September 2025, received Nuclear Regulatory Commission approval of its Principal Design Criteria topical report in May 2026 in under half the traditional review time (Oklo), and secured Department of Energy safety analysis approval in June 2026.

China is doing the same thing at state scale. Its 15th Five-Year Plan, approved in March 2026, targets 110 GWe of nuclear capacity by 2030 against roughly 56 GW operating today, with more than 30 units under construction and ten or more reactors approved every year since 2022 (CSIS, World Nuclear Association). The institutional method differs from the American one. The physical logic is identical.

The Electricity League Table Today

If power is the constraint, then national generation capacity is the scoreboard. Here is where the world actually stands, using Ember's 2025 actuals.

Rank	Country	2025 generation (TWh)	Share of world	What it runs on
1	China	10,580	33.4%	Coal, record solar, fast growing nuclear
2	United States	4,520	14.3%	Gas, nuclear, wind and solar
3	India	2,082	6.6%	Coal plus the fastest solar build in the world
4	Russia	1,193	3.8%	Gas, hydro, nuclear
5	Japan	1,030	3.2%	Gas, coal, restarting nuclear
6	Brazil	751	2.4%	Hydro plus wind
7	Canada	652	2.1%	Hydro plus CANDU nuclear
8	South Korea	625	2.0%	Nuclear, coal, gas
9	France	570	1.8%	Nuclear
10	Germany	500	1.6%	Wind, solar, gas

Source: Ember, Global Electricity Review 2026. The top ten account for roughly 71% of world generation of about 31,700 TWh. China alone generates more than the next three combined.

Ember's 2026 review carried a second finding that matters for this thesis. Renewables reached nearly 34% of global generation and overtook coal at 33% for the first time in a century, with solar rising a record 636 TWh to 2,778 TWh and meeting 75% of the net increase in demand. Solar is winning the growth. It is not yet winning the 24/7 profile that AI needs, which is why gas and nuclear are being contracted at the same time.

Top ten electricity producing countries in 2025 with projected 2036 output, showing Saudi Arabia and Indonesia entering the table while France and Germany drop out

The Same Table In Ten Years

Now the more useful question. Who is building, and where does that leave the league table in 2036?

The IEA's World Energy Outlook 2025 has global electricity demand rising about 40% by 2035, roughly 1,000 TWh added every year (IEA WEO 2025). I have allocated that growth country by country against stated national plans, then checked the total for consistency. The result below is a projection with its assumptions on the table, not a forecast dressed up as certainty.

Rank 2036	Country	Base (TWh)	Assumed growth	2036 (TWh)	What drives it
1	China	10,580	3.8% a year	16,000	15th Five-Year Plan, 110 GW nuclear by 2030, record solar and storage
2	United States	4,520	2.5% a year	5,900	Data centre load, gas build, nuclear restarts and uprates
3	India	2,082	5.9% a year	3,900	Fastest demand growth of any large economy, solar plus coal
4	Russia	1,193	1.1% a year	1,350	Slow domestic growth, nuclear exports
5	Japan	1,030	0.8% a year	1,120	Restarts plus new domestic AI capacity
6	Brazil	751	3.2% a year	1,060	Hydro, wind, industrial onshoring
7	Canada	652	2.0% a year	810	Hydro, CANDU refurbishment, data centre siting
8	Saudi Arabia	455 (2024)	5.1% a year	785	Humain AI programme, gas and solar, first nuclear
9	South Korea	625	1.9% a year	770	Nuclear expansion policy, chip fabrication load
10	Indonesia	372 (2024)	5.9% a year	700	Industrialisation, minerals processing, urbanisation
11	France	570	1.7% a year	690	Falls out of the top ten despite the EPR2 programme
12	Vietnam	308 (2024)	7.0% a year	660	Power Development Plan 8, manufacturing relocation, export ambitions
13	Germany	500	1.2% a year	570	Falls out of the top ten on flat demand

Base year figures are Ember 2025 actuals except Saudi Arabia and Indonesia, which are 2024. Projection is Live Trading News analysis.

Three things fall out of that table.

Europe leaves the top ten. France and Germany are displaced by Saudi Arabia and Indonesia. Flat demand used to read as efficiency. In an AI economy it reads as a ceiling on how much compute you can host.
The gap widens rather than closes. China moves from 33% of world generation to roughly 36%. India roughly doubles. Everyone else is fighting for the remainder.
The new entrants are the ones with a plan and a balance sheet. Saudi Arabia is targeting 1.9 GW of AI capacity by 2030 and 6.6 GW by 2034 alongside its first nuclear programme. The UAE has Barakah operating and the first 200 MW of Stargate UAE going live this year inside a campus scoped at 5 GW (G42).

Who Is Building For The Future

Four routes to firm power for AI: restarts and uprates, advanced reactors, the gas bridge, and frontier options including geothermal and orbital compute

There are only four routes to a guaranteed kilowatt for an AI campus, and they resolve on very different timelines.

Restarts and uprates deliver 2027. Three Mile Island Unit 1 and Susquehanna are the template. An existing licensed site, an investment grade counterparty on a twenty year contract, and power inside three years. This is the only route that solves the near term.

Advanced reactors deliver the 2030s. Meta's 6.6 GW portfolio, Google's Kairos agreement and Oklo's Aurora build are options on factory fabricated capacity that mostly lands between 2029 and 2032. Valuable, but not a 2027 answer.

Gas is the bridge and it is fully booked. 116 GW of backlog at GE Vernova, a record book at Siemens Energy, Mitsubishi sold out into 2028. Anyone who did not reserve a turbine slot by 2025 is queuing behind someone who did.

Frontier options are optionality, not supply. Enhanced geothermal is the most credible of them. Fervo Energy held 658 MW of contracted PPAs at the end of 2025 and a 3 GW framework agreement with Google, and listed on Nasdaq in May 2026 raising about $2.2bn (Fervo S-1, SEC EDGAR). Space is the most radical. SpaceX unveiled its AI1 orbital data centre satellite in June 2026 at 120 to 150 kW of compute payload each, having filed with the FCC in January 2026 for a constellation of up to one million units, with launches targeted from 2027 (DCD). Sunlight is nearly continuous in orbit and land and water constraints disappear. Thermal rejection, radiation hardening and debris management do not.

Elon Musk has put the endpoint of all this more directly than most energy ministers would dare. Writing on X in February 2026, he argued that once the loop from solar generation to robot manufacturing to chip fabrication to AI closes, "conventional currency will just get in the way. Just wattage and tonnage will matter, not dollars" (X). Whether or not you accept the monetary conclusion, the direction of travel is measurable in every capacity auction and every turbine order book.

The Equity Map: Where The Winners Sit

The AI energy trade is wider than a handful of nuclear tickers. It runs from uranium in the ground to the switchgear in the hall. Below is how I map the chain. Established operators with contracted offtake earn today. Pure play developers are options on execution. Those are not the same instrument and should not be sized the same way.

Contracted firm power, earning now

Ticker	Company	Position in the chain	Why it can win	What to watch
CEG	Constellation Energy	Largest US commercial nuclear operator	20 year Microsoft PPA on the 835 MW Crane restart, contracted cash flow visibility that generation peers lack	Restart on schedule for 2027, new offtake signings, free cash flow into the late 2020s
VST	Vistra	Nuclear plus a large Texas and PJM gas fleet	Captures both scarcity pricing today and clean power contracting tomorrow	Uprates, new data centre contracts, ERCOT and PJM capacity outcomes
TLN	Talen Energy	Susquehanna nuclear, restructured and focused	Roughly 1,920 MW contracted to Amazon from a single high quality site	Contract extensions, co-location approvals
1816.HK	CGN Power	China's largest nuclear operator	28 units and about 31.8 GW managed at end 2025 inside a 110 GW national target	Unit availability, new approvals, tariff policy
601985.SS	China National Nuclear Power	CNNC listed operator	26 units and about 25 GW at end 2025, direct exposure to the state build programme	Construction starts, Linglong One commercial operation

Developers and the fuel cycle, options on execution

Ticker	Company	Position in the chain	Why it can win	What to watch
OKLO	Oklo	Sodium cooled micro reactors for campus scale load	First Aurora under construction at Idaho, regulatory reviews closing faster than precedent	First criticality timeline, fuel supply, conversion of interest into firm orders
SMR	NuScale Power	Light water small modular reactor	The US SMR with regulatory design approval in hand	Orders that convert into steel in the ground
BWXT	BWX Technologies	Nuclear components, fuel and naval reactors	Sells into commercial, defence and SMR programmes at once	SMR component awards, DOE and navy work
CCJ	Cameco	Uranium mining and fuel services	2026 guidance of $3.13bn to $3.37bn at a realised $85 to $89 per pound, plus a nine year, 22 million pound agreement with India	Contracting cycle, spot uranium, Cigar Lake and McArthur River output
LEU	Centrus Energy	Enrichment, including HALEU	$900m Department of Energy task order to expand HALEU at Piketon	HALEU delivery milestones, additional DOE awards

Equipment, grid and frontier

Ticker	Company	Position in the chain	Why it can win	What to watch
GEV	GE Vernova	Gas turbines, grid equipment, BWRX-300 SMR	116 GW gas backlog at Q2 2026 and effectively sold out toward 2030	2031 slot contracting, electrification margin, SMR progress
ENR.DE	Siemens Energy	Turbines, grid technologies	Record backlog of about €136bn, grid business into a transmission supercycle	Grid technologies margin, execution on the book
7011.T	Mitsubishi Heavy Industries	Gas turbines, nuclear services	Sold out into 2028 while expanding capacity	Capacity additions, Japanese restart work
VRT	Vertiv	Data centre power and thermal management	Rack density is rising faster than the buildings, which is a liquid cooling problem	Orders and backlog, AI rack mix
ETN	Eaton	Electrical equipment and switchgear	Every megawatt reaching a rack passes through equipment like this	Data centre order growth, lead times
PWR	Quanta Services	Transmission and interconnection construction	Generation without transmission is stranded, and the queue is the constraint	Backlog, utility capital plans
FRVO	Fervo Energy	Enhanced geothermal	658 MW of contracted PPAs and a 3 GW framework with Google	Cape Station delivery, drilling cost curve
SPCX	SpaceX	Launch plus orbital compute	AI1 satellites at 120 to 150 kW each, launch cost advantage owned in house	FCC action on the constellation filing, first AI1 launches

This is analysis and opinion, not investment advice. Position sizes are yours.

What Could Go Wrong With This Thesis

Honest bull cases carry their own bear case.

Efficiency could outrun demand. If model and hardware efficiency improves faster than currently modelled, power intensity falls and some of this contracted capacity becomes surplus. Scenario analysis needs a high efficiency case, not only a high demand case.
Speculative capacity may not arrive. PJM's monitor found roughly $6.2bn of capacity cost tied to data centres that have not been built. If a share of the announced pipeline is duplicative or cancelled, capacity prices normalise and the merchant power trade unwinds fast.
Advanced nuclear is still pre revenue. Regulatory delay, first of a kind cost overruns and fuel cycle bottlenecks are the historical pattern in this industry, not the exception.
Public acceptance and water. Cooling water and local opposition are the two constraints that do not appear in any capex forecast until they stop a project.

AI And AEO: The Quick Answers

How much electricity do data centres use?

About 415 TWh globally in 2024, roughly 1.5% of world electricity, rising 17% in 2025 and projected by the IEA to reach around 945 TWh by 2030.

Which country produces the most electricity?

China, at about 10,580 TWh in 2025, roughly a third of world generation and more than the next three countries combined.

Why are AI companies buying nuclear power?

AI training and inference need continuous high density power. Nuclear delivers high capacity factor, low carbon base load, which matches that load shape better than intermittent sources on the timelines hyperscalers are working to.

How much nuclear capacity have the hyperscalers contracted?

Around 9.8 GW across more than a dozen disclosed agreements, with Meta the largest single procurer at up to 6.6 GW.

What is the real bottleneck in 2026?

Delivery, not intent. Gas turbine order books are full to 2030, PJM capacity auctions are clearing at the cap, and interconnection queues run for years. The constraint is firm capacity that can be energised, not announcements.

Kilowatts Are The New Constraint

Compute and electricity have fused into one problem. The companies and countries that solve for abundant, firm, low carbon power will set the pace at which artificial intelligence scales, and everyone else will buy capacity from them at whatever the clearing price turns out to be.

Nuclear has returned to the centre of energy strategy because it matches the physics of the demand. Gas is the bridge and the bridge is fully booked. Geothermal, fusion and orbital compute widen the solution set later this decade and into the next, without changing what happens in 2027.

The race is no longer only about model parameters or process nodes. It is about kilowatts delivered reliably, at scale, under geopolitical and technological uncertainty. Power equals progress, and the league table in ten years will be written by whoever starts building now.

For the mapped view of how compute, chips, capital and energy connect across the AI supply chain, the KXCO ontology is public at kxco.ai/ontology-live.

Shayne Heffernan, Ph.D. is the founder of Live Trading News and KXCO.