Chapter 1 · Layer 0 · Power Supply
The Megawatts Nobody Can Plug In
Texas has 474 gigawatts of power requests in its queue, mostly data centres, and less than half are close to real. Why the grid connection, not the chip, is AI's first ceiling.
Sanket Kulkarni · · 7 min read
Part of The Seven-Layer Tokenomics Stack
474 gigawatts in a queue
In June 2026, the Electric Reliability Council of Texas (ERCOT), which runs most of the Texas power grid, said it was tracking about 474 gigawatts (GW) of requests from large electricity users wanting to connect to its network. About 90% of them were data centres. Three months earlier, in March, the figure was about 410 gigawatts. (A gigawatt is a million kilowatts. One gigawatt running for an hour is a million units on an Indian electricity bill.)
For scale: India's entire installed generation capacity was 548.86 gigawatts as of June 2026, according to the Ministry of Power. One American state's waiting list is close to the size of everything India can generate.
Here is the catch. When the Texas grid operator checked which of those requests had studies far enough along to enter "Batch Zero", the first round of its new connection study process, its preliminary count (as of 28 July 2026) was about 205 gigawatts. Less than half. The rest is, for now, paper megawatts: announced, applied for, not yet real.
That gap between asking for power and getting it is what this chapter is about.
The queue, in gigawatts (GW)
One state's waiting list, next to all of India
Sources: Electric Reliability Council of Texas (ERCOT), April and July 2026 (Batch Zero figure is preliminary, as of 28 July 2026); Ministry of Power, India, June 2026.
Sanctioned load: the oldest rule in electricity
Every Indian household knows the idea of sanctioned load. Your electricity connection is approved for a certain number of kilowatts. Add three air conditioners to a connection sized for two, and you will learn the limit the hard way. It does not matter how good the new air conditioners are.
A data centre works the same way, at a scale of hundreds of megawatts. In the Tokenomics Stack, Layer 0 begins "where power and water arrive at the facility." Its first component, Power Supply, covers three things:
- Grid connection and its capacity. The line from the grid, and how much power the electricity company has agreed to deliver. This is the sanctioned load.
- Delivered power by site and hour. Not power in general, but power at this address, at this hour. A megawatt at 3 pm in a Texas summer is a different product from a megawatt at 3 am.
- Backup power. Generators and batteries that keep the site running when the grid does not.
Back in our kitchen, this is the gas line. You can buy the finest tandoor in the city. If the pipe cannot deliver enough gas at dinner time, the tandoor sits cold, and you still pay for it.
Illustrative site, in megawatts (MW)
A site runs at the rate of its tightest constraint
Chips bought
Kitchen: the finest tandoor
Building and cooling
Kitchen: the kitchen
Grid connection
Kitchen: the gas line
What the site can run
and so, the tokens it can make
Illustrative numbers. The point holds at any scale: power that is not connected caps every layer above it.
Why power is now the bottleneck
Three numbers tell the story.
Three numbers
Why power is now the bottleneck
Queues · Europe
7 to 10 years
Average wait to connect a data centre in Europe's long-established hubs.
Ember, June 2025
Concentration · Ireland
5% to 23%
Data centres' share of the country's metered electricity, 2015 to 2025.
Central Statistics Office, July 2026
Planning · India
13.6+ GW
Data centre demand the Central Electricity Authority (CEA) projects by 2032. It has asked four states to plan for it.
ThePrint, 2026
First, connection queues. Ember, the energy think tank, found in June 2025 that it takes an average of 7 to 10 years to connect a data centre to the grid in Europe's long-established data-centre hubs (Frankfurt, London, Amsterdam, Paris and Dublin), and up to 13 years in some markets. Reporting on the International Energy Agency (IEA) 2025 Energy and AI report, Latitude Media noted the agency's estimate that around 20% of planned data centre projects could be at risk of delays.
Second, concentration. Ireland's Central Statistics Office reported in July 2026 that data centres used 23% of the country's metered electricity in 2025, up from 5% in 2015. When one industry takes nearly a quarter of a national grid, every new connection becomes a national decision.
Third, India. In 2026, ThePrint reported that the Central Electricity Authority (CEA), India's technical power planner, had asked Maharashtra, Andhra Pradesh, Telangana and Tamil Nadu to plan for data centre demand, strengthen their networks and build adequate resources. Jitesh Patil, chairman and managing director of Telangana's southern electricity distribution company, put it plainly: "Data centres suck power suddenly. A sudden surge can lead to half a gigawatt of power consumed." The same report cites the authority's projection of more than 13.6 gigawatts of data centre demand by 2032.
Operators feel it too. Uptime Institute's 2025 global survey of more than 800 owners and operators described "major constraints on power availability" that are worsening.
Price lever, volume lever, or something else?
In the language of this series, Layer 0 is a price lever. The framework is explicit: the lower layers "set what a token costs, in dollars and in watts," and Layer 0 "sets the price and supply of power."
Power Supply adds something sharper than price: a ceiling. A site cannot produce more tokens than its connection can feed. If your capacity is not switched on, your cheapest token is the one you cannot make at all. Scarce power shows up further down the chain as higher cloud prices, tighter limits on how much computing you can book, longer waits for capacity, and caps on how many requests you can send per minute.
The framework names the key decisions:
- Siting: place capacity "where delivered power is available and priced well."
- Time-shifting: work that can wait, such as large overnight AI jobs and model testing runs, "shifts to the hours when power is cheaper or cleaner."
- Who controls the energy decision: "Each step away from owning the facility hands more of the energy decision to someone else." If you own the site, you see every meter. If you rent cloud computing, energy is folded into the price list.
Time-shifting matters more than it sounds. A February 2025 study from Duke University's Nicholas Institute, reported by POWER Magazine, estimated that the US grid could absorb 76 gigawatts of new demand if those new users accepted curtailment (being asked to power down) for 0.25% of the year, with curtailment events averaging about 1.7 hours. Flexibility, in other words, is a way to buy a connection sooner.
Illustrative day
Move "can wait" work to the hours when power is cheaper
Grid demand
Now: live chat and customer answers
Can wait
- 1Bulk processing jobs
- 2Model testing runs
- 3Overnight reports
Illustrative demand curve. Duke University's Nicholas Institute (2025, reported by POWER Magazine): the US grid could absorb 76 gigawatts of new demand if new users accepted powering down for 0.25% of the year, about 22 hours, in events averaging about 1.7 hours.
As a chemical engineer, I learnt early that a plant runs at the rate of its tightest constraint. Supply chain taught me the same lesson in commercial terms. For AI, that constraint is increasingly the grid connection, not the chip.
What I would do on Monday morning
- Find out where your tokens are made. Ask your cloud provider, or the company that hosts your servers, which regions and sites run your AI work, and how much of that capacity is switched on today versus only contracted for later. If you simply pay per use for someone else's AI model, track two things by region: how close you are to your usage limits, and how often requests get turned away for hitting them. They are among the few warnings you get when capacity further up the chain runs short.
- Split your workload into "now" and "can wait". Tag everything that can wait: bulk processing jobs, model testing, document backlogs, overnight reports. Move it to off-peak hours, or to the cheaper, slower bulk-processing options many AI providers offer. Work that can wait is work you can buy cheaper, and it is the flexibility grids are starting to reward.
- Treat power like any critical raw material. Have a second source. For anything customer-facing, have a second region you have actually tested. And when you read a vendor's capacity announcement, ask one question: how many of those megawatts are actually switched on today?
Next
Getting power is the first problem. Paying for it is the second. In Chapter 2, we look at Price and Procurement: electricity tariffs, time-of-day pricing, long-term power purchase agreements (PPAs), and why the same megawatt-hour (MWh, a thousand units) costs very different amounts in different regions.
This series builds on The Seven-Layer Tokenomics Stack by the Tokenomics Foundation, used under a Creative Commons Attribution 4.0 licence (CC BY 4.0). The commentary, and any mistakes, are mine. Views are my own and do not represent my employer.
Sources
Every figure was checked against its source on 4 November 2026.
- Tokenomics Foundation, The Seven-Layer Tokenomics Stack (full paper), October 2026, Creative Commons Attribution 4.0 licence (CC BY 4.0). Layer 0 definition, Power Supply component (utility feed and connection capacity, delivered power by site and hour, backup power), siting and time-shifting decisions, control by access model, "set what a token costs" quote.
- Tokenomics Foundation, The Seven-Layer Tokenomics Stack, 14 October 2026, Creative Commons Attribution 4.0 licence (CC BY 4.0). Layer 0 "sets the price and supply of power"; primary levers.
- Electric Reliability Council of Texas (ERCOT), Update to the Texas Senate, Panel 1: Assessing the Grid, 29 July 2026. About 474 gigawatts (GW) of large connection requests as of June 2026; about 90% data centres; about 205 gigawatts eligible for Batch Zero (preliminary, as of 28 July 2026).
- Electric Reliability Council of Texas (ERCOT), Large Load Update, Senate Committee on Business and Commerce, 1 April 2026. About 410 gigawatts as of 26 March 2026; about 87% data centres.
- Asian News International via Devdiscourse, India's power supply position remains adequate, installed capacity reaches 548.86 GW, 4 August 2026. Ministry of Power statement to the Rajya Sabha: 548.86 gigawatts installed as of June 2026.
- ThePrint, Data centre gold rush risks blackouts: central electricity body warns states, 2026. Central Electricity Authority (CEA) communique to four states; Jitesh Patil quote; the authority's projection of more than 13.6 gigawatts of data centre demand by 2032.
- Central Statistics Office, Ireland, Data Centres Metered Electricity Consumption 2025, 7 July 2026 (also reported by the Irish Examiner, 7 July 2026). 23% of metered electricity in 2025, up from 5% in 2015.
- Ember, Grids for data centres in Europe, 19 June 2025. Average 7 to 10 years to connect in long-established hubs; up to 13 years in some primary markets.
- Latitude Media, Report: Global grid congestion puts 20% of data center projects at risk, 10 April 2025, reporting the International Energy Agency (IEA) Energy and AI report (the agency's own page could not be opened when this was checked). Around 20% of planned projects at risk of delays.
- POWER Magazine, Duke researchers: grid flexibility key to accommodate load growth, 19 February 2025, reporting Duke Nicholas Institute, Rethinking Load Growth. 76 gigawatts at 0.25% curtailment; average event about 1.7 hours.
- Uptime Institute, 15th Annual Global Data Center Survey press release, 30 July 2025. More than 800 operators; "major constraints on power availability".