Artificial intelligence has spent the last few years being sold as software.
It is the helpful assistant on your phone, the clever button inside your laptop, and the slightly overenthusiastic chatbot being added to almost every service with a login screen. AI feels digital, invisible and almost weightless.
The infrastructure behind it is none of those things.
Every AI answer begins somewhere inside a data centre filled with specialized processors, cooling equipment and electrical hardware. Those facilities require land, water, transmission lines and an extraordinary amount of electricity. The cloud may sound wonderfully light and fluffy, but the machinery holding it up is becoming difficult to ignore.
Texas is discovering that faster than most.
Earlier this month, Governor Greg Abbott ordered regulators to conduct a comprehensive audit of data-centre projects seeking access to the Texas electrical grid. Under the directive, the audit must be completed before those projects can move forward through the state’s grid-connection process.
The decision became a larger technology story again this week after Abbott said the data-centre industry had “dug their own grave” by failing to earn public support. Reuters reported Monday that growing political opposition to AI infrastructure was even contributing to nervousness around technology stocks.
Texas is hardly known for discouraging either energy development or large corporate investment. That is precisely what makes the shift interesting. The AI industry has not suddenly encountered a government that hates technology. It has encountered a physical system that cannot expand at the same speed as a venture-capital presentation.
The numbers stopped looking hypothetical
In June, the Electric Reliability Council of Texas said it was tracking more than 438,000 megawatts of requests from proposed large electricity users. Nearly 89 per cent of that requested load came from data centres.
That does not mean all those projects will be built. Grid-connection queues frequently contain speculative proposals, competing projects and facilities that will never make it beyond a planning document. ERCOT itself has warned that its forecasts are planning tools rather than predictions carved into stone.
Even with that rather large grain of salt, the scale is remarkable.
ERCOT’s preliminary forecast projects approximately 367,790 megawatts of demand in its region by 2032. Its all-time peak demand was 85,508 megawatts in 2023. The forecast includes economic growth and other large industrial users—not just AI—but it illustrates how dramatically the expectations surrounding electricity have changed.
Abbott’s August 3 directive asks data-centre developers to disclose their expected electricity and water use, tax incentives, ownership, cooling technology and plans for generating their own power. Projects that fail to satisfy the requirements can be denied a connection.
Those are not unreasonable questions. In fact, they are the sort of questions that probably should have been answered before hundreds of enormous facilities lined up at the electrical door.
The International Energy Agency expects global data-centre electricity consumption to roughly double by 2030, reaching about 945 terawatt-hours. That would still represent just under three per cent of worldwide electricity demand, so AI is not about to consume every available electron on Earth.
The problem is concentration.
Data centres do not spread their demand evenly across countries and continents. They arrive in clusters, often near existing transmission infrastructure, and require huge amounts of continuous power in very specific locations. The IEA notes that data centres can be built in two or three years, while the power plants and transmission lines needed to support them often take considerably longer.
AI moves at software speed. The electrical grid does not.
Canada is already having the same conversation
This is not merely Texas wrestling with a Texas-sized problem.
In January, the Province of British Columbia and BC Hydro announced a competitive process for AI and data-centre projects. The stated goal is to manage rising electricity demand while creating a transparent process for deciding which projects receive access to clean power.
That matters in a province where hydroelectricity is abundant but certainly not infinite. Clean electricity still requires generating capacity, substations, transmission corridors and years of construction. A data centre cannot simply arrive beside a dam with an extension cord.
Ontario is preparing for similar pressure. Its Independent Electricity System Operator now expects data centres to account for 8.6 per cent of provincial electricity demand by 2050, roughly 60 per cent more than its previous forecast.
None of this means data centres are inherently bad.
They support the online services people already use and the AI tools governments and businesses increasingly consider essential. New facilities can generate investment, construction work and tax revenue. Better hardware, more efficient software, flexible power use and on-site generation could also reduce some of the pressure.
AI is not the only reason electricity demand is rising, either. Electric vehicles, manufacturing, population growth and the wider shift away from fossil fuels all require more power. Blaming every future grid problem on chatbots would be convenient, but it would not be particularly honest.
The issue is whether the companies building AI infrastructure are paying the full cost of their ambitions—or quietly expecting everyone else to absorb it through public subsidies, new transmission projects and higher pressure on local utilities.
For years, the technology industry treated computing power as though it could expand forever. Need a larger model? Add more processors. Need more processors? Build another data centre. Need electricity for the data centre? Somewhere, somehow, the grid would provide it.
That assumption is beginning to crack.
Texas has chosen an audit and a temporary pause. British Columbia is creating a competition for access. Ontario is rewriting its long-term demand forecasts. Other governments will almost certainly face the same choices as AI development accelerates.
The next phase of the AI debate will not be decided entirely by faster chips, smarter models or whichever company delivers the most impressive demonstration. It will also be decided by utility planners, community meetings and the unglamorous question of who builds the next transmission line.
Artificial intelligence may live in the cloud, but its future is increasingly being decided at the substation.
