Will AI data centers outgrow nations in power demand?

AI data centers: How the race to AGI could consume the world’s resources

AI data centers are emerging as the beating heart of the artificial intelligence revolution. Unlike traditional facilities that stream video, host websites, or back up cloud storage, AI data centers are designed to train and run models that aim for Artificial General Intelligence (AGI). These centers are not simple server farms. They are supercomputers on an unprecedented scale, consuming colossal amounts of electricity, water, copper, silicon, lithium, gold, and silver.

The pace of development is so rapid that entire construction projects become obsolete before completion. The most striking example occurred in Temple, Texas, where Meta began building a traditional H-type data center in 2022 only to demolish it months later. By 2025, the same site housed two high-density AI facilities designed from the ground up for liquid cooling and extreme power draw. This episode shows how quickly AI requirements outstrip even the most advanced plans, and why global infrastructure is being reshaped around them.

Trading platforms pexels tima miroshnichenko 7567443.jpg? nc cat=111&ccb=1 7& nc sid=6ee11a& nc ohc=QwTDklpLC6wQ7kNvgEmg1GS& nc zt=23& nc ht=scontent.fpos1 1

TradingView — Track All Markets

Where the world charts, chats and trades markets. We’re a supercharged super-charting platform and social network for traders and investors. Free to sign up.

From server rooms to AI supercomputers

In the early days, data centers were often just converted office basements. As the internet grew, companies like Google, Microsoft, Amazon, and Meta built hyperscale facilities to store and distribute data worldwide. Location mattered. Datacenters clustered around hubs like Ashburn, Virginia, near Washington Dulles Airport, where connectivity and fibre-optic density guaranteed low latency for streaming and cloud services.

AI data centers have overturned this model. Their main task is either training large language models or running inference. Training is an internal process where terabytes of data are processed for weeks or months on end. It does not matter whether the training cluster is in Texas, Ohio, or Mongolia, as no customers interact directly with it. Even inference, the act of responding to prompts such as a ChatGPT query is not as latency-sensitive as streaming video. Adding half a second of delay to generate a text answer has little impact on the user experience.

The result is that location now matters less than raw compute capacity, access to high-voltage power, and cooling infrastructure.

The compute arms race

At the heart of AI data centers are GPUs, particularly those produced by Nvidia. Each new generation consumes exponentially more power. The Volta GPU drew 250 watts. The Hopper architecture rose to 700 watts. The current Blackwell GB200 superchip, combining two GPUs with a Grace CPU, demands 2,700 watts on a single board. Industry roadmaps already point toward 2,000-watt GPUs in the near future.

Density is the key metric. Modern racks such as Nvidia’s NVL72 pack 72 GPUs into a single system drawing 132 kilowatts. By comparison, a traditional enterprise rack typically consumes between 3 and 7 kilowatts. Even high-performance hyperscale racks rarely exceed 20 kilowatts. AI racks are not marginally hungrier; they are 10 to 40 times more power-intensive.

This scale forces designers to maximise copper interconnects within racks, since optical transceivers consume too much additional energy. Two miles of copper can be found inside a single Nvidia rack. Yet copper only works efficiently at short distances, meaning racks must be densely packed with GPUs to keep performance and energy costs manageable.

Cooling: from air to liquid

Power density brings heat density. Traditional air-cooled designs cannot cope with thousands of GPUs crammed into tight enclosures. AI accelerators like AMD’s MI300X are already mostly heatsink, with the chip hidden beneath. Liquid cooling is now essential.

Liquid can absorb 4,000 times more energy per unit of volume than air. It enables tighter rack layouts, reduces the size of heatsinks, and prolongs component lifespan by keeping chips cooler. Google has already switched its tensor processing units (TPUs) to liquid cooling, and Nvidia’s Blackwell systems are built with liquid loops in mind.

This requires massive plumbing at the facility level, water pipes, cooling towers, and redundant systems. But it is the only viable way to operate clusters with hundreds of thousands of GPUs.

Powering gigawatt campuses

AI data centers are not measured by floor space but by critical IT power. Retail facilities often provide less than 10 megawatts. Wholesale clusters around Dulles Airport run at 10–30 megawatts. Traditional hyperscalers range from 40 to 100 megawatts.

AI campuses exceed them all. Microsoft operates 300-megawatt facilities for OpenAI. Meta is building sites in Ohio and Louisiana aiming for gigawatt levels. To put that into perspective, Germany’s average power usage is around 60 gigawatts. A single AI supercluster could one day draw the same electricity as a small country.

Unlike consumer data centers, AI facilities run near full load constantly. A Netflix server farm might spike during evenings and weekends, but an AI training cluster runs flat out for months. This sustained draw requires direct access to high-voltage transmission lines and entire fleets of transformers, which are now in global short supply.

The energy industry reshaped

The hunger for electricity is pushing hyperscalers into the energy business. Microsoft has struck a deal to restart the decommissioned Three Mile Island nuclear plant to supply a 20-year stream of power for AI facilities. Amazon has acquired a 2,500-megawatt nuclear-adjacent site in Pennsylvania to secure capacity for its cloud. Google is investing in advanced nuclear reactors to fuel its next-generation clusters.

Gas-fired turbines are already in use at Meta’s Prometheus site in Ohio, while retrofitted Bitcoin mines with dedicated power plants are being transformed into AI campuses by companies like CoreWeave.

This trend signals a future where major AI players are also energy producers, not simply consumers. The scale of their demand rivals that of industrial nations.

Water, metals, and the hidden costs

Electricity is not the only resource consumed. Liquid cooling systems demand vast quantities of water, often competing with agriculture and residential needs in dry regions. The construction of each facility requires thousands of tonnes of steel, aluminium, and especially copper.

AI chips rely on silicon wafers produced in energy-intensive foundries. They also depend on rare earths, gold, and silver for connections, and lithium for backup power systems. The volume of materials required for a single 2-gigawatt AI campus is staggering, raising concerns about mining, environmental impacts, and supply chain security.

Why Meta demolished a data center

The Temple, Texas case illustrates how quickly requirements evolve. Meta’s original H-type design could provide 60 megawatts of critical IT power. By the time ChatGPT launched, it was already obsolete. The company razed the half-built structure and replaced it with two new buildings supporting 170 megawatts and liquid cooling.

This shows the unforgiving pace of the AI race. Facilities planned in 2022 are outdated by 2023. Delays of even a few months can cost billions in lost opportunity as rivals scale faster.

The race to AGI

Whether or not AGI is achievable remains debated. But the largest technology companies clearly believe it is both possible and immensely profitable. They are committing trillions of dollars to reach it first. Each GPU rack, each megawatt, and each nuclear plant brought online is another step in that race.

AI data centers are no longer infrastructure hidden behind the internet. They are becoming one of the largest consumers of power, water, and metals on Earth. Within a decade, individual campuses may exceed the electricity use of megacities, while global clusters could match the demand of industrialised nations.

What happens next

If current trends continue, AI will reshape not only computing but also energy, manufacturing, and geopolitics. Nations with abundant renewable or nuclear capacity may become global hubs for AI training. Mining industries will expand to meet demand for copper, lithium, and rare earths. Local communities may face competition for water and land.

The pursuit of AGI is effectively the pursuit of resources. Whoever builds the largest, fastest, most efficient AI data centers will have the best chance of reaching it. And as long as the potential reward is measured in trillions of dollars, the consumption of natural resources on a planetary scale will continue.

____________________