Data Centers: Obsolete Before Online

The AI industry is selling the public a fantasy. The story being sold is: artificial intelligence is the future, the future needs more computing power, and therefore the country must build more data centers as fast as possible. More chips. More electricity. More land. More cooling. More transmission lines. More substations. More transformers. More public support. More money.

But the reality is that when this story is looked at closely through the lens of actual physics, engineering, and economics, it starts to fall apart.

In fact, many of these AI data centers may be obsolete before they are even online.

The problem is that the industry is trying to build slow, expensive, maintenance-heavy infrastructure around a technology that is changing too quickly. A large data center can take years to plan, permit, finance, build, connect to the grid, equip, and test. But AI chips, cooling systems, memory, networking equipment, and model designs can change much faster than that. By the time a facility is finished, the technology it was designed around may already belong to an older generation.

That is the first crack in the fantasy.

The public is encouraged to think of a data center as a kind of modern factory, a permanent building that will quietly produce the future. But an AI data center is not just a building. It is a giant industrial machine. Electricity enters the site, chips process data, and almost all of that electricity becomes heat. That heat has to be removed constantly. The more powerful the chips, the more heat they create. There is no way around this. It is basic laws of physics.

The industry talks about intelligence, automation, productivity, and national competitiveness. Engineers see something more grounded: power loads, heat, voltage, wiring, transformers, cooling failures, pump failures, battery aging, water demand, corrosion, leaks, replacement cycles, and constant maintenance schedules.

That is the real physical machine behind the word “AI.”

These facilities do not simply “use electricity.” They convert enormous amounts of electricity into heat, then require more equipment to move that heat out of the building. That means cooling systems, pumps, fans, chillers, liquid-cooling loops, sensors, backup systems, water systems, and constant monitoring. The denser the racks, the harder the cooling problem becomes. The bigger the facility, the bigger the stress, lots of stress.

This is where the fantasy begins to collide with thermodynamic reality.

A 100-megawatt AI data center is not a normal commercial building. It is a major industrial power load. If the country tried to build 5,000 of them, even at only 100 megawatts each, the total demand would be 500 gigawatts of continuous 24/7 power. Over a year, data centers would require roughly as much electricity as the entire United States currently uses.

That is not a serious infrastructure plan. That is a fever dream dressed up as national strategy.

And electricity is just only one part of the problem. These facilities also need advanced GPUs, CPUs, high-bandwidth memory, networking switches, fiber optics, power supplies, transformers, substations, copper, aluminum, cooling equipment, backup generators, batteries, trained technicians, and huge amounts of construction material. The shortage is not just chips. The shortage is the whole industrial stack.

Even if enough GPUs are made, that does not mean there is enough memory. Even if there is enough memory, that does not mean there is enough advanced packaging. Even then, there is not enough resource materials or manufacturing capacity. Even if the servers are ready, that does not mean the transformers, substations, and transmission lines are ready. Even if the equipment exists, that does not mean the local grid can handle the load. Every link in the chain has to work, and every link has its own physical limits.

The fantasy version says these facilities can be announced, financed, built, and plugged in. The real-world version says something much different and uglier.

Power has to be generated. Transmission lines have to be built. Transformers have to be manufactured. Cooling systems have to be installed. Land has to be available. Water has to be available. Workers have to be available. Local communities have to accept the project. The grid has to survive the added demand. None of this happens because a press release says it will.

Then comes wear and tear, as dictated by the laws of physics.

AI servers do not live easy lives. They run hot, hard, and continuously. Fans fail. Pumps fail. Power supplies fail. Memory fails. Drives fail. Batteries wear out. Connectors degrade. Liquid-cooling systems need inspection and repair. Backup generators need testing, fuel, filters, and parts. Electrical equipment must be maintained continuously. Cooling systems must be watched constantly. A data center is not a one-time purchase. It is a permanent maintenance burden.

This is another part of the AI bubble that does not get enough attention. The public hears about construction costs and investment totals, but the real cost is much larger. The real cost is building the facility, powering it, cooling it, repairing it, upgrading it, securing it, staffing it, replacing parts, and paying the debt attached to it.

And while all of this is happening, the technology is quickly aging.

AI chips depreciate quickly. A GPU system that is top-of-the-line today may be ordinary in two years and inefficient in four or five. It may still work, but that does not mean it is still worth running. In AI, performance per watt matters. Performance per dollar matters. If a newer chip can do more work using less electricity, the older system becomes less competitive and less profitable.

That is a serious problem when electricity is one of the biggest costs.

A data center filled with older GPUs may still function, but it may produce too little value for the power it consumes. The building may be finished. The grid connection may be active. The cooling systems may be running. The debt may still be owed. But the equipment inside may already be second-rate.

That becomes stranded compute.

Stranded compute is what happens when the infrastructure exists but no longer produces enough value to justify its cost. The building is real. The machines are real. The electric bill is real. The maintenance bill is real. The debt is real. But the economics no longer work.

That is the danger hiding inside the AI data-center fantasy.

The industry is acting as if the future definitely requires endless scaling: more chips, more power, more land, more cooling, more data centers, more debt. But that assumes today’s AI path continues without major disruption. It assumes bigger models keep winning. It assumes companies can make enough money from AI to pay for all this infrastructure. It assumes users and businesses will keep paying. It assumes the grid can expand fast enough. It assumes new technology will not make today’s hardware inefficient.

Those are not facts. They are bets, and many of those bets may not pay off.

Future AI models may become smaller and more efficient. Specialized chips may replace today’s general-purpose GPUs for many tasks. Edge AI may move more computation onto local devices. New memory systems, optical computing, analog chips, neuromorphic chips, or other designs may reduce the need for massive centralized GPU farms. Better software may also reduce the need for brute-force computing.

If any of that happens, many of today’s planned data centers could become yesterday’s infrastructure before they ever reach completion or full use.

That is what makes this look like a bubble. A potentially costly bubble.

A bubble does not mean everything involved is fake. Railroads were real during railroad bubbles. The internet was real during the dot-com bubble. Houses were real during the housing bubble. AI is real too. The problem is when a real technology gets inflated into an unrealistic financial story. The real thing becomes the seed of the fantasy. Then the fantasy becomes the business model, until it bursts.

That appears to be happening with AI infrastructure.

Companies announce massive projects. Politicians repeat the numbers. Investors chase the growth. Utilities plan upgrades. Land gets grabbed or bought. Public incentives are offered. Local governments are pressured to cooperate. The public is told this is necessary for national competitiveness and security. But many of these projects depend on future power, future chips, future customers, future revenue, future grid capacity, and future technology assumptions that may not arrive on time, or may not arrive at all or even exist.

This is the classic structure of a speculative bubble. The promise comes first. The financing follows. The infrastructure race begins. The public is told not to question it. Then, later, reality sends the bill, and then boom!

The public should ask a simple question: who benefits even if the long-term economics fail?

Landowners benefit. Construction firms benefit. Utilities may benefit. Chip companies benefit. Equipment suppliers benefit. Cloud companies may benefit. Financiers benefit. Politicians get to claim they are building the future. Consultants benefit. Defense contractors may benefit. But ordinary citizens may be left with higher electricity bills, grid stress, water conflicts and deficit, public subsidies, and infrastructure built around technology that has already moved on.

That is why this looks less like a clean technology plan and more like a transfer-of-wealth mechanism.

The physics says the energy and heat problem is enormous. The engineering says the systems are incredibly complex, fragile, and maintenance-heavy. The economics says the hardware depreciates very quickly and may become obsolete before the debt is paid.

Put those together, and the conclusion is hard to avoid.

Many AI data centers are not the future. They are a race against obsolescence.

They are being built around hardware that may be outdated before installation, powered by grids that may not be ready, cooled by systems that never stop working, financed by assumptions that may not survive, and justified by revenue projections that are still unproven.

The industry wants the public to believe this buildout is inevitable. But inevitability is not proof. A press release does not generate electricity. An executive order does not manufacture transformers. A market forecast does not remove heat. A national-security slogan does not make the economics work.

The physical world still gets the final vote.

Chips wear out. Cooling systems fail. Transformers are limited. Power grids overload. Hardware depreciates. Technology changes. Debt remains.

That is the reality behind the hype.

Data centers are becoming obsolete before they are even online, not because computing is unimportant, but because the AI industry is trying to build slow, expensive, maintenance-heavy infrastructure around a fast-changing technological target.

The fantasy is that America can build its way into AI dominance by covering the country with giant machine warehouses.

The reality is that many of those warehouses may be outdated, underpowered, overleveraged, and economically weak before they ever become useful.

That is not a stable foundation for the future or a civilization.

It is an infrastructure fantasy bubble waiting to pop, and the population will be left holding the bag of debt, while the rich get richer.