Half of AI's New Servers Will Never Turn On

The Irony No One Is Talking About

For the last three years, the story of artificial intelligence has been almost entirely a story of construction. Companies have poured hundreds of billions of dollars into data centers, GPUs, cooling systems, and power contracts. Every quarter brings announcements of new "AI campuses," hyperscale facilities, and multi-gigawatt projects meant to train and run the models of tomorrow. The implicit promise is simple: more compute equals smarter, cheaper, better AI.

But there is a flaw hiding inside that promise, and it is not a software problem. It is a plug problem. Across the world, the single most scarce resource for an AI company is no longer chips, capital, or talent. It is electricity — and enough of it, reliably, at the right place and the right time. The result is a risk that financial markets are only beginning to price in: a large share of the infrastructure being built today will never actually run.

This article explains why that is happening, how the stranded-asset risk works, and why it may be the most important — and most overlooked — bubble forming in AI today.

Why this matters: We tend to think of the AI bubble as a question of stock prices or overhyped models. But the deepest risk may be physical. If half of the new data centers cannot get the power they need, the money spent building them is gone — and the companies that built them may not survive it.

What Is Actually Happening

Data centers are not magic boxes. Behind every chatbot, image generator, and language model is a rack of servers that must be powered and cooled every single second they are alive. Training a single large model can consume as much electricity as a small city over its lifetime, and running those models day after day — serving billions of queries — is a steady, unrelenting demand that does not sleep. AI compute is one of the fastest-growing sources of electricity demand in history, and the growth is outpacing the grids that feed it.

Here is the core of the problem. Building a data center takes months. Connecting it to the power grid — a process called interconnection — can take years. Securing a reliable, dedicated supply of electricity can take even longer. Meanwhile, the financial model of the facility was built on the assumption that it would be fully operational and selling its compute to AI buyers almost immediately. When the power does not arrive on time — or does not arrive at all — the building stands empty, the debt does not go away, and the revenue the business plan depended on never shows up.

This is what analysts at firms such as Morgan Stanley and S&P Global have begun flagging: the AI buildout is running into a power bottleneck that could slow or stall projects already under construction. Reports from Reuters, Fortune, and others have documented data centers waiting years on grid queues, utilities redirecting power away from residential neighborhoods to serve data centers, and developers hedging their bets with natural-gas turbines because they cannot rely on the grid. The physical reality is catching up to the financial optimism.

The core mismatch: data centers are built and financed on a timeline of 12 to 24 months, but the grid operates on a timeline of 5 to 10 years. When those two timelines collide, the facility that cannot get power is a "stranded asset" — expensive, unfinished, and useless.

What "Stranded Assets" and "Power Shortage" Actually Mean

A stranded asset is an investment that has lost value — or become a liability — before the end of its expected life. In energy, the classic example is a power plant built for expensive fossil fuels that never earns back its cost because cheaper alternatives undercut it. The same logic can apply to AI data centers: a building that cannot get the power it needs to operate is a plant that cannot generate revenue.

The "half" figure you may have seen is not a precise census — it is a directional warning from analysts modeling how much new data-center capacity is being announced versus how much new generation and grid connection can realistically be delivered. The exact number varies by forecast, but the shape of the argument is consistent: announced capacity is growing far faster than the power supply that can support it. Even a modest share of facilities that cannot come online is enough to turn a boom into a bust on a company's balance sheet.

Concrete Examples: What This Looks Like in Practice

Numbers stay abstract until you picture the physical reality. Here are a few scenarios that make the risk tangible.

🏗️

The building that never boots

A company signs a $1 billion deal to build a 500-megawatt data center on land it leased cheaply. The servers arrive. The racks are wired. But the grid interconnection queue is three years deep. The data center sits dark, the loan payments are due, and no customer can use it.

⚡

The neighborhood that gets cut off

A utility, overwhelmed by data-center demand, reroutes power to keep a facility running. Nearby homes face rolling blackouts. The public and political backlash often triggers new rules that freeze or delay exactly the projects that needed the power.

🛢️

The gas turbine fallback

Because the grid cannot be trusted, developers build their own natural-gas turbines on site. This adds cost, pollution, and another reason the project's economics collapse if AI demand turns out to be lower than expected.

📉

The model that priced in power that never came

An AI startup raises money to train and run models depends on renting compute. If their assumed infrastructure slips, their runway evaporates faster than their code can be written.

Real-World Signals Already Showing Up

This is not a hypothetical scenario playing out in a spreadsheet. The signals are visible across several industries and regions:

Grid queues that stretch for years: In the United States, thousands of gigawatts of data-center and generation projects are waiting in interconnection queues — a backlog that analysts compare to the delays that stalled renewable energy projects.
Utilities fighting back: Several US utilities have paused or restricted new data-center connections, citing grid capacity limits, and regulators have opened investigations into whether data centers are driving up household electricity bills.
The "it's all based on guesswork" problem: Reporting has shown that even the estimates of how much power AI data centers will consume are highly uncertain — meaning financial models may be wrong in both directions, but the downside risk is what creates stranded assets.
International pressure: Bodies such as the UN and the European Union's security institute have flagged that AI's rising electricity demand is becoming a strain on power systems worldwide, not just in the United States.

Who Feels This First? Four Sectors

The power constraint does not hit everyone equally. The risk concentrates where capital intensity and continuous operation matter most.

🏢

Hyperscalers

The big cloud owners (Amazon, Microsoft, Google, Meta) are the largest buyers of AI compute. They have the balance sheets to weather delays, but even they are racing to secure power — and their overbuilding may create excess capacity if demand does not materialize.

🏭

Data-center operators & REITs

Companies whose entire business is owning and operating data centers are the most exposed. A single stranded facility can weigh down an entire portfolio and shake investor confidence across the sector.

🤖

AI model startups

Startups that raise money to train and run models depend on renting compute. If their assumed infrastructure slips, their runway evaporates faster than their code can be written.

🏠

Homeowners & local communities

The ultimate downstream cost falls on anyone sharing the grid. Higher bills, rolling blackouts, and delayed projects are the price of a power system stretched to serve machines that outpace the wires feeding them.

Future Scenarios: One to Ten Years Out

Short term (1–2 years)

Delays multiply. Projects that were supposed to launch slip. Developers lean harder on on-site gas turbines and battery storage. The first high-profile stranded-asset write-downs make headlines, and investors start asking harder questions about power in their AI investments.

Mid term (3–5 years)

The market begins to sort winners from losers. Facilities without firm power contracts struggle to attract customers. A wave of consolidations and bankruptcies hits the most over-leveraged operators. Power availability becomes a core part of any AI company's valuation, not an afterthought.

Long term (10+ years)

If the buildout continues on today's trajectory, a significant share of mid-century data-center capacity could be stranded — buildings built for an AI demand that never fully arrived, retired early and recycled. If the buildout is reined in and paired with real generation growth, the AI infrastructure of the 2030s could be both smaller and more durable than the headlines suggest.

How This Technology Is Evolving

The pressure of the power bottleneck is itself driving innovation. Several trends could reshape the landscape:

Smarter, more efficient chips: New AI accelerators are being designed to do more work per watt, reducing the electricity cost of each inference.
On-site generation & microgrids: Developers are pairing data centers with natural-gas, nuclear, and solar-plus-storage projects to become power-independent.
Cooling breakthroughs: Liquid cooling and other techniques reduce the enormous amount of electricity that cooling systems themselves consume.
Location rethinking: Companies are considering where to build based on where cheap, clean power already exists — rather than where land is cheapest.

But none of these fully solve the core math. Efficiency buys time; it does not eliminate the fundamental gap between how fast data centers can be built and how fast the power can arrive.

Implications: The Upside and the Downside

Potential Positives

Forces a healthier market: A correction could prune overhyped, overleveraged projects and leave a more sustainable AI industry.
Accelerates clean energy: The demand for reliable power could accelerate nuclear, solar, and storage investment.
Spurs innovation: The constraint is already driving more efficient chips, better cooling, and smarter grids.

Risks and Negatives

Stranded assets: Billions spent on facilities that never run could trigger write-downs and bankruptcies across the sector.
Higher costs for everyone: Power shortages and grid strain can push up electricity bills for households and businesses.
A bubble that bursts loudly: If investors suddenly realize how much capacity is stranded, valuations for AI-infrastructure companies could fall sharply and fast.
Wasted resources: Hardware built and then scrapped is not just lost money — it is embodied carbon and materials discarded.

Conclusion: The Bubble Is Physical, Not Just Financial

The AI bubble that Wall Street worries about is usually about valuations and hype. But the deeper, more durable risk is physical: a growing share of the data centers being built today may never have enough electricity to turn on. When that happens, the money is not just paper losses — it is concrete and steel left empty, debt left unpaid, and hardware left to be scrapped.

The good news is that the same constraint is forcing a more honest, more efficient, and potentially more sustainable AI industry. The bad news is that markets price in optimism, and optimism is exactly what power shortages will puncture.

So the real question for anyone investing in, building, or cheering on AI may not be "how smart can the model get?" but "where is the power coming from?" Until that answer is solid, the biggest risk in AI may be the buildings that never switch on.

What this means for you: Whether you are an investor, a founder, or just someone curious about AI's future, watch the power — not just the models. The companies that solve the electricity problem will define the next decade of AI. Those that ignore it may end up as cautionary tales of the bubble.