The most expensive "idle item" in the AI ​​circle was born: Nvidia's top-end GPU, worth billions of dollars, is lying in Microsoft's warehouse eating dust.The reason is incredibly simple:No power. Just recently, Microsoft CEO Satya Nadella and OpenAI CEO Sam Altman were guests on Brad Gerstner's "BG2" podcast interview show. They rarely "exposed themselves" and personally confirmed this shocking dilemma:"Our biggest problem now is not computing power, but electricity... In fact, I have a bunch of chips piled in inventory, but there is no power available."

As soon as this statement came out, the market was shocked. While all companies are still worried about not being able to get chips, Microsoft, which has a large number of GPUs, has encountered a more fundamental problem:Not enough "sockets" found.

From "computing power shortage" to "power shortage", the rapid development of AI seems to have hit the hardest wall in the physical world.


The warehouse is full of chips, but there is no "socket"

For a long time, the scientific and technological community has generally believed that "computing power" is the biggest obstacle to the development of AI. Therefore, companies such as Microsoft and OpenAI have been purchasing GPUs frantically.

But now it seems they were running too fast.

Nadella admitted in the podcast,The number of chips Microsoft ordered has obviously exceeded the power capacity that their contract can guarantee.

"In this particular situation, you simply can't predict the cycle of demand and supply," Nadella said. "Our biggest issue is power and whether we can build data centers quickly close to the power source."

The so-called "warm shell" is a commercial real estate term, which refers to buildings with complete infrastructure that allow tenants to "move in". in other words,Microsoft currently has enough chips and management, but it lacks a "live house".

This is not just Nadella’s personal opinion. Microsoft Chief Financial Officer Amy Hood also confirmed to analysts during last week's earnings call: Obtaining computing hardware has not been Microsoft's bottleneck. "What we lack is space and power."

She even added: "This has been going on for several quarters. I thought we would catch up, but no. Demand is still growing."

Financial report data shows that Microsoft invested US$11.1 billion in leasing data centers in the first quarter of fiscal year 2026.In 2025 alone, Microsoft will add approximately 2 gigawatts of new computing capacity globally, bringing the total number of data facilities to more than 400.

A report released by TD Cowen last month showed that in the third quarter of 2025, the data center capacity leased by ultra-large-scale enterprises in the United States exceeded the total in 2024. While most of the increase came from Oracle and OpenAI, Google, Meta, Microsoft, Amazon and Anthropic also saw significant increases in computing power.

The S&P Global report also released last month pointed out that by the end of 2025, the demand for grid power from data centers across the United States will increase by 22% compared with 2024, and by 2030, this number will triple.

AI "hungry for electricity" is too fierce, and software giants encounter "hard" challenges

Why does this situation occur?

The answer is simple:Compared to heavy assets such as large power plants, chips and codes are technologies that are easy to scale and deploy quickly.Technology companies that have become accustomed to these two technologies have hit the "south wall" of the energy world.

Over the past decade or so, U.S. electricity demand has been roughly equal to supply. But over the past five years, data center power demand has begun to soar, growing at a rate far faster than utilities can plan new generating capacity.

This trend has caused residential energy bills to skyrocket, showing that AI infrastructure construction is having a negative impact on ordinary people.to this end,OpenAI has even called on the federal government to build 100 gigawatts of new generation capacity every year, and said that this is a strategic asset for the United States to maintain an advantage in the AI ​​competition with China.

This has also forced data center developers to start looking for "off-grid" solutions, that is, skipping the public power grid and delivering power directly to the data center.

Altman, who also participated in the podcast conversation, believes that hidden dangers are brewing: If AI continues to develop at an alarming rate of "reducing the cost of unit intelligence by 40 times every year," then "from an infrastructure construction perspective, this exponential change is really alarming."

The head of OpenAI has already set his sights on the future energy field.He has personally invested in nuclear fission startup Oklo, nuclear fusion startup Helion, and a solar startup called Exowatt.

The cheaper it is, the hungrier it is: AI’s “unlimited” appetite

Although solar photovoltaic technology is being adopted by technology companies due to its low cost, rapid deployment and zero emissions, it requires construction time just like data centers. In fact, new gas turbine orders placed now may not be delivered until the end of this decade.

The demand for AI changes much faster than the completion speed of any project.

Altman admitted that if AI energy efficiency improves or demand growth is less than expected, some companies may face the dilemma of idle power generation facilities.

But he himself seems to be a firm believer in the Jevons Paradox. The theory is that improvements in resource usage efficiency will lead to increased usage, thereby pushing up overall demand.

"Suppose tomorrow, the computing price per unit of intelligence drops by 100 times," Altman said, "you will find that the increase in usage is far more than 100 times. Many computing power applications that are not economically feasible at the current cost will usher in an explosion by then."

in other words,The cheaper and more efficient AI becomes, the greater the world’s demand for it will be, and the more insatiable the world’s thirst for power will be.

The Real Risk: The Energy Revolution and Localized Computing

However, the development and changes of energy supply also expose all AI practitioners to huge risks. Future energy revolutions may make today's massive power investments in vain.

As Altman warned during the talk: “If a very cheap form of energy becomes available on a large scale very quickly (such as nuclear fusion), many companies with existing expensive power contracts will be hit hard.”

The conversation also revealed another “ticking time bomb” that bets on hyperscale AI data centers must face: the rise of localized computing.


Altman imagined: "One day, we will create a revolutionary consumer device that can run GPT-5 or GPT-6 level models locally with low power consumption."

If AI models can really run locally and efficiently on personal computers or mobile phones, then the inference needs of the huge AI data centers that technology giants have spent billions or tens of billions to build may not be fulfilled at all.

The podcast host also immediately commented: "This is indeed remarkable, but it will obviously also worry companies that have invested heavily in building centralized computing clusters."

By then, this situation could hasten the bursting of the AI ​​bubble. And when the bubble bursts, what will be exposed to the risk is a terrifying market value of nearly 20 trillion US dollars.

From "computing power shortage" to "power shortage", AI development has obviously entered the deep water zone. Those expensive chips in Microsoft's warehouse that can't light up may be just the beginning of the pain of this huge transformation.