It is said that the end of AI is energy, and Microsoft CEO Nadella indirectly confirmed this viewpoint in a recent interview. Due to power shortages, many of Microsoft’s GPUs are lying in the warehouse and not working. ”Nadella said so.
Google recently came up with a clever trick of sending TPU into space and using the sun to generate electricity for machines, which seems to be the “echo” of Nadella’s words.
But strangely, Nadella’s words seem to be favorable for the energy industry, but neither Big A nor Nasdaq’s energy sector has risen because of Nadella’s words. From early November to press, Big A rose 0%, while the largest company in the Nasdaq energy sector rose 0.77%.
On the one hand, Silicon Valley giants have repeatedly called for power shortages and even come up with solutions like “heaven”, but on the other hand, such clear signals have been ignored by the market and have not responded for a long time.
This raises a question: Is the AI industry really short of electricity?
OpenAI CEO Sam Ultraman’s view is: yes, and no.
Yes, it’s because there is indeed a shortage of electricity now; Saying no, it’s because the essence of the problem is actually an excess of AI. Although he’s not sure how many years it will take, at most within 6 years, AI will exceed people’s needs and also lead to a decrease in AI’s demand for electricity.
That is to say, the AI industry may experience a short-term power shortage, but in the long run, as AI energy consumption decreases, the power shortage problem will be resolved.
one
Google announced a project called “Project Suncatcher” in early November 2025, which operates by sending TPU chips into space and generating electricity from solar energy.
The energy radiated by the sun every second is approximately 3.86 times 10 to the power of 26 watts, which is over 100 trillion times the total global electricity generation of human society. Satellites deployed in sun synchronous orbits during dawn and dusk can receive almost uninterrupted sunlight on their solar panels, receiving eight times more energy per year than solar panels of the same area in mid latitude regions of the Earth.
The Sun Catcher program collaborates with satellite company Planet Labs to deploy an AI computing cluster consisting of 81 satellites in a low Earth orbit at a distance of 650 kilometers from the ground. According to the design, these satellites will work together within a radius of 1 kilometer in the airspace, maintaining a distance of 100 to 200 meters between each other. The plan is expected to launch the first two experimental satellites in early 2027 to verify the feasibility of the plan.
Although Google has previously stated that it has reduced the single query energy consumption of its Gemini model by 33 times within a year, it is clear that Google still needs electricity.
The use of solar power in space is not a new concept, but it has long been plagued by a core challenge of how to efficiently and safely transmit the generated electricity back to the ground. Whether using microwave or laser beams, the energy loss during transmission and potential impact on the ground environment make it difficult to implement on a large scale.
The idea of the “Japanese Catcher Plan” chose to bypass this link. It does not intend to transmit electricity back to Earth, but to directly use this electricity for calculations in space, and only transmit the calculated results back to the ground.
The TPU supercomputer cluster on the ground uses customized low latency optical chip interconnect technology, with a throughput of several hundred gigabits per second (Gbps) per chip.
However, the current commercial satellite to satellite optical communication links typically have data rates in the range of 1 to 100Gbps, which is far from meeting the large-scale data exchange needs within AI computing clusters. The solution proposed by Google is to use dense wavelength division multiplexing technology, which theoretically can make the total bandwidth of each satellite link reach about 10 terabits per second (Tbps).
Google has explained many challenges and solutions related to the “Sun Catcher Program” to the public, such as how to control cluster formation, how to resist radiation, and so on.
But Google did not explain how to dissipate heat.
This is a very tricky physics problem. There is no air convection in a vacuum, and heat can only be dissipated through radiation. Google once mentioned in a paper that advanced thermal interface materials and heat transfer mechanisms need to be used, preferably passive to ensure reliability, in order to efficiently conduct the heat generated by the chip to a dedicated heat sink surface for radiation. There is not much information provided in the paper regarding the technical details of this section.
In fact, the idea of sending data centers into space is not exclusive to Google. Just a few days before Google announced its plans, a startup called Starcloud had launched a satellite carrying the Nvidia H100 chip and announced plans to build a space-based data center with 5 gigawatts of power. Elon Musk has also stated that SpaceX “will do” space data centers.
In May 2025, the first batch of 12 computing satellites of the “Three Body Computing Constellation” jointly developed by China’s Zhijiang Laboratory and China National Space Administration have been successfully launched and networked.
So when it comes to sending AI into space, although it sounds novel, everyone’s goal is the same – if you want to use electricity, go up there and get it. The electricity on the ground is not enough for you to use.
two
The situation of AI’s thirst for electricity is mainly attributed to Nvidia. This company’s GPU products, from Ampere architecture to Blackwell architecture, have seen a several fold increase in power consumption in just 4 years.
A server rack using Hopper architecture GPU with a rated power of approximately 10 kilowatts; In the Blackwell architecture, due to the increase in the number of GPUs, the rack power approaches 120 kilowatts.
Moreover, since the units of GPUs are now in tens of thousands, when tens of thousands of GPUs communicate with each other, NvLink technology from Nvidia is also used to improve communication efficiency. Each NvLink has a power consumption of 4 to 6 watts, and there are 18 links between two GPUs. These NvLinks need to be centralized on an NvSwitch to achieve non blocking connections, while the power consumption of an NvSwitch is 50 to 70 watts.
If a GPU cluster has 10000 H100 blocks, it would require 157 NvSwitches and 90000 NvLink links. The power consumption is approximately between 730 kW and 1100 kW.
It’s not over yet, GPU is also a major power consumer in terms of heat dissipation. The most common 8-card H100 server, if using an air-cooled system, would require a power consumption of 150 watts. Therefore, for a 10000 card cluster, the heat dissipation alone would require 187 kilowatts.
Currently, the competition among large technology companies has shifted its measurement standards from traditional computing power units to energy consumption units such as gigawatts (GW). Companies like OpenAI and Meta plan to increase their computing power by over 10 gigawatts in the coming years.
As a reference, the AI industry consumes 1 gigawatt of electricity, which is enough to supply daily electricity to approximately 1 million American households. In a report by the International Energy Agency in 2025, it is estimated that energy consumption in the field of artificial intelligence will double by 2030, with a growth rate almost four times that of the power grid itself.
Goldman Sachs predicts that global data center electricity demand is expected to increase by 50% to 92 gigawatts by 2027. The proportion of data center electricity demand in the total electricity demand in the United States will increase from 4% in 2023 to 10% in 2030. In addition, Goldman Sachs also pointed out that power access requests for some large data center parks can indeed reach the level of 300 megawatts to several gigawatts per project.
However, something interesting has come.
NextEra Energy is the largest renewable energy company in North America, and the representative industry ETF that tracks the performance of the US utility sector is called XLU. In the past 52 weeks, NextEra has risen by 11.62%, ETF XLU has risen by 14.82%, but the S&P 500 index has risen by 19.89% during the same period.
If the artificial intelligence industry really faces severe power shortages, then energy companies and utility sectors, as power suppliers, should receive excess market returns instead of being unable to even outperform the market.
Nadella revealed a key clue regarding this. He said, ‘The approval for grid access takes 5 years,’ and ‘the construction of transmission lines takes 10 to 17 years.’.
At the same time, the procurement cycle of GPUs is measured on a quarterly basis, while the construction cycle of data centers is usually 1 to 2 years. The explosive speed of demand for artificial intelligence varies on a quarterly basis.
There is an order of magnitude difference between these time scales, and the resulting time mismatch is precisely what Nadella said about the nature of AI’s power shortage.
And for Nadella, there is another unresolved issue at the moment. In 2020, Microsoft announced that it would achieve carbon negative emissions, net increase in water use, and zero waste while protecting its ecosystem.
However, the reality is that nearly 60% of the electricity used in Microsoft data centers still comes from fossil fuels, including natural gas. The annual carbon dioxide emissions generated by this are approximately equivalent to the total emissions of 54000 ordinary American households.
On the other hand, the International Energy Agency pointed out in its Renewable Energy Report released in October 2025 that the growth rate of global power generation capacity may exceed the new electricity demand, including artificial intelligence.
The report proposes that during the five-year period from 2025 to 2030, the global installed capacity of renewable energy will increase by 4600 gigawatts, which is roughly equivalent to the total installed capacity of the three economies of China, the European Union, and Japan. Furthermore, the report predicts that the newly installed capacity in these five years will be twice the increase in the previous five years.
What needs to be specifically mentioned here is nuclear energy. Nuclear energy is the only option that can provide stable, large-scale, low-carbon electricity. The problems with traditional large-scale nuclear power plants are long construction periods, high costs, and high risks. But small modular reactors (SMRs) are changing this situation. SMR can mass produce standardized modules in factories like manufacturing airplanes or cars, and then transport them to the site for assembly by railway or road, similar to the construction method of “Lego bricks”.
The single unit capacity of SMR is only 50-300 megawatts, much smaller than the 1000-1600 megawatts of traditional nuclear power plants, but this is precisely its advantage. Smaller scale means shorter construction period, lower initial investment, and more flexible site selection. SMR can be mass-produced in the factory and then transported to the site for assembly, significantly reducing costs and risks.
SMR is currently the most popular and trendy power generation method. Google once signed an agreement with Kairos Power to purchase 500 megawatts of SMR nuclear power, marking the first time a technology company has directly invested in SMR technology. Microsoft hired the former Director of Nuclear Strategy and Projects at Ultra Safe Nuclear Corporation (USNC) as the Director of Nuclear Technology in January 2024. The purpose is to develop SMRs and smaller miniature modular reactors (MMRs).
In other words, what Microsoft lacks is not electricity, but time.
three
Compared to energy, reducing the power consumption of AI itself is also an important development direction.
Ultraman’s viewpoint is that the cost per unit of intelligence decreases by 40 times annually, and it is likely that we will not need so much infrastructure in a few years. And if the breakthrough continues, personal level general artificial intelligence may run on laptops, further reducing power generation demand.
Ultraman once wrote an article explaining this issue using his own product as an example. The article states that in just one year, from the GPT-4 model in early 2023 to the GPT-4o model in mid-2024, the cost per token has decreased by approximately 150 times. Under the premise of constant computing power, the same business will consume less electricity at different stages of AI development.
He said that such a significant price drop cannot be achieved solely through linear reduction in hardware costs, as it inevitably involves a comprehensive effect of algorithm optimization, model architecture improvement, and inference engine efficiency enhancement.
The Stanford University 2025 Artificial Intelligence Index (HAI) report confirms this statement, stating that within 18 months, the cost of calling AI models at GPT-3.5 level (MMLU accuracy of 64.8%) plummeted from $20/million token in November 2022 to $0.07/million token in October 2024, a 280 fold decrease in cost.
In terms of hardware, GPUs now have two new energy efficiency measurement units: TOPS/W (trillions of operations per watt) and FLOPS per Watt (floating-point operations per watt). This type of unit is designed to provide a more intuitive view of breakthroughs in energy efficiency.
For example, Meta’s fifth generation AI training chip Athena X1 achieves an energy efficiency ratio of 32 TOPS/W with low precision, which is 200% higher than the previous generation, and reduces no-load power consumption by 87%. Even in the low precision range of FP8, the Nvidia H100 has an energy efficiency ratio of only 5.7 TFLOPS/W.
However, for some high-precision training tasks, H100 is still needed, which is why Meta has to purchase hundreds of thousands of Nvidia GPUs on a large scale.
Epoch AI’s research data shows that the energy efficiency of machine learning hardware is improving at a rate of 40% per year, doubling every 2 years. The energy efficiency of the new generation of AI chips has significantly improved.
Nvidia’s H200 GPU has increased energy efficiency by 1.4 times compared to the previous generation H100. It seems that there is still a lot of room for improvement.
From a macro perspective, the energy efficiency of the data center itself is the most noteworthy number. PUE (Energy Efficiency) is commonly used to measure the energy consumption of data centers.
The ideal value of PUE is 1.0, which means that all electricity is used for calculations and not wasted on cooling and other auxiliary systems. Ten years ago, the average PUE of data centers was 2.5, but now it is 1.5. Google’s latest data center has dropped to 1.1. This means that the same computing task now only requires half of the electricity used in the past. Liquid cooling technology, free cooling, and AI driven energy management systems are continuing to push down this number.
But regardless of the outcome, the energy industry has been reshaped by AI, and even if the demand for AI decreases in the future, the prosperity of the energy industry will drive the development of other industries.
No comment content available at the moment