Google has gone to great lengths for power generation, but does AI really lack electricity?

It is often said that the end of AI lies in energy. Microsoft CEO Satya Nadella indirectly confirmed this view in a recent interview. "Due to power shortages, many of Microsoft's GPUs are sitting idle in the warehouse," Nadella said.

Google recently came up with the ingenious idea of sending TPUs into space and using solar power to generate electricity for the machines, which seems to be an "echo" of Nadella's words.

Strangely enough, although Nadella's words seem to be good news for the energy industry, neither the A-share market in China nor the energy sector on the NASDAQ has seen an increase because of them. From early November to the time of writing, the A-share market had a 0% increase, and the largest company in the NASDAQ energy sector only had a 0.77% increase.

On the one hand, Silicon Valley giants are complaining about power shortages and even coming up with solutions like sending equipment into space. On the other hand, the market has turned a blind eye to such a clear signal and has not responded for a long time.

This inevitably raises a question: Does the AI industry really face a power shortage?

OpenAI CEO Sam Altman's view is: Yes and no.

It's yes because there is indeed a power shortage at present. It's no because the essence of the problem is actually an over - supply of AI. Although he is not sure exactly how many years it will take, within at most six years, AI will exceed people's needs, which will also lead to a decrease in AI's power demand.

In other words, the AI industry is facing a short - term power cut, but in the long run, as the energy consumption of AI decreases, the power shortage problem will be solved.

01

In early November 2025, Google announced a project called "Project Suncatcher". The operation mode of this project is to send TPU chips into space and use solar energy to generate electricity for them.

The sun radiates about 3.86×10²⁶ watts of energy per second, which is more than one hundred trillion times the current total global power generation of human society. Satellites deployed in a dawn - dusk sun - synchronous orbit can receive sunlight almost continuously, and the energy they receive in a year is eight times that of solar panels of the same area in the mid - latitudes on Earth.

Project Suncatcher is collaborating with satellite company Planet Labs to deploy an AI computing cluster consisting of 81 satellites in a low - Earth orbit 650 kilometers above the ground. According to the design, these satellites will work together in an airspace with a radius of 1 kilometer, maintaining a distance of 100 to 200 meters from each other. The project plans to launch the first two test satellites in early 2027 to verify the feasibility of the plan.

Although Google once said that it had reduced the energy consumption of a single query of its Gemini model by 33 times within a year, it is obvious that Google still needs electricity.

Generating electricity using solar energy in space is not a new concept, but it has long been plagued by a core problem, that is, how to efficiently and safely transmit the generated electricity back to the ground. Whether using microwave beams or laser beams, the energy loss during transmission and the potential impact on the ground environment make it difficult to implement on a large scale.

The idea of "Project Suncatcher" bypasses this step. It does not plan to transmit data back to Earth but directly uses the electricity for computing in space and only sends the computed results back to the ground.

The TPU supercomputer cluster on the ground uses customized low - latency optical chip interconnection technology, and the throughput of each chip can reach hundreds of gigabits per second (Gbps).

Currently, the data rate of commercial inter - satellite optical communication links is usually only in the range of 1 to 100 Gbps, which is far from meeting the needs of large - scale data exchange within the AI computing cluster. Google's proposed solution is to use dense wavelength - division multiplexing technology, which theoretically can increase the total bandwidth of each inter - satellite link to about 10 terabits per second (Tbps).

Google has explained many problems and solutions regarding "Project Suncatcher" to the public, such as how to control the cluster formation and how to resist radiation.

But Google has not explained how to dissipate heat.

This is a very tricky physical 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, preferably passive ones to ensure reliability, are needed to efficiently conduct the heat generated by the chips to the surface of a dedicated heat sink for radiation. The paper did not provide much information about the technical details of this part.

In fact, Google is not the only one with the idea of sending data centers into space. Just a few days before Google announced its plan, a startup called Starcloud had launched a satellite equipped with NVIDIA H100 chips and claimed to build a space - based data center with a power of 5 gigawatts. Elon Musk also said that SpaceX "will build" a space data center.

In May 2025, the first 12 computing satellites of the "Three - Body Computing Constellation" jointly developed by China's Zhijiang Laboratory and Guoxing Aerospace were successfully launched and networked.

So, although the idea of sending AI into space sounds novel, everyone has the same goal: if you need electricity, go get it from there because there isn't enough on the ground.

02

The main culprit for the AI industry's hunger for electricity is NVIDIA. The power consumption of this company's GPU products has increased several times in just four years from the Ampere architecture to the Blackwell architecture.

The rated power of a server rack using Hopper - architecture GPUs is about 10 kilowatts. By the Blackwell architecture, due to the increase in the number of GPUs, the rack power is close to 120 kilowatts.

Moreover, since the number of GPUs is now in the tens of thousands, when tens of thousands of GPUs communicate with each other, they need to rely on NVIDIA's NvLink interconnection technology to improve communication efficiency. The power consumption of each NvLink link is 4 to 6 watts, and there are 18 links between two GPUs. These NvLinks are then concentrated on the NvSwitch to achieve non - blocking connections, and the power consumption of an NvSwitch is 50 to 70 watts.

If a GPU cluster has 10,000 H100s, it will require 157 NvSwitches and 90,000 NvLink links. Then its power consumption will be approximately between 730 kilowatts and 1100 kilowatts.

That's not all. GPUs are also big power consumers in terms of heat dissipation. For the most common 8 - card H100 server, if an air - cooling system is used, the power consumption will reach 150 watts. So, for a 10,000 - card cluster, the heat - dissipation alone will require 187 kilowatts.

Currently, the competition among large technology companies has shifted from the traditional unit of computing power to the unit of energy consumption, the gigawatt (GW). Companies like OpenAI and Meta plan to increase their computing power by more than 10 gigawatts in the next few years.

For reference, the AI industry consuming 1 gigawatt of electricity is enough to supply the daily electricity needs of about 1 million American households. A report by the International Energy Agency in 2025 estimated that by 2030, the energy consumption in the field of artificial intelligence will double, and its growth rate is almost four times that of the power grid itself.

Goldman Sachs predicts that by 2027, the global data - center power demand is expected to increase by 50% to 92 gigawatts. The proportion of US data - center power demand in the total power demand will increase from 4% in 2023 to 10% in 2030. In addition, Goldman Sachs also pointed out that the power - access requests of some large data - center parks can indeed reach the level of 300 megawatts to several gigawatts for a single project.

But here comes the interesting part.

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's increase was 11.62%, and the increase of ETF XLU was 14.82%, but the increase of the S&P 500 index during the same period reached 19.89%.

If the artificial - intelligence industry is really facing a severe power shortage, then energy companies and the utility sector, as power suppliers, should receive excess market returns instead of underperforming the market.

In response, Nadella revealed a key clue. He said that "it takes five years to get grid - access approval" and "it takes 10 to 17 years to build transmission lines".

Meanwhile, the procurement cycle of GPUs is measured in quarters, the construction cycle of data centers is usually one to two years, and the explosion speed of artificial - intelligence demand changes on a quarterly basis.

The differences in these time scales are of an order of magnitude, and the resulting time mismatch is the essence of what Nadella said about the AI power shortage.

Moreover, Nadella has a current problem that cannot be solved. In 2020, Microsoft announced that it would "achieve negative carbon emissions, a net increase in water use, and zero waste" while protecting the ecosystem.

However, the reality is that nearly 60% of the electricity used in Microsoft's data centers still comes from fossil fuels, including natural gas. The annual carbon - dioxide emissions generated are approximately equivalent to the total emissions of 54,000 ordinary American households.

On the other hand, the "Renewable Energy Report" released by the International Energy Agency in October 2025 pointed out that the growth rate of global power - generation capacity may exceed the new power demand, including that of artificial intelligence.

The report stated that from 2025 to 2030, the global renewable - energy installed capacity will increase by 4600 gigawatts, which is roughly equivalent to the current total installed capacity of the three economies of China, the European Union, and Japan. Furthermore, the report predicts that the new installed capacity in these five years will be twice that of the previous five - year period.

It is particularly worth mentioning here is nuclear energy. Nuclear energy is the only option that can provide stable, large - scale, and low - carbon electricity. The problems with traditional large - scale nuclear power plants are long construction cycles, high costs, and high risks . But small modular reactors (SMRs) are changing this situation. SMRs can be mass - produced in factories as standardized modules, just like manufacturing airplanes or cars, and then transported to the site by rail or road for assembly, similar to a "Lego - style" construction method.

The single - unit capacity of an SMR is only 50 - 300 megawatts, much smaller than the 1000 - 1600 megawatts of traditional nuclear power plants, but this is precisely its advantage. A smaller scale means a shorter construction cycle, lower initial investment, and more flexible site selection. SMRs can be mass - produced in factories and then transported to the site for assembly, significantly reducing costs and risks.

SMRs are the hottest and most trendy way of generating electricity at present. Google once signed an agreement with Kairos Power to purchase 500 megawatts of SMR nuclear power, which is the first time a technology company has directly invested in SMR technology. In January 2024, Microsoft hired the former nuclear - strategy and project director of Ultra Safe Nuclear Corporation (USNC) to serve as its nuclear - technology director. The purpose is to develop SMRs and even smaller micro - modular reactors (MMRs).

In other words, what Microsoft lacks is not electricity but time.

03

Compared with the energy aspect, reducing the power consumption of AI itself is also an important development direction.

Altman's view is that the cost per unit of intelligence decreases by 40 times every year. It is very likely that we won't need so much infrastructure in a few years. Moreover, if the breakthroughs continue, personal - level general artificial intelligence may run on laptops, further reducing the demand for power generation.

Altman once wrote an article, taking his company's products as an example, to explain this problem. The article stated that from the GPT - 4 model in early 2023 to the GPT - 4o model in mid - 2024, in just one year, the cost per token decreased by about 150 times. With the same computing power, the power consumption required for the same business will decrease at different stages of AI development.

He said that such a significant price reduction cannot be achieved by simply linearly reducing hardware costs. It must involve a combination of factors such as algorithm optimization, model - architecture improvement, and reasoning - engine efficiency enhancement.

The 2025 Artificial Intelligence Index Report (HAI) of Stanford University confirmed this statement. The report stated that within 18 months, the cost of invoking an AI model at the GPT - 3.5 level (MMLU accuracy of 64.8%) dropped sharply from $20 per million tokens in November 2022 to $0.07 per million tokens in October 2024, a cost reduction of 280 times.

In terms of hardware, GPUs now have two new energy - efficiency measurement units: TOPS/W (trillions of operations per second per watt) and FLOPS per Watt (floating - point operations per second per watt). These units are used to more intuitively observe breakthroughs in energy efficiency.

For example, Meta's fifth - generation AI training chip, Athena X1, has an energy - efficiency ratio of 32 TOPS/W under low - precision conditions, a 200% improvement over the previous generation, and the no - load power consumption has decreased by 87%. Even in the low - precision range of FP8, NVIDIA's H100 only has an energy - efficiency ratio of 5.7 TFLOPS/W.

However, for some high - precision training tasks, the H100 is still needed, which is why Meta is purchasing hundreds of thousands of NVIDIA GPUs on a large scale.

Research data from Epoch AI shows that the energy efficiency of machine - learning hardware is increasing at a rate of 40% per year and doubling every two years. The energy efficiency of new - generation AI chips has been significantly improved.

NVIDIA's H200 GPU has a 1.4 - fold improvement in energy efficiency compared to the previous - generation H100. There still seems to be a lot of room for improvement.

From a macro perspective, the energy efficiency of the data center itself is the most worthy number to pay attention to. Usually, the Power Usage Effectiveness (PUE) is used to measure the energy consumption of a data center.

The ideal value of PUE is 1.0, which means that all electricity is used for computing without being wasted on cooling and other auxiliary systems. Ten years ago, the average PUE of data centers was 2.5, and now it is 1.5. Google's latest data center has even dropped to 1.1. This means that for the same computing task, only half of the electricity is needed compared to ten years ago. Liquid - cooling technology, free - cooling, and AI - driven energy - management systems are continuing to push this number down.

But no matter what the outcome is, the energy industry has been reshaped because of AI. Even if the demand for AI decreases in the future, the prosperity of the energy industry will drive the development of other industries.

This article is from the WeChat public account "Facing AI", author: Miao Zheng