Google has gone to great lengths for power generation, but is AI really facing a power shortage?
TechFlow Selected TechFlow Selected
Google has gone to great lengths for power generation, but is AI really facing a power shortage?
AI lacks not electricity, but time.
Navigating Web3 tides with focused insightsAuthor: Facing AI
It's often said that the ultimate bottleneck for AI is energy, and Microsoft CEO Nadella recently indirectly confirmed this view. "Due to power shortages, many of Microsoft's GPUs are sitting idle in warehouses," Nadella stated.
Google’s recent move of sending TPUs into space and using solar power to run them seems like an echo of Nadella’s statement.
Yet curiously, although Nadella’s comment appears favorable to the energy sector, neither China's A-share market nor Nasdaq has seen any significant rise in their energy sectors as a result. From early November to publication time, the A-share energy index gained 0%, while the largest company in the Nasdaq energy sector rose only 0.77%.
On one hand, tech giants in Silicon Valley are loudly declaring power shortages—even resorting to space-based solutions—but on the other, despite such a clear signal, financial markets remain indifferent and unresponsive.
This naturally raises a question: Is the AI industry truly facing an electricity shortage?
According to OpenAI CEO Sam Altman, the answer is both yes and no.
Yes, because there indeed exists a current shortage of electricity; no, because the underlying issue is actually AI overcapacity. Although he can't pinpoint exactly when, Altman believes it won’t take more than six years before AI exceeds human demand, which would then reduce AI’s overall power consumption.
In other words, the AI industry faces short-term power constraints, but long-term, with decreasing energy demands per unit of computation, the power shortage problem will resolve itself.
01
In early November 2025, Google announced a project called "Project Suncatcher," which involves launching TPU chips into space and powering them via solar energy.
The sun radiates approximately 3.86 × 10²⁶ watts of energy every second—over a hundred trillion times more than total global human electricity generation. Satellites deployed in dawn-dusk sun-synchronous orbits can receive nearly uninterrupted sunlight, capturing up to eight times more energy annually than solar panels of the same area located at mid-latitudes on Earth.
Project Suncatcher collaborates with satellite company Planet Labs to deploy an AI computing cluster composed of 81 satellites in low Earth orbit, 650 kilometers above the surface. These satellites are designed to operate cooperatively within a 1-kilometer radius airspace, maintaining distances between 100 and 200 meters from each other. The plan aims to launch two test satellites by early 2027 to verify feasibility.
Although Google claims it reduced the energy cost per query of its Gemini model by 33 times within a year, clearly, Google still needs more power.
Generating power from solar energy in space isn’t a new concept, but it has long been hindered by a core challenge: efficiently and safely transmitting generated electricity back to Earth. Whether using microwave or laser beams, energy loss during transmission and potential environmental impacts have prevented large-scale implementation.
Project Suncatcher bypasses this issue entirely. Instead of transmitting data down, it uses the power directly in orbit for computation, only sending the final results back to Earth.
On-ground TPU supercomputer clusters use customized low-latency optical interconnects, achieving throughput of hundreds of gigabits per second (Gbps) per chip.
Current commercial inter-satellite optical communication links typically operate between 1 and 100 Gbps—far below what’s needed for massive internal data exchange within an AI computing cluster. Google proposes using dense wavelength division multiplexing (DWDM), theoretically enabling inter-satellite link bandwidth of about 10 terabits per second (Tbps).
Google has publicly addressed many technical challenges of Project Suncatcher, including formation control and radiation resistance.
But Google hasn’t explained how they will manage heat dissipation.
This is a critical physical challenge—there’s no air convection in vacuum, so heat can only be dissipated through radiation. In one paper, Google mentioned the need for advanced thermal interface materials and passive heat transfer mechanisms to reliably conduct chip-generated heat to dedicated radiator surfaces. However, the paper provided little detail on these technologies.
In fact, Google isn’t alone in considering orbital data centers. Just days before Google’s announcement, a startup named Starcloud launched a satellite equipped with Nvidia H100 chips, claiming plans for a 5-gigawatt space-based data center. Elon Musk has also indicated that SpaceX “will do” space-based data centers.
In May 2025, China’s Zhijiang Laboratory, in collaboration with Guoxing Aerospace, successfully launched and networked the first 12 satellites of its "Three-Body Computing Constellation."
So while sending AI into space may sound novel, the motivation is universal: if you need power, go where it’s abundant—because Earth simply doesn’t have enough to spare.
02
The primary culprit behind AI’s insatiable power hunger is Nvidia. In just four years, from the Ampere architecture to Blackwell, the power consumption of the company’s GPUs has increased several-fold.
A server rack using Hopper-architecture GPUs has a rated power of about 10 kilowatts; with Blackwell, due to increased GPU count, rack power approaches 120 kilowatts.
Moreover, given that modern deployments involve tens of thousands of GPUs, efficient communication relies heavily on Nvidia’s NvLink technology. Each NvLink connection consumes 4–6 watts, and two GPUs connect via 18 such links. These links converge onto NvSwitch units to enable non-blocking connectivity, with each NvSwitch consuming 50–70 watts.
In a GPU cluster with 10,000 H100s, 157 NvSwitches and 90,000 NvLink connections are required, resulting in a total auxiliary power draw between 730 and 1,100 kilowatts.
And that’s not all—cooling these GPUs also consumes substantial power. An 8-GPU H100 server using air cooling consumes around 150 watts just for cooling. Thus, a 10,000-GPU cluster requires approximately 187 kilowatts solely for thermal management.
Today, competition among major tech firms is increasingly measured not in raw computing power, but in energy units—gigawatts (GW). Companies like OpenAI and Meta plan to add over 10 GW of computing capacity in the coming years.
For context, 1 GW of power consumed by the AI industry could supply roughly 1 million average American households. According to a 2025 International Energy Agency (IEA) report, AI’s energy consumption is expected to double by 2030, growing nearly four times faster than grid expansion.
Goldman Sachs forecasts that global data center power demand will increase by 50% by 2027, reaching 92 GW. In the U.S., data centers’ share of total electricity demand is projected to rise from 4% in 2023 to 10% by 2030. Additionally, Goldman notes that individual large data center campus power requests now range from 300 megawatts to multiple gigawatts.
Here’s the twist:
NextEra Energy, North America’s largest renewable energy provider, and XLU, the representative ETF tracking U.S. utilities, have underperformed the broader market. Over the past 52 weeks, NextEra rose 11.62%, XLU gained 14.82%, while the S&P 500 surged 19.89%.
If the AI industry were truly facing a severe power shortage, energy suppliers and utility companies should be seeing outsized returns—not trailing the market.
Nadella offers a crucial clue: “Grid connection approvals take five years,” he said, “and building transmission lines takes 10 to 17 years.”
In contrast, GPU procurement cycles are measured in quarters, data center construction typically takes 1–2 years, and AI demand surges occur on a quarterly basis.
The mismatch in these timelines—orders of magnitude apart—is the real essence behind Nadella’s claim of AI power shortages.
Nadella also faces another immediate challenge. In 2020, Microsoft pledged to achieve carbon-negative operations, water-positive impact, and zero waste—all while protecting ecosystems.
Yet today, nearly 60% of the electricity used in Microsoft’s data centers still comes from fossil fuels, including natural gas. This results in annual CO₂ emissions equivalent to those of about 54,000 average U.S. households.
Meanwhile, the IEA’s October 2025 *Renewables Report* suggests that growth in global power generation capacity may outpace new electricity demand—including that from AI.
The report projects that between 2025 and 2030, global renewable capacity will grow by 4,600 GW—roughly equal to the combined current installed capacity of China, the EU, and Japan—and twice the growth seen in the previous five-year period.
Notably, nuclear energy stands out as the only source capable of providing stable, large-scale, low-carbon power. Traditional large reactors face long construction times, high costs, and significant risks. However, small modular reactors (SMRs) are changing this equation. SMRs can be mass-produced in factories like airplanes or cars, then transported by rail or road for on-site assembly—akin to “Lego blocks.”
SMRs have capacities of 50–300 MW, much smaller than traditional reactors (1,000–1,600 MW), but this is precisely their advantage. Smaller size means shorter build times, lower upfront investment, and greater siting flexibility. Factory production and modular assembly significantly reduce costs and risks.
SMRs are currently the hottest and most promising form of power generation. Google signed an agreement with Kairos Power to purchase 500 MW of SMR nuclear power—the first direct investment in SMR technology by a tech company. Microsoft, in January 2024, hired the former Nuclear Strategy and Projects Director from Ultra Safe Nuclear Corporation (USNC) as its Nuclear Technology Director, aiming to develop SMRs and even smaller micro-modular reactors (MMRs).
In other words, what Microsoft lacks isn’t electricity—it’s time.
03
Compared to energy infrastructure, reducing AI’s own power consumption is another crucial path forward.
Altman argues that the cost per unit of intelligence drops by 40x annually, suggesting we may soon need far less infrastructure. If progress continues, personal-level AGI might run on laptops, further reducing power needs.
Altman once wrote an article using his company’s products to illustrate this point. He noted that from early 2023’s GPT-4 to mid-2024’s GPT-4o, the cost per token dropped by about 150x in just one year. With fixed computational resources, the same task consumes significantly less power across different stages of AI development.
He emphasized that such dramatic cost reductions cannot stem solely from linear hardware cost declines—they must involve algorithmic optimization, architectural improvements, and enhanced inference engine efficiency.
The 2025 Stanford AI Index Report (HAI) supports this: the cost to run an AI model at GPT-3.5 level (MMLU accuracy 64.8%) plummeted from $20 per million tokens in November 2022 to $0.07 in October 2024—a 280x drop in 18 months.
On the hardware front, GPUs now feature two new energy-efficiency metrics: TOPS/W (trillion operations per watt) and FLOPS per Watt. These allow clearer tracking of efficiency breakthroughs.
For example, Meta’s fifth-generation AI training chip, Athena X1, achieves 32 TOPS/W under low precision—200% higher than its predecessor—with 87% lower idle power. Even under low-precision FP8, Nvidia’s H100 reaches only 5.7 TFLOPS/W.
Still, high-precision training tasks still require H100s, explaining why Meta is purchasing hundreds of thousands of Nvidia GPUs.
Epoch AI research shows machine learning hardware efficiency is improving at 40% annually—doubling every two years. New-generation AI chips show marked gains.
Nvidia’s H200 GPU delivers 1.4x better efficiency than the H100, indicating considerable room for improvement remains.
From a macro perspective, the most telling metric is data center efficiency itself, commonly measured by PUE (Power Usage Effectiveness).
An ideal PUE is 1.0, meaning all power goes to computation with none wasted on cooling or auxiliary systems. Ten years ago, average data center PUE was 2.5; today it’s 1.5, and Google’s latest facilities have reached 1.1. This means today’s data centers require only half the power for the same workload. Liquid cooling, free-air cooling, and AI-driven energy management systems continue to drive PUE downward.
Regardless of the outcome, the energy sector has already been reshaped by AI. Even if future AI demand declines, the momentum in energy innovation will fuel broader industrial advancement.
Join TechFlow official community to stay tuned
Telegram:https://t.me/TechFlowDaily
X (Twitter):https://x.com/TechFlowPost
X (Twitter) EN:https://x.com/BlockFlow_News
