By 2028, 98% of the power required for artificial intelligence (AI) compute expansion is expected to rely on fossil fuels (87% natural gas, 11% coal), driving fuel competition among data center operators and creating new energy bottlenecks.
Waking up in the morning to chat with an artificial intelligence (AI) model on your smartphone or having complex business documents summarized in just a few seconds has now become part of our daily lives. The AI on the screen appears to operate as effortlessly as invisible digital magic. Press a single button, and vast knowledge pours out in real time.
However, behind this dazzling and seamless experience lies a massive physical infrastructure that is hard to imagine in our daily lives. Countless data centers (facilities housing large-scale computer servers and storage devices) around the world operate day and night, consuming staggering amounts of electricity.
Recently, through public documents and research findings from Anthropic (a leading AI research company developing conversational AI models), one of the largest and harshest realities facing AI technology has surfaced: the power issue.
Why This Matters
As the AI services we use daily become more advanced, the scale of the data centers supporting them snowballs. By analogy, this is much like a high-performance sports car continuously burning massive amounts of fuel to run at explosive speeds. The problem is that as this AI-exclusive compute capacity surges, power demand is exploding uncontrollably.
| Studies estimate that power consumed exclusively for AI purposes will reach 165 to 326 terawatt-hours (TWh, a unit representing power consumption comparable to the total electricity usage of a massive city or entire country) annually by 2028 [We did the math on AI’s energy footprint. | MIT Technology Review](https://www.technologyreview.com/2025/05/20/1116327/ai-energy-usage-climate-footprint-big-tech/). This is a critical issue that goes beyond mere cost concerns for tech companies, directly impacting global energy supply-demand structures and climate change response. |
In particular, according to Anthropic’s S-1 filing (registration statement submitted to the U.S. Securities and Exchange Commission) ahead of its initial public offering (IPO, the process of listing shares on a stock exchange for the public), securing power has emerged as a core bottleneck determining the success or failure of the entire AI industry Natural gas emerges as key AI bottleneck, Anthropic S-1 ….
Understanding Made Easy: The Vast Shadow of AI Factories and Fossil Fuels
Simply put, training and operating massive AI models is similar to running a gigantic factory where tens of thousands of scientists gather to perform computations day and night. Naturally, electricity is required to keep the factory running.
Imagine this: when you comfortably ask questions to an AI from home, hundreds of thousands of high-performance semiconductor chips emit intense heat and run tirelessly in the massive data center behind the scenes. Cooling and operating these servers consumes power equivalent to what hundreds of thousands of ordinary households use.
To meet this massive power demand, Anthropic’s 2028 power mix (the composition ratio of energy sources used to generate electricity) is projected to rely overwhelmingly on fossil fuels. Looking at the detailed figures, surprisingly, 87% will come from natural gas and 11% from coal 87% of the Anthropic power will come from gas and 11% from coal in 2028. Combined, this is a structure where a staggering 98% depends on hydrocarbon (fossil fuel) sources 87% of the Anthropic power will come from gas and 11% from ….
Even in an era where the transition to renewable energy is a hot topic, practical constraints are significant in handling exploding AI demand. New power sources such as wind, solar, batteries, nuclear power, and geothermal energy face issues with availability, timing, and cost, making it difficult to rapidly support the surging demand of AI data centers Anthropic’s New IPO Filing Puts AI and Data Center’s Energy …. As a result, natural gas—with its relatively high availability and abundant domestic U.S. infrastructure—has emerged as a key alternative and primary bottleneck Natural gas emerges as key AI bottleneck, Anthropic S-1 ….
Current Situation: Power Securing Competition and Warnings in Anthropic’s S-1 Filing
Currently, global big tech companies and AI research labs are staking their lives not only on grid power but also on projects to procure power independently (‘Bring Your Own Power’, abbreviated as BYOP). Industry-wide, AI data centers are trending to induce approximately 377 gigawatts (GW) of new fossil fuel generation Anthropic’s 2028 Power Mix Skewed 87% Toward Natural Gas.
| In the past, when the White House requested power-related information, Anthropic proposed that the U.S. should additionally build 50 gigawatts (GW) of dedicated power by 2027 [We did the math on AI’s energy footprint. | MIT Technology Review](https://www.technologyreview.com/2025/05/20/1116327/ai-energy-usage-climate-footprint-big-tech/). This vividly illustrates how immense energy is essential to elevate AI model performance to the next level. |
However, in the S-1 filing issued ahead of its initial public offering (IPO), Anthropic candidly warns that this power dependency could become a managerial risk Anthropic’s New IPO Filing Puts AI and Data Center’s Energy …. The filing explains that as competition among data center operators to secure natural gas intensifies, risks of unstable fuel supplies or soaring costs loom large Anthropic’s New IPO Filing Puts AI and Data Center’s Energy ….
For Anthropic, which is pursuing a massive listing targeting a valuation of $1 trillion (over 1,300 trillion KRW) Is the AI bubble bursting just asAnthropicgoes public?, solidifying a stable and cost-effective energy supply chain has become a matter of survival, just as much as securing technological prowess.
What Lies Ahead?
Over the next few years, the AI industry must simultaneously solve two massive challenges: ‘building smarter models’ and ‘securing more power.’
In the short term, reliance on natural gas will remain high 87% of the Anthropic power will come from gas and 11% from coal in 2028, but in the long term, movements to explore integration with nuclear power or next-generation clean energy sources will intensify further to cope with rising power cost pressures and environmental regulations. For general readers as well, it is worth noting that the cost of the convenient AI services we enjoy is not limited to monthly subscription fees, but is deeply tied to global energy infrastructure and cost structures.
AI’s Perspective
The dazzling growth of AI technology appears to take place in infinite digital space, but in reality, it stands upon a very physical and traditional foundation of massive power plants and natural gas pipelines. The reality that 98% of power must rely on fossil fuels in 2028 clearly demonstrates the physical limits facing AI. Moving forward, not only technological innovation but also the maximization of energy efficiency and the securing of clean power sources will become the starting points determining the true competitiveness of AI companies.
References
- 87% of the Anthropic power will come from gas and 11% from coal in 2028
-
[We did the math on AI’s energy footprint. MIT Technology Review](https://www.technologyreview.com/2025/05/20/1116327/ai-energy-usage-climate-footprint-big-tech/) - Dylan Patel – Anthropic & OpenAI will have most of the…
- 87% of the Anthropic power will come from gas and 11% from …
- Anthropic’s 2028 Power Mix Skewed 87% Toward Natural Gas
- Anthropic’s New IPO Filing Puts AI and Data Center’s Energy …
- Natural gas emerges as key AI bottleneck, Anthropic S-1 …
- Is the AI bubble bursting just as Anthropic goes public?
- Solar and wind
- Natural gas
- Geothermal and nuclear
- Shortage of senior developers
- Access to power such as natural gas and fuel supply competition
- Limits in semiconductor chip manufacturing processes
- 5 gigawatts (GW)
- 20 gigawatts (GW)
- 50 gigawatts (GW)
- 165 to 326 terawatt-hours (TWh)
- 10 to 50 terawatt-hours (TWh)
- 500 to 1000 terawatt-hours (TWh)