Trump Finds a Solution to the AI Power Crisis, and Google’s Counterattack Begins

This morning, I came across some truly intriguing news. The Trump administration is investing $80 billion to construct eight large nuclear power plants to address the AI data center power shortage. While the surge in power consumption due to the AI boom was anticipated, the idea of tackling it head-on with such a massive nuclear investment is quite surprising.

Even more interesting is Google’s decision to sell its TPUs, which were previously used internally, to external parties. In a market where NVIDIA almost monopolizes AI chips, Google’s move could significantly alter the landscape. Notably, the recent achievement of Gemini 3.0 topping the LM Arena leaderboard without NVIDIA GPUs is highly symbolic.

On the other hand, South Korea is facing growing concerns as it decides to exclude exceptions to the 52-hour workweek in the semiconductor special law. In the fiercely competitive global environment, I wonder if such regulations are appropriate. Let’s delve deeper into each issue.

The U.S. Nuclear Renaissance: An AI-Driven Energy Revolution

The AP1000 nuclear power plant construction project, pursued by the Trump administration in collaboration with Westinghouse, is noteworthy in many aspects. First, the scale is overwhelming. An $80 billion investment, equivalent to about 117 trillion won, marks the largest nuclear investment in decades. It’s understandable why industry giants like Brookfield and Cameco emphasize it as the “largest in decades.”

What personally intrigues me is that part of this investment utilizes Japan’s $550 billion investment fund in the U.S. Seeing the economic dimension of the U.S.-Japan alliance materialize this way suggests that global supply chain restructuring is transcending mere trade. Especially with the intensifying tech competition with China, strategic cooperation among allied countries seems to be extending to energy infrastructure.

Each AP1000 reactor can produce 1,100 MW of power, which is an astounding figure. It can supply electricity to 500,000 households, and from the perspective of AI data centers, this is even more significant. Considering that the power consumption of data centers operating large language models like ChatGPT or Claude is akin to that of small cities, the need for such large-scale nuclear power is evident.

Since the AI boom began, the power demand of data centers has increased exponentially. Big tech companies like Microsoft, Google, and Amazon are rushing to invest in nuclear power for this reason. Microsoft is investing in the reactivation of the Three Mile Island nuclear plant, Google has signed a contract with Kairos Power, a small modular reactor (SMR) developer, and Amazon has formed various partnerships with nuclear companies.

Interestingly, China is also responding to these moves. According to reports, China has launched the “Burning Plasma” international scientific program, aiming to demonstrate nuclear fusion power with scientists from over ten countries. The joint signing of the “Hefei Nuclear Fusion Declaration” by scientists from major countries like France, the UK, and Germany is also significant.

Google’s TPU Counterattack: Cracking NVIDIA’s Monopoly

Google’s announcement to sell TPUs (Tensor Processing Units) externally is major news. Until now, Google had used TPUs exclusively for its cloud services, but now it’s opening sales to external companies like Meta. Reports have emerged that Meta is discussing a multi-billion dollar investment to adopt Google’s TPUs starting in 2027.

The market reaction was immediate. Following the news, Alphabet’s stock surged by $18.82 (6.28%) to close at $318.47 on the 24th. This shows how positively investors view Google’s strategic shift.

NVIDIA’s dominance in the AI chip market has been overwhelming, with a market share exceeding 80%, almost monopolistic. As NVIDIA GPUs like the H100 and A100 became the standard for AI training and inference, prices skyrocketed. With the H100 priced between $30,000 and $40,000, alternatives were desperately needed.

Google’s Gemini 3.0 topping the LM Arena leaderboard with 1,501 points using only TPUs, without NVIDIA GPUs, is highly symbolic. It demonstrates that Google’s TPUs have reached a level where they can compete with NVIDIA GPUs. It’s particularly impressive that both training and inference were handled entirely by TPUs.

I personally believe this change will positively impact the overall AI ecosystem. As NVIDIA’s monopoly breaks, chip prices could drop, creating an environment where more companies can participate in AI development. With AMD introducing AI chips like the MI300X and Intel aiming to enter the market with the Gaudi series, the competition will likely intensify with Google’s TPUs joining the fray.

Google’s strategic shift is also intriguing from a cloud business perspective. Google’s cloud market share has been lower compared to Amazon AWS and Microsoft Azure. However, with TPUs as a differentiated hardware asset, Google could strengthen its position in the AI cloud market. As the performance of TPUs, specialized for training and inference of large language models, is proven, more AI companies might migrate to Google Cloud.

The fact that Meta is considering adopting Google’s TPUs is also significant. Meta leads the open-source AI model market with the Llama series, and combining it with Google’s TPUs could create substantial synergy. If Meta introduces TPUs to its data centers, it could be a win-win situation, providing hardware sales revenue for Google and AI infrastructure cost savings for Meta.

The Dilemma of South Korea’s Semiconductor Industry: Between Regulation and Competitiveness

The situation in South Korea is somewhat disappointing. Seeing the news that both ruling and opposition parties agreed to exclude exceptions to the 52-hour workweek in the semiconductor special law leaves me with mixed feelings. While protecting workers’ rights is important, the potential impact of such regulations on competitiveness in the fiercely competitive semiconductor industry is concerning.

The impact of the 52-hour workweek on R&D could be more significant than expected. Semiconductor development often involves continuous 24-hour processes, and immediate analysis of experimental results is frequently required. In time-sensitive projects, restrictions on working hours could inevitably slow down development.

Looking at the semiconductor investment competition between the U.S. and China, these concerns grow. The U.S. is injecting $52.7 billion into the semiconductor industry through the CHIPS Act, and China is also pouring massive funds at the national level. In such a scenario, competing with regulatory shackles could be a significant burden for South Korea.

However, there’s also an understandable aspect. South Korea’s work culture is notoriously intense, and the overtime culture in the semiconductor industry was particularly severe. Ensuring workers’ health and quality of life is also an important value, so these regulations aren’t entirely meaningless.

Personally, I think a more flexible approach is needed. For example, introducing optional work schedules for the R&D sector or allowing project-based flexible work hours could be viable solutions. The Democratic Party’s original proposal includes an ancillary opinion stating that “efforts will be made in the National Assembly to consider the realities of the semiconductor industry’s R&D regarding working hours,” and I hope it develops in this direction.

Considering the global competitiveness of domestic semiconductor companies like Samsung Electronics and SK Hynix, this is even more crucial. While they remain strong in the memory semiconductor market, they are lagging behind companies like TSMC and NVIDIA in new areas like AI chips and foundries. In such a situation, additional regulatory burdens could lead to weakened competitiveness.

In fact, the semiconductor industry is closely tied to the global supply chain, making it not just a domestic issue. Various factors are intertwined, such as supply chain restructuring due to the U.S.-China trade war, Japan’s material export restrictions, and the Netherlands’ EUV equipment export limitations. In this complex scenario, further internal regulation could make South Korea’s semiconductor industry’s position even more challenging.

However, there are hopeful aspects as well. South Korea’s semiconductor technology remains world-class, and it maintains a dominant position, especially in the memory semiconductor field. As the AI era fully unfolds, the demand for high-bandwidth memory (HBM) is surging, and Samsung and SK Hynix have strengths in this area. Korean-made HBM is used in NVIDIA’s AI chips like the H100 and H200.

Ultimately, finding a balance seems crucial. It’s about seeking ways to protect workers’ rights while maintaining global competitiveness. Benchmarking how other advanced countries resolve such issues could also be beneficial.

Summarizing the news I read today, it feels like the rules of the game in the AI era are rapidly changing. The U.S. is trying to solve power issues with nuclear energy, Google is challenging NVIDIA with TPUs, and China is aiming to lead next-generation energy technology with nuclear fusion. In this global competition, South Korea will have much to consider in terms of strategy.

Personally, I believe these changes will eventually have a positive impact on the overall AI ecosystem. As power supply stabilizes, competition in the chip market arises, and technological advancement accelerates, more innovations can occur. However, for South Korea not to fall behind in this wave of change, a more strategic approach seems necessary.

This article was written after reading a news article, adding personal opinions and analysis.

Disclaimer: This blog is not a news outlet, and the content reflects the author’s personal views. The responsibility for investment decisions lies with the investor, and no liability is assumed for investment losses based on this article’s content.