Did you know that the computing power required to train AI has been doubling every 5.5 months? Since 2010, that growth rate has reached roughly 4.6x a year. This is not a passing boom but a seismic transformation of the semiconductor industry, and here is what is really behind it.

The Core of the "Bubble" Debate: Why AI Demand Is Real

Skepticism about generative AI persists. Many compare it to the real estate bubble of the 1990s or the dot-com bubble of the 2000s. However, the data tells a fundamentally different story about today's semiconductor demand.

According to the World Semiconductor Trade Statistics (WSTS), the global semiconductor market reached $772.2 billion in 2025, recording a remarkable 22.5% year-over-year growth. What's striking is that this growth is concentrated in memory and logic semiconductors. AI-focused GPUs and HBM (High Bandwidth Memory) are driving the market, while other segments remain largely flat, a clear "bifurcation" in the industry.

Unprecedented Computing Demand: The Numbers Tell the Truth

The reason AI semiconductor demand cannot be called a "bubble" lies in the continuous and predictable growth of computing requirements.

According to analysis by Preferred Networks (PFN), the computing power needed to create state-of-the-art AI models has been doubling every 5.5 months since 2010. This translates to approximately 4.6x annually, or about 1 million times over a decade. Even more striking is that AI inference demand exploded in late 2024, recording an 80x year-over-year increase.

Google's usage exemplifies this trend: their token consumption grew from 480 trillion in May 2025 to 980 trillion just two months later, nearly doubling. This demonstrates the rapid adoption of AI in software development.

NVIDIA's Dominance and the "Winners' Circle"

NVIDIA has benefited most from this transformation. Commanding approximately 90% of the AI semiconductor market, the company topped the global semiconductor rankings in Q2 2025. Their Data Center segment alone achieved an operating profit margin exceeding 75%, an extraordinary figure for any manufacturing business.

However, industry-wide, the "AI boom" benefits only a handful of companies. As Masahiko Todoroki, Managing Executive Officer at Shin-Etsu Chemical, noted: "Today's semiconductor industry cannot be described in a single phrase." The divide between "AI and everything else" is becoming increasingly stark.

Test equipment giant Advantest doubled its profits riding AI semiconductor demand, while many Japanese manufacturers focusing on MCUs and analog semiconductors continue to struggle.

Japan's Semiconductor Industry: Where the Real Strengths Lie

Japan's semiconductor industry, which once commanded over 50% global market share, has shrunk to below 10%. However, the country maintains overwhelming presence in manufacturing equipment and materials:

  • Silicon Wafers: Over 50% global share
  • Photoresist: Approximately 90% global share
  • Cleaning and Inspection Equipment: 60-80% global share

Global semiconductor manufacturing equipment sales reached a record $33.66 billion in Q3 2025. The Semiconductor Equipment Association of Japan (SEAJ) forecasts Japanese equipment sales will reach ¥6 trillion ($40 billion) by fiscal year 2027.

Companies like Tokyo Electron, Advantest, Disco, and SCREEN Holdings are essential to AI semiconductor manufacturing. As AI proliferates, demand for their equipment will only intensify.

The Rapidus Challenge: Domestic 2nm Semiconductor Production

Rapidus, established in 2022, has become the symbol of Japan's semiconductor strategy. The company is building a state-of-the-art 2nm process semiconductor facility in Chitose, Hokkaido, targeting mass production by 2027.

This government-led project has attracted investment from eight major Japanese corporations including Toyota, Sony, NTT, and Denso. The government has already committed substantial public funds to Rapidus, and across the wider AI and semiconductor sector, roughly ¥10 trillion ($67 billion) in public support is planned over the next decade.

The 2nm generation that Rapidus will adopt employs a new technology called GAA (Gate-All-Around) structure. This evolution from the conventional FinFET structure surrounds transistors with nanosheets from all directions, achieving both performance and power efficiency.

Notably, Rapidus is developing an AI-assisted design tool called "Raads." Leveraging large language models (LLMs), it aims to reduce design time by 50% and design costs by 30%.

TSMC Kumamoto: Revival of "Silicon Island Kyushu"

The other pillar of Japan's strategy is the entry of TSMC, the world's largest foundry, into Kumamoto Prefecture. Following the first factory that began operations in 2024, construction of a second factory with an investment of began in October 2025.

The second factory initially planned to manufacture at 6nm to 12nm processes, but reports suggest it may shift to the more advanced 3nm process in response to surging AI semiconductor demand. If realized, this would also serve as a competitive response to Rapidus.

TSMC's entry has attracted over ¥6 trillion ($40 billion) in semiconductor-related capital investment to Kyushu, with more than 40 companies from Kanto and Chubu regions and 11 from overseas announcing expansion plans. The revival of Kyushu's "Silicon Island" semiconductor industry is becoming increasingly tangible.

The Coming "Inference Economy"

Current AI demand centers on the "training" phase, but a shift toward the "inference" phase is underway. This involves deploying models trained in data centers to edge devices like PCs, smartphones, and automobiles.

This shift could fundamentally restructure the semiconductor market. Inference doesn't necessarily require NVIDIA's GPUs, opening opportunities for more power-efficient and affordable processors. This could present new opportunities for Japanese companies specializing in MCUs and analog semiconductors.

For OpenAI's GPT-4, training costs were estimated at approximately $150 million, while annual inference service costs reach roughly $2.3 billion, over 15 times higher. The expansion of the inference market is certain to become the center of future semiconductor demand.

Energy: The New Constraint

The biggest constraint on AI growth is no longer semiconductor supply; it's energy. Next-generation AI data centers consume power on the scale of several gigawatts. One gigawatt is equivalent to a large nuclear power plant, roughly the electricity consumption of one million people.

Microsoft CFO Amy Hood has noted that the biggest constraint on AI infrastructure is no longer semiconductors. Securing power and transitioning to green energy is becoming the new battleground in AI competition.

Japan's "Third Path"

AI is not a bubble. The continuous growth in computing demand proves this. The question is how to ride this massive wave.

Japan leads the world in manufacturing equipment and materials while lagging significantly in cutting-edge logic semiconductor manufacturing. The country is attempting to fill this gap with two pillars: Rapidus and TSMC Kumamoto.

Challenges remain substantial. Talent shortages are severe, and competition with TSMC and Samsung for skilled workers is fierce. Rapidus's 2027 mass production target faces significant technical hurdles.

Still, amid rising geopolitical risks, Japan's semiconductor revival holds significant meaning for a world seeking alternatives to Taiwan concentration. Japan is being called upon to pursue a "third path," leveraging strengths in equipment and materials while securing a meaningful presence in advanced manufacturing.

In Japan, discussions like these are unfolding around AI and semiconductors. What initiatives are progressing in your country regarding AI industry and semiconductor policy? We'd love to hear about government support measures, industry trends, and the level of public interest in your region.

References