This is Part 1 of a five-part series on semiconductor demand in the age of AI warfare.

The Setup

In late April 2026, NVIDIA briefly touched all-time highs above $200 per share. Within days, the stock pulled back — not on any specific news, but on the weight of its own narrative: the fear that the greatest infrastructure build-out in tech history might already be slowing.

The bear case makes intuitive sense. Hyperscalers are building their own chips. GPU utilization at many data centers reportedly runs below 30%. Microsoft paused some data center lease commitments. A financial press that spent two years writing about NVIDIA’s unstoppable growth has begun writing about “the inevitable peak.”

The bull case is equally coherent. The world’s largest companies have collectively committed over $600 billion in AI capex for 2025 and 2026. A new computing paradigm — inference-time scaling — is emerging that may require orders of magnitude more compute than training did. Jensen Huang has projected the global compute market will reach $1 trillion.

Both cases are based on real evidence. This article examines each honestly.

The Bear Case: Three Structural Concerns

1. GPU utilization is low.

The bear case often starts here: if utilization at major data centers is running at 5-30%, why would hyperscalers keep buying? The argument implies that the initial buying wave was driven by competitive panic — “we must have AI capabilities or fall behind” — rather than actual economic demand. Fear-of-missing-out, measured in racks and megawatts.

This is not an unreasonable reading of 2023-2024. Companies bought compute before they had workloads to run on it. The installed base now exceeds the near-term application layer.

2. Hyperscalers are building their own silicon.

Google’s TPU, Amazon’s Trainium, Meta’s MTIA, Microsoft’s Maia — every major cloud company is designing chips to run AI inference at lower cost than NVIDIA’s H100/H200 can deliver. The argument is that NVIDIA’s 80%+ gross margins are a structural vulnerability: they give customers enormous economic incentive to verticalize.

The precedent is real. Apple moved away from Intel. Amazon designed the Graviton server chips that now power significant portions of AWS. When the margin differential is large enough, customers become competitors.

3. The capex mix is shifting.

Of the committed hyperscaler capex, a growing share is going to HBM memory, NAND, networking, and power infrastructure — not GPUs. NVIDIA captures a declining fraction of the AI infrastructure dollar as the stack matures. This is the normal economics of platform commoditization.

The Bull Case: Three Structural Supports

1. Inference-time compute changes the math.

Models like OpenAI’s o3 and DeepSeek R1 use “extended thinking” — they run additional inference passes to reason through complex problems before returning an answer. A single query to o3 may consume as much compute as thousands of standard queries.

This matters because the prevailing bear logic assumes AI compute demand is proportional to training runs — and training runs have a ceiling (you only train a model once). But inference does not. If every production workload shifts to extended thinking, the compute curve bends upward again. The question of how much compute AI needs may have been systematically underestimated.

2. The enterprise cycle hasn’t started.

ChatGPT launched in late 2022. Three and a half years later, enterprise AI deployment remains nascent. The hyperscalers built for an application wave that is still materializing. The current capex pause — to the extent it is real — may simply reflect the lag between infrastructure build-out and the applications that monetize it. Historical tech buildouts (broadband, mobile networks, cloud) all followed this pattern.

3. Geopolitical demand is sovereign, not economic.

The United States has restricted chip exports to China. This has done two things: it made every other government acutely aware that access to advanced semiconductors is a geopolitical asset, and it triggered aggressive domestic AI investment programs across Southeast Asia, the Middle East, and Europe. Saudi Arabia and the UAE are buying Blackwell clusters at scale, not because the ROI pencils out today, but because they have decided that AI sovereignty is a national security priority.

This demand does not follow the utilization logic of a McKinsey ROI model. It follows the logic of arms acquisition.

The Third Thesis: A Hype Cycle, Not a Demand Cycle

There is a third interpretation that the bull/bear framing misses.

The training phase for frontier AI models is largely complete among the top labs. GPT-5, Gemini 2.0, Claude 4 — the major architectural leaps of the first generation have been made. The companies that needed to train at scale have trained. The first wave of demand was real; it was front-loaded into 2023-2025.

What follows is not collapse. It is the trough of a hype cycle — the period when the technology is real but the commercial applications are still incomplete, when the builders are exhausted, and when the narrative turns from euphoria to skepticism.

This is the phase before the productivity deployment wave. Enterprise software running on AI, industrial process automation, healthcare diagnostics — these are the applications that will drive the next demand wave. They run inference, not training. They require different chips in different configurations.

The question for investors is not whether AI demand continues. It is what that demand looks like in the inference era, and which companies capture it.

What This Means for Japan

Japanese semiconductor exposure is overwhelmingly upstream: equipment (Tokyo Electron, Lasertec, Advantest), materials (Shin-Etsu, SUMCO), and components (Murata, TDK, Rohm).

The upstream thesis is structurally agnostic to which AI company wins. Tokyo Electron benefits whether NVIDIA, AMD, or Google’s TPU program drives the next capacity expansion — all three require the same etching, CVD, and testing equipment. Shin-Etsu supplies the silicon wafers that every chip requires.

This is both a comfort and a ceiling. Upstream players do not capture the full economics of AI demand, but they are insulated from NVIDIA-specific narrative risk.

The more interesting question — explored in the remaining parts of this series — is whether AI chips are even the right frame. As the semiconductor market enters its next cycle, the incremental demand may be coming from somewhere the financial press has barely discussed: the battlefield.

Part 2 — “Beyond GPUs: Why Semiconductor Demand Doesn’t End Even If NVIDIA Slows” — is available here.

Source: NVIDIA investor relations and public earnings materials | 日本語版

Disclaimer | This article is for informational purposes only and does not constitute investment advice.