This is Part 5, the conclusion of a five-part series on semiconductor demand in the age of AI warfare. Start at Part 1 or Part 4 for drone chip supply chains.

The Question Everyone Is Asking

Is this the dot-com bubble?

The parallel is appealing. A transformative technology attracts massive capital. Valuations of the enabling infrastructure companies reach historically extreme levels. A narrative of “everything changes” justifies multiples that require decades of compounding to rationalize. Then the music stops.

NVIDIA’s market cap at its peak was above $3 trillion — larger than every listed Japanese company combined. The comparison to Cisco in 1999, which briefly became the world’s most valuable company on networking infrastructure demand, is not idle. Cisco’s stock did not recover its 2000 peak for over twenty years.

The honest answer is that we do not know whether semiconductor stocks in 2026 are in a bubble. We can, however, distinguish between the types of demand that are durable and the types that are not.

The Bubble Test: Three Questions

Question 1: Is the application layer monetizing?

The dot-com crash happened because the infrastructure (bandwidth, servers) was built before the applications that would use it existed. The valuations priced in an application wave that was real but arrived ten years later than the infrastructure buildout implied.

In AI, the application question is partially answered. ChatGPT has 200 million weekly users. Enterprise software is being retooled with AI copilot functionality at every major vendor. AI-assisted coding (Copilot, Cursor, Claude) has demonstrably increased developer productivity in measurable ways. The applications exist and are generating revenue.

The remaining question is whether they monetize at the scale required to justify $600 billion in annual infrastructure capex. This is genuinely uncertain. The case for “yes” is large and growing; it is not yet proven.

Question 2: Is the demand multi-vector or single-source?

The dot-com bubble was a single-vector demand story: internet adoption. When internet adoption growth slowed, the infrastructure overcapacity was exposed.

This series has argued that 2026 semiconductor demand is multi-vector:

  • AI training and inference (GPU cycle)
  • Memory and storage buildout (HBM, NAND — the Kioxia signal)
  • EV power electronics (SiC, IGBTs)
  • Military drone production (MCUs, sensors, RF, power management)
  • Industrial automation and robotics (MCUs, sensors, motion control)

The semiconductor companies best positioned are those with broad exposure across multiple demand vectors — not pure-play NVIDIA derivatives. Japanese equipment and materials companies (TEL, Advantest, Shin-Etsu) are structurally diversified across end-markets in a way that a stock like Marvell or Arm is not.

Question 3: Is the geopolitical driver durable?

The CHIPS Act, the Japan-TSMC JV at Kumamoto, the EU Chips Act, South Korea’s K-semiconductor strategy — these are government commitments to semiconductor domestic production that are not driven by market economics. They are driven by the lesson of COVID (single-point-of-failure supply chains) and the US-China decoupling (semiconductor access as a geopolitical weapon).

Government-mandated capex is slower to respond to demand cycles than private capital. A recession does not defund a national security program the way it defunds a startup’s cloud budget. This structural floor on semiconductor equipment demand is not present in most other technology sectors.

The Drone Thesis Revisited: Bull Case and Risks

The drone demand argument made in Parts 3 and 4 has a bull case and several risks worth stating explicitly.

Bull case:

  • Military drone production is in the early stages of a multi-year ramp
  • The chip content per drone ($50-150) at scale (millions of units annually, globally) is a material semiconductor demand increment
  • The same components (MCUs, image sensors, RF components) that serve military drones also serve commercial drones, industrial robots, and autonomous vehicles — demand vectors reinforce
  • Japan’s supply chain exposure is at the component and sensor level, which is less geopolitically constrained than advanced logic

Risks:

  • Export control changes could restrict sensor/MCU exports to military customers
  • Consumer drone manufacturers (primarily Chinese) may vertically integrate component production, reducing reliance on Japanese suppliers
  • Military drone technology may evolve toward custom ASICs (like AI accelerators evolved away from GPUs in some applications), eroding commodity component demand
  • Scale estimates are uncertain — production rates are not publicly audited

The honest position is that drone demand is an emerging tailwind, not a confirmed megatrend. Its materiality for companies like Murata, Rohm, and Renesas depends on how rapidly military drone production programs scale globally, which is a function of geopolitics that is inherently unpredictable.

A Framework for Japanese Semiconductor Exposure

After five articles, the investment framework can be stated simply:

Layer 1: Structural demand (hold for the cycle) Tokyo Electron, Advantest, Shin-Etsu — upstream equipment and materials that benefit from any expansion of global semiconductor production. Multiple demand drivers; relatively insulated from NVIDIA-specific narrative.

Layer 2: Multi-vector component demand (evaluate on valuation) Murata, TDK, Rohm — passive components and power devices that serve AI infrastructure, EV, and defense applications simultaneously. Cyclical exposure remains, but demand floor is higher than historical semiconductor cycles due to EV transition and defense ramp.

Layer 3: Emerging drone/defense exposure (early optionality) Sony Semiconductor (via Sony Group), Renesas — image sensors and MCUs with nascent but growing military end-market exposure. This is not priced into consensus estimates. Monitor for order trends and any company disclosure of defense/aerospace revenue growth.

Layer 4: NVIDIA derivative plays (highest volatility) Companies with concentrated exposure to NVIDIA’s GPU fab supply chain — TSMC vendors, HBM suppliers — carry the full narrative risk of the NVIDIA bull-bear debate. Japanese investors with indirect exposure through equity funds should be aware of this concentration.

Conclusion

The question “is the NVIDIA era over?” turns out to be less interesting than it first appears.

Whether NVIDIA’s next earnings beat or miss consensus by 10%, whether hyperscaler capex accelerates or decelerates in the next two quarters — these are questions for traders, not for investors building positions in a technology transition that will unfold over a decade.

The structurally durable question is this: the world has decided that semiconductors are strategic infrastructure. Governments are spending to ensure domestic supply chains. Militaries are rearming with electronics-intensive systems. Enterprises are rebuilding their software stack around AI. All of this requires semiconductors.

Japan’s upstream semiconductor supply chain — built over forty years of patient specialization — sits at the foundation of that transition. The companies discussed in this series are not exciting in the way NVIDIA is exciting. They are essential in a way NVIDIA depends on.

That is a different investment case. It is, arguably, a more durable one.

This concludes the five-part semiconductor demand series. Part 1: The NVIDIA Bull-Bear Debate | Part 3: The Drone Army Revolution

Source: Company IR materials; defense industry public data | 日本語版

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