In June 2026, a single trading session erased over $1.3 trillion from global semiconductor stocks. Headlines declared the AI bubble had burst. Michael Burry — the investor who famously called the 2008 housing collapse — was short. The phrase “semiconductor bubble” was trending across financial media.

The same week, largely without fanfare, a historic threshold was crossed: data centers are on track to overtake smartphones as the single largest consumer of NAND flash memory in the world. For the first time in the industry’s history.

These two facts are connected. But not in the way most investors think.

The Debate Is Asking the Wrong Question

The semiconductor bubble argument is, at its core, a question about NVIDIA’s valuation. Did a stock that rose 450% in two years go too far? That is a legitimate debate. NVIDIA briefly touched a $5 trillion market capitalization in early 2026. Burry called it “the final months of the 1999 dot-com bubble.” JPMorgan countered in December 2025 that AI investment is “linked to actual enterprise revenue” and does not meet classic bubble criteria. TSMC’s CEO flatly denied the thesis, pointing to chip supply shortages expected to last years.

Both sides are arguing about compute demand — GPU clusters, AI training runs, inference capacity. That is the visible debate, because NVIDIA is public and its numbers are auditable. ChatGPT is not listed. Claude is not listed. The AI companies where the hype visibly concentrates are privately held. So the market anchors its bubble argument to the one place it can see: semiconductor stocks.

This is the wrong focal point.

A New Continent Was Discovered

When the Americas were discovered, the first fortunes went to shipbuilders and navigation companies. The ships were the enabling infrastructure. But whether any individual voyage was profitable or not, the continent did not disappear. Demand shifted — from ships to ports, roads, and warehouses.

Artificial intelligence has opened the equivalent of a new digital continent. The first phase of infrastructure build-out favored compute: GPU clusters, training facilities, inference farms. This is what NVIDIA supplies. This is what the bubble debate is about.

But the continent is not going away. And what a continent needs, after the ships arrive, is somewhere to put things.

Demand Is Migrating to Storage

The structural shift is already visible in the data.

In 2026, data centers are set to surpass mobile phones as the single largest end-market for NAND flash memory — a first in the history of the industry. Enterprise SSD demand is growing at 41% year-over-year. NAND demand is expanding at 20-22% annually while supply bit growth is constrained to 15-17%, creating a structural shortage that TrendForce expects to persist through 2027. Nearline HDD supply is critically short, with lead times extended across major vendors.

The reason is not complicated. Every piece of AI-generated content must be stored — permanently. And unlike GPU cycles that can be powered down, stored data persists. It accumulates.

One Slide Deck Per Day. Now Ten Per Hour.

In 2024, building a presentation took half a day. In 2026, the same person asks an AI and gets ten versions in an hour. One gets sent. Nine get saved.

Consider what generative AI has done to individual data creation. Before 2023, one person wrote one email, built one presentation, produced one draft. Now the same person prompts an AI for three presentation variants, refines each, saves all versions, and sends the best one. GitHub Copilot writes 46% of all code produced by its 20 million users — generating far more lines per human-hour than any developer would alone. 34 million AI-generated images are produced every day. Each one is stored. None disappear.

This data multiplier effect was not captured in pre-2025 demand forecasts. Existing models were built on historical correlations between human activity levels and data generation. When AI detaches data output from human effort — producing more per person-hour without increasing the hours — those correlations break. The new continent is larger than the old maps suggested.

The Evidence You Can Observe

Abstract data volume forecasts are hard to verify. But observable proxies exist.

Power consumption is the clearest signal. The International Energy Agency projected in early 2024 that global data centers, AI, and cryptocurrency combined would consume over 1,000 terawatt-hours of electricity by 2026 — roughly equivalent to Japan’s entire annual electricity consumption, up from 460 TWh in 2022. Data center power demand grew 17% in 2025, against global electricity demand growth of just 3%.

Subsea cable investment tells the same story. Capital committed to new submarine cable projects is expected to reach $13 billion between 2025 and 2027 — nearly double the previous three-year period. Meta completed its 45,000-kilometer 2Africa cable. AWS launched its first independently commissioned submarine cable. Google announced a new trans-oceanic route connecting Australia and Thailand. Physical infrastructure of this scale does not get built speculatively.

The three largest hyperscalers — AWS, Microsoft Azure, and Google Cloud — combined to commit more than $500 billion in infrastructure capital expenditure for fiscal year 2026. These are not paper valuations. They are concrete pours.

What Is the Real Risk?

The real risk is timing, not existence. Semiconductor fabrication plants take two to three years to build. If the capex decisions made in 2024 and 2025 result in simultaneous capacity additions in 2027 and 2028, an inventory correction cycle could follow. That risk is real — Marumae (TSE:6264), the Kagoshima precision parts maker whose order book functions as a real-time read on semiconductor equipment demand, has seen its FY2026 revenue track toward nearly tripling from the FY2024 trough. Cycles do not move in a straight line.

But the floor is materially higher than in previous downturns. Storage demand — unlike compute demand — does not decommission itself. The data generated in the first wave of AI infrastructure build-out stays on disk. The new continent keeps needing warehouses.

Japan’s Quiet Beneficiaries

For investors in Japanese equities, the migration from compute to storage has specific implications. Kioxia (TSE:285A), the world’s largest NAND-focused chipmaker by volume, is a direct beneficiary of the structural shift. HOYA (TSE:7741), whose glass substrates underpin HDD production and which holds the leading global market share in that segment, benefits as nearline HDD demand outstrips supply. Marumae (TSE:6264) supplies precision components to equipment makers on both sides of the equation — compute and memory alike.

These companies do not appear in the same headlines as NVIDIA. They are outside the bubble debate. That may be precisely why they are worth watching.

Conclusion

The semiconductor bubble argument asks whether the ships that carried explorers to the new continent were overpriced. That is worth debating. But the continent itself is not in question. The data being generated at scale — emails, images, code, documents, and synthetic training sets — must be stored somewhere. The infrastructure to store it is being built. The demand does not turn off.

The question is not whether AI is a bubble. The question is whether the new continent will disappear.

It will not. And the data born on the new continent flows back to the warehouses of the old one. AI-generated content lands on enterprise storage systems. Training datasets accumulate in existing cloud infrastructure. Inference outputs return to on-premise servers. The demand wave is hitting both new and legacy infrastructure at the same time. Both continents are being washed by the same tide.

Source: IEA Electricity 2024 | TrendForce NAND Q1 2026 | 日本語版

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