Key Points
・On July 17, 2026, Kioxia Holdings shares hit their daily-limit floor at 52,110 yen, down more than 16 percent, landing below half of the all-time high of 112,700 yen the stock had set less than a month earlier. In mid-June the company had briefly overtaken Toyota as Japan’s most valuable listed firm.
・The business itself was not faltering. Kioxia’s own forecast still pointed to April-June operating profit roughly 30 times the year-earlier level, with production sold out for the year. The selloff was triggered by a US patent verdict, a new Chinese AI model, and rising supply, none of which touch this quarter’s earnings.
・A stock that kept climbing while profit multiplied by 30 was pricing in more than one strong quarter, it was pricing in a story that the memory industry’s decades-old boom-bust cycle had ended for good. That story is what came under challenge this week.
News
On July 17, 2026, Kioxia Holdings shares opened for sale on the Tokyo Stock Exchange with no buyers willing to match the asking price, and by around 9:30am the stock had fallen to its daily-limit floor, down 10,000 yen, or 16.1 percent, to close at 52,110 yen. That left the stock below half of the all-time high of 112,700 yen it had reached on June 22. Kioxia’s market capitalization fell below 30 trillion yen, dropping the company from Japan’s largest listed firm by value to fifth place, overtaken by Tokyo Electron. In the roughly one month since the June 22 peak, an estimated 31 trillion yen in market value had evaporated.
The immediate trigger was a US patent ruling. On July 16, a federal jury in Texas found that Kioxia had infringed a patent held by the satellite communications company Viasat, covering technology that corrects data errors on memory chips while reducing power consumption, and ordered Kioxia to pay approximately $229 million (roughly 37 billion yen) in damages. Kioxia said on July 17 that “Viasat’s claims and the jury’s decision are wholly unacceptable” and that it planned to pursue legal options, including an appeal.
The selling was not confined to Kioxia. Weakness in US semiconductor stocks the previous session, combined with fresh wariness after Chinese AI startup Moonshot AI unveiled a new model called Kimi K3 on July 16, added to the pressure. Japan’s Nikkei average fell 2,694.42 yen on July 17, its fifth-largest point drop on record, closing at 64,141.12, with selling concentrated in AI and data-center-related names. Attention has since turned to how far the unwinding of Kioxia’s record margin-buying balance will run, and whether this pattern, strong earnings paired with a collapsing share price, spreads further across AI-linked stocks.
Background
Toshiba’s Fire Sale Becomes Japan’s Most Valuable Stock, Then Loses Half Its Value in a Month
Kioxia is what remains of Toshiba’s memory chip business, spun off during one of Japanese industry’s more dramatic corporate crises. In 2017, Toshiba, pushed toward insolvency by massive losses at its US nuclear power business in the wake of an earlier accounting scandal, carved out its semiconductor memory division as Toshiba Memory and sold it to a consortium led by the American private equity firm Bain Capital. The unit was renamed Kioxia in 2019. The sequence since then:
– December 18, 2024: Kioxia lists on the Tokyo Stock Exchange’s Prime Market. The IPO price is set at 1,455 yen, but the stock opens at 1,440 yen, a rare case of a Japanese IPO opening below its offer price. Market capitalization is roughly 784.3 billion yen.
– 2025: AI data-center demand improves memory market conditions, and the stock begins to climb.
– June 12, 2026: Market capitalization reaches roughly 44 trillion yen, overtaking Toyota Motor to become Japan’s most valuable listed company.
– June 22: The stock hits an all-time high of 112,700 yen, about 77 times the IPO price. Market capitalization briefly tops 60 trillion yen.
– June 26: Reports that OpenAI is considering delaying a planned IPO send the stock down 12 percent.
– July 17: The stock hits its daily-limit floor at 52,110 yen. Market capitalization falls below 30 trillion yen and Kioxia drops to fifth place among Japanese listed companies.
Kioxia’s core product is NAND flash memory, used in smartphones and solid-state drives (SSDs) for long-term data storage, and the company is one of the world’s largest producers, with major fabrication plants in Yokkaichi and Kitakami, Japan. Before its IPO, Kioxia was sometimes described as a symbol of Toshiba’s breakup and decline. Eighteen months after listing, it briefly became Japan’s most valuable company, then lost half its value within a month. That swing is the shape of the story.
From Work Desk to Warehouse: Why NAND Became an AI Stock
Memory chips split broadly into two categories, and NAND was a late arrival to the AI boom. DRAM functions as a computer’s work desk, holding data temporarily while it is being processed, while NAND functions as a warehouse, storing data for the long term. Early in the AI boom, the spotlight fell on the work-desk side: high-bandwidth memory (HBM), a specialized form of DRAM wired directly to GPUs, was treated as a scarce component sitting next to AI’s core processing power and its price surged first. NAND, with many producers and a reputation as a competitive commodity, drew far less attention.
What changed was the shift in how AI is used, from training models to running them at scale for millions of users, a phase generally called inference. Conversation histories and generated outputs need to be stored, and data centers responded by buying large-capacity SSDs, built from NAND, in bulk. Demand tightened quickly, and reported NAND prices rose sharply as a result.
The scale of that shift shows up in Kioxia’s own numbers. The company’s forecast for the April-June quarter of its fiscal year projects sales of 1.75 trillion yen and operating profit of 1.298 trillion yen, an operating margin of 74 percent and roughly 30 times the year-earlier profit. Kioxia’s production capacity for 2026 has reportedly sold out entirely. That combination of numbers is what pushed the company close to the center of the “AI stock” category in investors’ minds.
The Silicon Cycle, and Why Buying on Margin Amplifies It
Memory chips have a boom-bust pattern well known enough in the industry to have its own name, the silicon cycle. Prices rise, producers respond by investing heavily in new capacity, that capacity comes online two to three years later, supply outpaces demand, prices collapse, producers cut output and post losses, supply tightens again, and the cycle repeats. The most recent trough came in 2022-23, when falling memory prices pushed Kioxia itself into production cuts and losses.
This time, the supply side had a specific gap. Over the past two years, Samsung Electronics and SK Hynix, the two Korean memory giants, had actually reduced NAND wafer investment, redirecting capital toward the higher-margin HBM and DRAM side of the business. When AI-driven demand for storage arrived suddenly, the industry’s warehouse-building capacity had been thinning rather than growing, and the resulting price spike was correspondingly sharp. Signs of a supply response are now emerging: SK Hynix announced a roughly $51 billion NAND investment plan in July, China’s Yangtze Memory Technologies (YMTC) reportedly grew revenue more than fourfold in a year, and China’s ChangXin Memory Technologies (CXMT), the world’s fourth-largest DRAM maker, opened subscriptions on July 16 for a Shanghai STAR Market IPO seeking roughly 29.5 billion yuan (about 600 billion yen).
Japan adds its own amplifier on top of this industrial cycle: margin trading, in which an investor puts up cash or existing shares as collateral and a brokerage lends additional funds to buy more stock than the investor’s own capital would allow. If the price falls, the loan does not shrink, and once the collateral value drops below a set threshold, the brokerage issues a margin call or, if it goes unmet, sells the position outright regardless of the investor’s wishes. The outstanding balance of shares bought this way, known as the margin buying balance, is tracked across the Tokyo Stock Exchange as a whole. That balance topped 7 trillion yen for the first time in early July, and roughly 60 percent of the prior week’s increase in that balance was concentrated in Kioxia alone, a warning sign regulators and analysts had flagged before the crash.
Analysis
The Stock Priced In the End of the Cycle, Not Just This Quarter’s Profit
A company with 30 times its prior year’s profit losing half its market value in a month is only strange if the stock price was tracking profit. It was not.
The underlying demand is real. A 74 percent operating margin means Kioxia is selling everything it can produce at prices that are not eroding. Production for 2026 is reportedly sold out, and Taiwan Semiconductor Manufacturing Company (TSMC), the world’s largest contract chipmaker, reported on July 16 that its own April-June net profit rose 77 percent to a record, driven by AI chip demand. Forecasters expect both DRAM and NAND contract prices to post further double-digit gains in the July-September quarter. Judged purely on business fundamentals, this week’s selling looks like an overreaction.
What actually broke was the premise behind a valuation 77 times the IPO price. In a normal silicon cycle, the market discounts a memory maker’s peak-year profits, because everyone knows the peak does not last. Kioxia’s stock had been priced as though that discount no longer applied, on the theory that AI demand had permanently changed the structure of the memory business and the old cycle was not coming back. Once that theory is in place, it is not this quarter’s earnings that move the stock, it is evidence against the theory: a competitor’s new capacity, a rival’s faster-than-expected model, a cheaper way to run AI. None of it touches current profit, and all of it cuts directly at the premise. That the roughly 5 trillion yen in market value lost on the day of the patent verdict was more than a hundred times the size of the damages award itself is a sign that the selling was never really about the lawsuit.
The theory has not been disproven, only challenged. If demand keeps outrunning new capacity as it comes online over the next two to three years, the “this time is different” story regains the upper hand.
The Tremors Are Starting at the Edge of the AI Chain, Not the Center
Lined up together, the past week’s shocks in Asian markets share a pattern: they are hitting the edges of the AI supply chain, not its center. A closely related episode in South Korea days earlier saw a reversal in debt-fueled buying concentrated in two memory stocks; the Kioxia move is a memory-specialist stock losing half its value. Both are the part of the AI buildout most exposed to the traditional industry cycle, and therefore most dependent on the “the cycle is over” story holding up. Nvidia and OpenAI, sitting closer to the center of the AI trade, have not seen comparable swings.
It is telling that the June 26 plunge in Kioxia shares was triggered by reports that OpenAI, a private company with no direct link to Kioxia’s factories, was weighing whether to delay its own IPO. A piece of news with nothing to do with the utilization rate of Kioxia’s Yokkaichi plant moved the company’s market value by trillions of yen in a single day. The chain of AI infrastructure investment has built a structure in which a component maker at the far end of the supply chain can swing on the funding calendar of a single company near the center. The longer and more tightly wound that chain becomes, the harder a small movement at the center shakes the edges.
Japan’s own version of that fragility shows up in the margin buying balance described above. The Tokyo Stock Exchange’s total margin buying balance passed 7 trillion yen for the first time in early July, with about 60 percent of the week’s increase concentrated in Kioxia alone. Analysts had warned days before the crash that a high-volatility stock carrying a large margin balance is exposed to forced selling once prices turn down. The specific instrument was different from the leveraged ETFs at the center of South Korea’s earlier turmoil, but the underlying mechanism, borrowed money amplifying a downturn once it starts, was the same.
China Just Sped Up the Clock on Both Sides of the Market
Three China-related developments landing in the same week was not a coincidence, they were evidence against the “cycle is over” story arriving simultaneously from the demand side and the supply side.
On the demand side, Moonshot AI’s Kimi K3, unveiled on July 16, claimed performance close to leading US models while using markedly less computing power to train, drawing comparisons to last year’s DeepSeek moment. The premise underlying much of the data-center buildout, and by extension the memory demand story, is that improving AI performance requires ever-larger amounts of compute. If capable models can be built more cheaply, the assumptions behind projected data-center and memory demand come into question, which is a large part of why AI-related stocks broadly sold off on July 17.
On the supply side, CXMT’s IPO and YMTC’s rapid growth point in the same direction. Both are backed by Chinese state capital and are expanding capacity, and an IPO raise, in particular, is capital that eventually becomes new factories. Combined with SK Hynix’s own NAND investment plans, these developments put a visible expiration date on the current tight-supply story that has been supporting prices.
China is not a one-directional factor here, though. A cheaper way to run AI models could just as easily expand who uses them and how much data they generate and store, increasing rather than decreasing memory demand over time. The same argument played out after the DeepSeek episode last year, and global AI investment kept expanding afterward. Cheaper AI can read as evidence the demand story is ending, or as evidence its base is about to widen. This week, the market chose the first reading.
Infrastructure Outlives Its Investors, as Fiber-Optic Cable Proved Once Before
Being right about demand and being rewarded as an investor are two different things, a separation the US fiber-optic buildout of the late 1990s demonstrated clearly.
Believing internet data traffic would explode, US telecom firms spent as much as an estimated $2 trillion over roughly five years laying more than 80 million miles of fiber-optic cable. The demand forecast was correct: data volumes kept rising and much of that cable eventually carried the video streaming and cloud computing of later decades. Yet by 2001, only an estimated 5 percent of the fiber laid was actually in use, telecom stocks collapsed, and several major industry names went bankrupt. Demand was real, the infrastructure endured as a societal asset, and investors who arrived too early were not made whole regardless.
Applied to AI and memory chips, data centers and fabrication plants persist whatever happens to their owners’ share prices. Kioxia’s Yokkaichi production lines kept running the same day the stock lost half its value. The gap is between the timeline on which demand materializes and the timeline the stock price had already borrowed against.
A few markers from this pattern are likely to remain useful even after this week’s headlines fade: a period where earnings stay strong but the stock falls is usually a sign the market is repricing the story, not the business; a wave of new capacity announcements or capital raises signals that supply is catching up to that story; and how sharply a downstream component maker’s stock reacts to financing news from a company near the center of the AI trade is a rough gauge of how taut that financing chain has become. All three showed up in the same week around Kioxia’s July 2026 crash.
Conclusion
What happened to Kioxia on July 17, 2026 was not evidence that AI demand is an illusion, nor that Japan’s semiconductor recovery is over. A company can post 30 times its prior year’s profit and still lose half its market value, because a stock price carries two layers, one tracking the business and one tracking a story about the business, and this week it was the story layer that came off. Whether that particular story, that AI has permanently ended the memory industry’s boom-bust cycle, turns out to be right is still an open question.
If demand keeps outrunning the new capacity now being announced, the story regains its footing. If the new factories arrive first, the industry’s older cycle runs its familiar course one more time. What is already clear is that the verdict will not come from a quarterly earnings report. It will come from financing announcements at unlisted companies, capacity plans from competitors, and product launches out of China, all of it sitting outside the business itself, and all of it, this week, pricing Kioxia’s stock more than its record profit did.
Reference Links
- キオクシアの株価、最高値の半値に 時価総額30兆円割り込む｜日本経済新聞
- キオクシア、約370億円の支払い命令に「到底容認できない」 控訴含む法的手段へ 米特許訴訟で｜ITmedia NEWS
- キオクシア、4～6月営業利益1.3兆円予想 「東芝」DRAM撤退がもたらした果実｜日経ビジネス
- 信用買い残が過去最高の7兆円超え キオクシアなど短期売買｜日本経済新聞
- China’s open-weight Kimi model stuns AI world with frontier-level results｜Axios
- China’s ChangXin launches $4.3B memory IPO into an AI-driven chip boom｜Cryptopolitan
- Boom and Bust in Telecommunications｜Federal Reserve Bank of Richmond
