Samsung's semiconductor division has seen a dramatic increase in profits, with Micron's quarterly revenue exceeding $40 billion and gross margins climbing to 85%.
However, just five weeks after these reports, the industry index faced a significant downturn: from July 24 to July 29, the top twenty semiconductor manufacturers lost $1.3 trillion in market capitalization.
Typically, record earnings drive stock prices up, but this time, the opposite occurred. On June 30, renowned investor Michael Burry, famously known as the protagonist of "The Big Short," revealed that he had taken short positions against the semiconductor ETF SOXX, as well as Nvidia and Applied Materials. Shortly after, he added Micron to his list, labeling it the "most cyclical" company.
In this article, we delve into Burry's reasoning behind his new short positions, why the unprecedented shortage isn't supporting stock prices, and the differing forecasts from analysts.
More Than the Dot-Com Era
In 2025, the largest American tech companies invested around $450 billion in infrastructure. This year, five major hyperscalers—Alphabet, Amazon, Meta, Microsoft, and Oracle—expect to spend about $730 billion.
The Bank for International Settlements (BIS) noted in its annual report that this investment cycle parallels historical booms like the British railway craze of the 1840s, the canal-building boom of the 1830s, the electrification of the 1920s, and the dot-com bubble of the late 1990s. The common thread in all these episodes was that real technological breakthroughs attracted more investment than they generated in profits, often leading to sharp declines in investment activity and recessions.
This current cycle is growing faster than any of those previously mentioned.
Source: Bank for International Settlements.Additionally, the BIS estimated that the sector needs to achieve around $600 billion in real annual revenues from end users for investments to be justifiable. Currently, artificial intelligence (AI) does not generate that level of revenue.
This backdrop contributed to the sell-off in July.
Volatility Beyond Bitcoin
On June 22, the Philadelphia Semiconductor index closed at a record high. Just three days later, Micron's stock hit a new peak at $1213.37 per share.
At that time, Reuters reported, citing BofA Global Research, that the bubble risk indicator for the PHLX Semiconductor index reached 0.91 out of 1, signaling potential overheating. The Technology Select Sector index was at 0.82. A significant drop soon followed.
Bubble risk indicators for the tech sector. Source: Reuters/BofA Global Research.On July 27, memory chip manufacturer CXMT debuted on the Shanghai Stock Exchange. The company's shares from Hefei surged 466% on their first trading day, giving it a market capitalization of approximately $487 billion. The IPO raised $8.6 billion, marking the largest in the history of China's semiconductor sector.
However, this was followed by a decline in stock prices within the sector, which analysts attributed to expectations of an impending increase in global supply. The Asian IT index MSCI fell by 4.7%, while Micron and SanDisk shares dropped by 5% and 12%, respectively, and SK Hynix's stock in Seoul decreased by 8.5%.
Interestingly, the companies that suffered the most were not the underperformers or the smaller players, but instead, those that had benefited the most from the acute shortage of key components.
In this context, one detail from CXMT's prospectus stands out. The company disclosed plans to allocate 29.5 billion yuan of the raised funds for upgrading DRAM technology, developing next-generation memory solutions, and modernizing silicon wafer production equipment. Notably, there were no projects mentioned related to HBM, the most promising segment of the market. The company did not reveal how the remaining approximately 28 billion yuan would be spent.
The acute shortage had allowed Micron and SK Hynix to maintain high margins. CXMT's market debut signaled that new capacities would eventually emerge, prompting the market to quickly factor this expectation into stock prices.
This scenario began to unfold in the following days. On July 28, SK Hynix reported a record operating profit of 60.54 trillion won ($41.25 billion), marking a year-on-year increase of 557%. However, the result fell short of the consensus forecast of 64 trillion won. The company's announcement of plans to increase capital expenditures to $31 billion led to a 15% drop in its stock price.
By the end of July, the South Korean KOSPI index fell by 16% over just two sessions, triggering trading halts on both days.
On July 28, Bitcoin mirrored the Korean market, losing about 2.7% and dropping below $63,000. The following session saw it gain
The impact of forced liquidations due to leveraged positions on this decline remains unclear, as neither exchanges nor management firms publish such breakdowns.
Subsequently, stock prices rebounded from their local lows. On July 30, Micron surged about 18%, and the next session opened with a further 6% increase after the CEOs of Apple and Amazon warned investors about steep increases in memory prices. However, trading concluded with a 5.9% drop, settling at $823.03. This marked a 32% decline from its June record over six weeks.
The sell-off quickly subsided: on August 3, stocks closed up to $829.50, while the Philadelphia Semiconductor index remained over 20% below its June peak. Burry's analysis suggests that this is the type of decline he anticipated when explaining his stance on Micron, evidenced by the phenomenon of "memflation."
Chronicle of volatility in the semiconductor sector. Source: ForkLog.A Curious Term
The term "memflation" was coined by Gartner. In an April forecast, the agency projected that the semiconductor industry would generate over $1.3 trillion in revenue, marking a 64% increase. This could represent the third consecutive year of double-digit growth for the sector and its strongest performance in two decades.
However, the majority of this increase is driven by price increases rather than heightened demand. Gartner estimates that the average annual prices for DRAM will rise by 125%, while NAND prices will increase by 234%, leading to a tripling of revenue in the memory segment.
The distinction between these two types of chips is crucial. DRAM is used for HBM, while NAND refers to flash storage used in laptops, smartphones, and servers.
The second type is seeing the steepest price increases, which is not directly related to AI training. Factories have shifted excess capacity to produce components for data centers, leading to retail price surges.
Gartner's senior analyst Rajiv Rajput described the potential consequences: "Memflation" could eliminate or delay all demand outside of AI until 2028, with prices unlikely to decrease significantly before the end of 2027.
Additionally, the final recommendation in the release advises IT directors to be cautious when signing contracts with unfavorable terms that extend beyond 2027. The forecast and caution coexist within the same document, as Gartner promises record revenues while preparing clients for a "cold shower."
On July 30, Tim Cook conducted his final meeting with investors as Apple's CEO, describing the price increases as a "once-in-a-hundred-years flood," justifying the company's price hikes. CFO Kevan Parekh attempted to "sweeten the bitter pill" by stating that without the expensive components, the adjusted gross margin for the quarter would not have declined at all.
On the same day, Amazon CEO Andy Jassy raised the company's capital expenditure forecast for 2026 from $200 billion to $220 billion due to rising memory prices. Just a week earlier, Elon Musk described the situation in one word—"madness."
Retail consumers will feel the effects later. Analysts are already predicting price increases for laptops and smartphones, while Microsoft anticipates a $25 billion increase in infrastructure costs.
The shortage does not appear to be temporary. SK Hynix CEO Kwak No-jun stated on July 10 to Reuters that 2027 will be the worst year in history for supply issues, with demand expected to exceed the company's capabilities even beyond 2030. UBS predicts an unbalanced DRAM market at least until the second quarter of 2028.
It is no surprise that given the current situation, Burry has opened short positions. However, if the shortage persists for another two years, manufacturers are guaranteed record margins.
Memory price increases. Source: Gartner.A New Short Selling Game
The second quarter of 2026 was the best for American stocks since 2020, with the S&P 500 gaining 14.9% and the Nasdaq increasing by 21.4%. During the same three months, SOXX surged by 94%, and Applied Materials' stock more than doubled.
On the last day of the quarter, Burry opened five short positions: Nvidia at $198.09, Applied Materials at $729.40, Tesla at $416.22, Caterpillar at $1060.98, and SOXX at $642.80. Concurrently, he rolled over puts on this ETF from January to March 2027, raising the strike prices to between $400 and $450.
Now, Burry reveals his trades not in SEC filings but on his blog, Cassandra Unchained, on Substack. His hedge fund, Scion Asset Management, deregistered with the regulator at the end of 2025. Since then, he has not filed Form 13F, leaving the size of his bets unknown—only entry prices and dates are public.
Two days later, he added Micron to his list at approximately $1052 per share.
Burry justifies his position on this company with statistics rather than forecasts. Over its 42-year public history, Micron has experienced 34 downturns of more than 30%. Its median return on invested capital is 4%, and on equity, it is 7%. Free cash flow has been negative in 48% of reporting periods, meaning that roughly every other quarter, the memory manufacturer destroys capital rather than creates it.
Additionally, he points to a technical argument. At the end of June, the PHLX Semiconductor index was trading about 65% above its 200-day moving average—a deviation not seen even at the height of the dot-com bubble.
Value “on Paper”
As of the close on July 31, the situation appeared mixed.
Burry's short positions and their performance. Source: Cassandra Unchained blog on Substack.The other stocks are showing similar behavior. According to TheStreet, out of the five positions disclosed at the end of June, only one is performing against the investor: as of July 31, it was trading slightly above the entry price.
Burry has increased his stake in this particular stock. By July 30, he bought additional put options with strike prices around $100–110 expiring on December 18. At a share price of $200.75, these contracts will only be profitable if the stock price drops by about half over the next 4.5 months. This no longer serves as a hedge but represents a standalone bet on a crash with a specific timeline.
One stock stands out in this list: Caterpillar. The company does not produce chips. Its Energy and Transportation division manufactures generators and turbines for data centers, with sales in this segment having increased by 17% to $8.4 billion.
Burry is betting against AI infrastructure construction but is approaching it from an unexpected angle—not through chips but through energy equipment. In his blog, he noted that he previously held only long positions in this stock, and always successfully.
A Diagnosis Without a Date
The logic behind the short positions is clear. The more challenging question is when exactly the bet will pay off. Burry's reputation rests on correctly identifying the cause, not on timing the market.
He played against the mortgage market nearly two years before the global financial crisis, and throughout much of that time, investors in his fund demanded their money back. In 2023, he lost about $1.6 billion in notional value on short index positions.
There are no external investors anymore; Scion Asset Management, through which Burry managed other people's money, has closed. The public short operates in two modes—both as a position and as material for his paid blog, Cassandra Unchained. It is not clear where one ends, and the other begins based on available data.
What is known is that as stock prices decline, the investor is increasing his already profitable position. He has added to Micron three times—from $1052 to $880 per share.
Burry's main argument centers not on stock prices but on how hyperscalers account for the costs of purchased chips.
Depreciation Mine
Burry's critique is easiest to understand through a simple analogy.
A company buys a laptop for $3000 and records it as an expense not all at once but over time—$500 per year for six years. This is how depreciation works: the cost of equipment is "spread out" over its useful life.
The problem arises if the technology becomes obsolete in three years. An honest schedule would then require writing off $1000 annually. This means that in each of those three years, reported profits would be inflated by $500, and the remaining $1500 would need to be recognized as a loss all at once when the device is sent to the landfill.
Now, imagine this on an industry scale. Instead of a laptop, consider Nvidia accelerators, instead of $3000, billions in capital expenditures, and instead of six years, the timeframe each company sets for itself.
This is where Burry's main argument against the sector lies. In November 2025, he calculated that five hyperscalers—Alphabet, Amazon, Meta, Microsoft, and Oracle—would understate depreciation by $176 billion for 2026–2028. He estimates that by 2028, Oracle will overstate profits by 26.9%, and Meta by 20.8%. He labeled the practice of extending asset lifespans as one of the most common forms of fraud in the modern era.
He has not disclosed his calculation methodology. However, in the financial reports of hyperscalers, there is a fact that is harder to dispute.
At the beginning of 2025, Amazon reduced the lifespan of some servers from six years to five, citing rapid technological advancement. This adjustment led to an additional expense of $677 million over nine months.
In the same quarter, Meta did the opposite, extending the write-off period from five years to 5.5 years. This single adjustment added $2.9 billion to profits.
Two approaches to equipment depreciation. Source: Cassandra Unchained.On July 8, Burry returned to the topic of lifespans and reframed it. Depreciation, in his view, does not answer the question of when a chip will cease to function physically; it indicates the timeframe over which investments are expected to pay off while the equipment remains technologically relevant. When one timeframe is extended and another shortened, it suggests that the issue is not in measurement but in a "lever" for managing profits.
The difference between the two approaches was highlighted by market participants themselves. Nvidia CEO Jensen Huang joked at a GTC conference that after the mass rollout of Blackwell, the previous generation of accelerators would be "impossible to give away."
Microsoft CEO Satya Nadella explained in a podcast that the company intentionally distributes chip purchases across generations:
"I didn't want to be stuck for four to five years with the depreciation of one generation."
Both the accelerator manufacturer and one of its largest customers are expressing the same sentiment. New-generation accelerators are being released faster than companies can depreciate the previous equipment.
There are counterarguments to this concern. Daniel Newman, head of The Futurum Group, agreed with the framing of the issue but questioned whether Burry possesses sufficient technical expertise to assess the actual lifespan more accurately than the engineers at Meta and Microsoft.
Practical evidence provides a stronger counterargument. At the cloud provider CoreWeave, A100 accelerators from 2020 are fully loaded, while H100 machines are being resold at about 95% of their original price after contract expiration. Huang himself softened his language in June, stating that older models are far from obsolete; rather, new ones offer better performance-to-price ratios.
Building on Credit
The debate over depreciation would not concern anyone if data centers were built using companies' own funds. However, the combined capital expenditures of the five hyperscalers have increased more than fivefold: from $95 billion in the 2020 fiscal year to $490 billion in the twelve months leading up to May 2026. In the coming year, all firms' infrastructure investments may exceed their free cash flow.
This strain has also manifested in the debt market. Five-year credit default swaps for Oracle surged to 212 basis points at the end of July : insuring $10 million of the company's bonds now costs about $212,000 annually. Subsequently, S&P downgraded the issuer's rating to BBB−, and Alphabet experienced a negative rating for the first time since its IPO.
The answer hinges on how quickly computing costs decrease.
Memflation vs. LLM-flation
On July 16, at the World Artificial Intelligence Conference in Shanghai, Beijing's Moonshot AI introduced Kimi K3—a model with a context window of one million tokens.
The market reacted swiftly: semiconductor manufacturers' stocks began to decline the very next session, and within a few days, the global semiconductor sector lost over $3.3 trillion in market capitalization.
The sell-off of Moonshot AI's shares is partially linked to expectations that Chinese developers could offer cheaper AI solutions. However, the cost of accessing the new model through API does not appear dramatically lower than the market: the company charges $3 for a million input tokens and $15 for a million output tokens—roughly three times more than the previous model and comparable to several Western competitors' prices.
The savings lie elsewhere: the model requires fewer tokens to solve a task. Nevertheless, it scores lower in independent rankings compared to flagship models from OpenAI and Anthropic.
Comparison of leading AI models. Source: Artificial Analysis.Eight days after the launch of the Chinese AI model, Anthropic released Claude Opus 5, claiming the top spot in the same ranking with a score of 61. The update cycle has shrunk to weeks.
The issue is not the demand for AI but rather how much computation is required to achieve the same result. This metric has been declining for five consecutive years.
This phenomenon has been termed "LLM-flation." The venture capital firm a16z calculated that the price of a model of comparable quality decreases by about ten times each year. The cost of GPT-3 at the end of 2021 was $60 per million tokens, but three years later, it dropped to $0.06.
While one aspect is becoming more expensive, another is decreasing in price. Data center equipment prices are rising by 125% per year, while the computations for which they are purchased are losing value at a similar rate over the same period.
Memflation and LLM-flation. Sources: Gartner, a16z.As early as the 19th century, English economist William Stanley Jevons observed that the rise in efficiency of steam engines did not reduce but rather increased coal consumption, as cheap energy opened up entirely new use cases.
Kimi K3 provided an opportunity to revisit this observation. Just days after its launch, Moonshot AI suspended new user registrations due to insufficient GPU capacity. The model, which initially frightened the market with its accessibility, could not handle the surge it itself had triggered.
Another detail contradicts the bearish scenario. To launch K3, about 1.4 TB of memory is required. The manufacturers most benefiting from the "devaluation of intelligence" are Micron, SK Hynix, Samsung Electronics, and the newly listed CXMT.
Laura Wang from Morgan Stanley maintained her forecast for the capital expenditures of the five largest cloud providers to exceed $800 billion this year and $1–1.2 trillion next year.
There is a flaw in this picture. All benchmarks at the time of the presentation were provided by Moonshot AI, and the model's weights were revealed only on July 27—eleven days later.
Burry does not dispute Jevons' logic. His objection concerns timing. While falling costs will indeed boost demand for computations, the depreciation clock is ticking: this is precisely when the "depreciation mine" comes into play.
How Long Will the Shortage Last?
The memory situation will normalize once new production lines come online. However, these do not come online overnight; it takes several years from the start of construction to the first deliveries.
The largest facility under construction is Samsung's P5 plant in Pyeongtaek, South Korea. Assembly is expected to be completed in the first half of 2027, with mass production slated for the second half of 2028.
On July 4, Micron began expanding its site in Hiroshima—$9.3 billion with support from the Japanese Ministry of Economy. Equipment there will be operational in the second half of 2028. Even those capacities that come online sooner will only reach full output closer to the end of 2027.
Sources: SemiWiki, TrendForce.Beyond this point, the paths for the two types of memory diverge. In a July 30 review, TrendForce forecasted that DRAM will remain in structural deficit until 2028, with the gap between supply and demand widening in 2027. NAND is expected to transition to surplus by the second half of next year.
Micron CEO Sanjay Mehrotra acknowledged that the company lacks clarity on when supply will catch up with demand. Goldman Sachs has also provided estimates: the memory needs of American data centers are projected to grow by 9–12% in 2027–2028, while regional capacities will expand by only 2–4%.
The situation with DRAM is further complicated by generational shifts. HBM4 requires 16 crystals for a stack instead of 12, resulting in each AI accelerator consuming approximately one-third more modules. Manufacturers are reallocating resources in favor of this standard, while supplies of memory for servers and consumer electronics remain limited.
This explains why the shortage isn't supporting stock prices. Investors are looking not at the current margins but at what they will be after 2027. The more manufacturers invest in expansion now, the closer the moment will be when supply meets demand, and excess profits will vanish.
AMD's vice president David McAfee described the cycle's logic more accurately than analysts when he stated that for decades, memory manufacturers have earned billions when demand outstripped supply and lost just as much during the opposite phase of the cycle.
There is a caveat to the forecast. TrendForce acknowledges that if agent-based AI is adopted more quickly, the market may fully utilize new capacities, preventing a surplus. No one has yet published reliable demand forecasts for such a scenario.
Three Diagnoses of One Market
No one knows exactly how much the industry will earn this year. In December 2025, WSTS anticipated $975.4 billion. Just six months later, it raised its forecast to $1.51 trillion, reflecting an increase of 89.9% compared to the previous period.
On January 15, Gartner placed artificial intelligence in the "valley of disappointment" for all of 2026. The agency's vice president John-David Lovelock explained the reasoning: large clients are less likely to launch separate expensive projects for neural networks and instead receive them as part of products for which they already pay contractors.
Analysts estimated the total investment in AI solutions at $2.5 trillion, up 44% from the previous year. Three months later, Gartner promised record revenues for semiconductors.
Views among investment banks are similarly divided. BofA's bubble risk indicator, where one indicates extreme overheating, approached the upper limit of the scale by the end of June. Conversely, JPMorgan maintained a recommendation of "overweight" on chipmakers on July 20 and called the downturn a convenient tactical buying opportunity.
Strategist Mislav Matejka's team explained why the July sell-off is unlikely to evolve into a prolonged downturn. The MSCI World index remained within 1–2% of its all-time high, despite some heavyweights in AI dropping by tens of percent. European semiconductor producers' stocks turned down even as profit forecasts for the next twelve months continued to rise.
The relative strength index for SOX approached oversold territory. Positioning metrics, which had reached their highest levels since 1999–2000, returned to normal by mid-July. This is why JPMorgan advises clients to increase their positions over the summer.
Charlie Dai, vice president and chief analyst at Forrester, offered a different perspective. He attributed the July sell-off to an overestimation of expectations following an exceptionally strong rally rather than a weakening demand for AI. According to him, investors are merely checking whether current revenues can justify record investments.
Comparing multiples provides a new angle. Nvidia gained 11.8% over the past 52 weeks, with a market capitalization of $4.86 trillion as of July 31. Investors are willing to pay about $30 for every dollar of annual profit from Nvidia. Micron, during the same period, has risen more than sixfold even accounting for the July decline, but its revenues have grown even faster: one dollar of profit costs buyers $20, while next year's expectations drop to $6.
In the past twelve months, the memory manufacturer earned $50.47 billion on revenues of $90.27 billion. The market values this performance more conservatively compared to Nvidia, which has seen little change in its stock price over the year.
Not only chips are expensive. The Buffett Indicator—the ratio of the market capitalization of the U.S. market to GDP—stood at 218% in the first quarter, down from a record 219% in the previous quarter. The price-to-sales ratio for the S&P 500 reached 3.22, compared to a historical average of 1.84.
Forecasts for the global semiconductor market. Sources: Gartner, WSTS.***
The market is seeking a new equilibrium. The DRAM shortage is expected to last until 2028, but the introduction of new capacities is constrained by construction timelines. Stock prices often reflect not the current margins of producers but the expectations of what they will be after the market saturates.
This gap explains the volatility of July.
The trend's vulnerability lies in leverage. Capital expenditures from hyperscalers are growing faster than operating cash flow. The difference is being covered by debt, and its cost is already reflecting increased risk. The higher the burden, the sharper the stocks react to any deviation from the plan—right down to adjusting server lifespans.
As computing costs decrease, demand surges while previously purchased hardware depreciates. Which of these processes will outpace the other in the next three years remains uncertain, as financial reports do not provide clear answers.
Unlike the dot-com boom, this cycle is characterized by real demand backing investments, rather than just promises. Some companies will not survive; bankruptcies and acquisitions have accompanied every previous investment boom. Burry identifies the cause but not the timing, and his calculations will only be verifiable over time.