The “Big Short” is crazy! Michael Berry snipers AI beliefs from underlying technology: when the parameter trap hits a peak in production capacity, the AI bull market ushered in a “liquidation time”
The Zhitong Finance App learned that Michael Berry, who has the title of “Big Short,” has been posting pessimistic remarks about “the end is coming” on his Substack subscription platform, and is shorting popular AI technology stocks at a time when global capital continues to flock to the theme of AI computing power infrastructure. Michael Burry, the prototype character of the movie “The Big Short,” has recently focused more clearly on crowded transactions related to AI computing power infrastructure and AI semiconductor capital expenditure cycles. For example, Burry has expressed bearish positions on Nvidia, Tesla, Micron, Applied Materials, and iShares semiconductor ETFs through put options or bears.
The latest series of AI-themed high-frequency shorting operations by this top global investor, famous for successfully and accurately predicting the 2008 US real estate crash and the full outbreak of the subprime mortgage crisis, was interpreted by the market as his bearish stance on extremely crowded storage chip themes driven by the AI infrastructure frenzy, semiconductor manufacturing equipment, and popular stocks related to AI chips such as AI GPU/AI ASIC, and systemic doubts about the return on investment expectations of increasingly huge AI computing power infrastructure capital expenses and the basic situation in high-valuation technology stocks . He also previously disclosed his put options positions on Nvidia and Palantir through 13F. As the Korean stock market plummeted again to collapse on Monday and popular technology stocks related to AI computing power in the global stock market fell into sharp fluctuations, some investors began to agree with the argument that the AI bubble was about to burst and joined Burry's shorting camp on the AI theme.
Bury sees SK Hynix and Samsung Electronics' recent unprecedented large-scale memory chip production expansion plan led by the South Korean government as one of the important signs of the “beginning to the end” of the AI investment cycle. Burry sees the Korean project as a “peak capital expenditure signal” rather than an “immediate production capacity signal,” and says it is the “beginning of the end” of the AI boom. His judgment is not that these factories will immediately cause oversupply next year, but is based on the semiconductor industry's classic cobweb cycle: the most optimistic demand forecasts, the highest chip prices, and the most relaxed financing conditions often prompt all manufacturers to announce huge expansion of production at the same time; since advanced fabs take many years from planning, land, electricity, and clean room construction, and large-scale introduction of semiconductor equipment to mass production, the real supply boom usually arrives when market demand has cooled down.
The “Big Short” latest shorting operation and extreme pessimism about the AI superbull market can be described as the complete opposite of the positive bullish stance of Wall Street financial giants such as Bank of America, Goldman Sachs, and Nomura. These top global financial giants agree that as corporate IT budgets flow more and more to application ecosystems and AI inference computing power budgets related to AI agents, plus the launch of OpenAI GPT-5.6 and ChatGPT Work smart devices, and Anthropic is expected to move from long-term losses In a new phase of fundamental expansion where overall profits are skyrocketing, global demand for AI computing power is almost endless.
“Big Short” Bury chanted AI's “parameter trap” and firmly used short positions to snipe the AI investment boom
Bury, who has the title of “Big Short,” quoted a classic line from Tim Burton's version of “Batman” clowns in a recent tweet: “Dance with the devil in the pale moonlight.” He called the AI boom a “mass addiction,” and predicted that it “could die in the midst of a myriad of attacks.” On July 1, he chose to short the stock when the stock price of US memory chip manufacturing giant Micron Technology (MU.US) reached an all-time high of $1,051.87. In the previous year, the stock had already risen by nearly 700%.
On July 10, he published more specific pessimistic and bearish views on the global AI investment boom on Substack. The content went beyond concerns about the bursting of the AI bubble and the expensive valuation of AI semiconductor themes. This is an empty attack on the technological foundation upon which the entire artificial intelligence industry is built.
According to Michael Berry, the artificial intelligence technology path was taken the wrong path from the beginning and is now stuck in it.
The article's argument revolves around what Burry called “it was wrong in the beginning.” He believes that the development of artificial intelligence chose language first, and should have chosen reasoning first; since then, the entire industry has been paying for this choice, but has never fully acknowledged this.
He framed this argument around the so-called “Ballard Test.” This is a philosophical case involving a character named Melville Ballard; he had deep reasoning skills even before he acquired strong language skills.
Barry used this to make a specific point: true understanding requires no language. Language is the output form of intelligence, not the source of intelligence.
“We mistook the output of artificial intelligence — language — as its engine — for reasoning.” Bury wrote on the Substack paid subscription platform.
According to his interpretation of the technology route of the artificial intelligence industry, the industry discovered that language is a problem that can be handled, so optimization is carried out around language; not because language can lead to general artificial intelligence, but because language is significantly expandable, and it is also easier to obtain financing from the huge technology industry.
Barry believes that models are becoming more and more capable of generating text. Investors are rewarded for this. The industry continued along this path. And in the process, the difference between generating language and actually making inferences was unwittingly overlooked.
What does Burry mean by an artificial intelligence “parameter trap”? The so-called “parameter trap” means that an industry is beginning to determine that larger scale equals better AI model performance: more model parameters, larger AI computing power infrastructure, more data input, and larger models.
Each scale upgrade brings visible improvements in benchmark scores and demonstration results. However, Barry's view is that expanding the scale of the language model will not solve the underlying logical reasoning problem; it will only make this simulation seem more convincing to users about language output.
This has real financial consequences. Companies that are investing hundreds of billions of dollars in AI computing power infrastructure are betting that scale expansion will eventually achieve general artificial intelligence goals.
Hyperscale cloud computing companies could spend $725 billion on artificial intelligence in 2026, according to TheStreet. The Philadelphia Semiconductor Index has risen 88% this year. Nvidia's market capitalization is around $5.45 trillion, and the price-earnings ratio over the past 12 months was 43 times. All of these valuations and expenses are based on the assumption that the current scaling path (that is, the scaling path) will work in the long term.
Burry believes that this path might not work. If the entire industry were investing money on such a huge scale just to improve something that wasn't the real goal, then the economic logic of the whole deal would be completely different.
This Substack article didn't come out of thin air. For months, Bury has been building short positions on popular tech stocks targeting artificial intelligence. The shorting targets he disclosed included Nvidia, Tesla, Micron Technology, Applied Materials, Caterpillar, and iShares Semiconductor Exchange Traded Fund. Just as the media revealed earlier, the Micron short position he disclosed on July 1 was established after the stock had already risen by nearly 700%.
His analytical opinion on the overvaluation of the AI semiconductor sector is that the Philadelphia Semiconductor Index is already close to its peak within the expected price-earnings ratio range for the past 15 years. His argument is that the trading price of AI semiconductor stocks is rising because hyperscale cloud computing companies are investing heavily in artificial intelligence training/inference computing power clusters; the reason why hyperscale cloud computing companies continue to invest heavily is also because they are in urgent need of these AI semiconductors and core infrastructure hardware related to AI infrastructure, which is driving global chip stocks to continue to rise. This feedback loop looks like real demand, but part of it is actually just market reflexivity.
In June 2026, the total market value of the US “Big Seven Tech” (the Magnificent Seven) stock evaporated by more than $2.2 trillion. Bury maintained a short position during this round of decline, and instead of easing his shorting position, it further strengthened his shorting position.
Why is Burry's AI technology route criticism harder to ignore than valuation theory? Most people who are bearish on artificial intelligence put forward valuation arguments: stocks are too expensive, expenses related to AI computing power are too huge, and returns have yet to be realized. There is a real basis for these arguments, but they are already familiar to the market, and the market has heard these opinions many times in the past, but they have not been greatly affected.
Bury's July 10 article was different because he attacked not price tags, but AI technology architectures. He believes that the current generation of artificial intelligence may be getting better at accomplishing something that is not really the goal of general AI.
A system that can generate more persuasive language output isn't necessarily getting closer to true artificial intelligence reasoning capabilities. If this judgment holds true, then the collapse facing the AI computing power expansion theory will be more fundamental than what valuation critics have described.
He predicted that artificial intelligence narratives “may die in the midst of a myriad of attacks.” This wording was chosen deliberately. Instead of predicting a single moment of collapse, he is describing a process where confidence is gradually eroding: as the gap between the actual capabilities of artificial intelligence and investors' previous assumptions is becoming increasingly difficult to conceal with larger parameters, more expensive AI computing power clusters, and better benchmarks, market beliefs will collapse little by little.
What does Bury's warning mean for investors who have been tracking popular stocks related to artificial intelligence for a long time? Bury has also made premature judgments in the past. He called for a stock sale in August 2023, but since then the US stock market has risen 66%. He may have been shorting Nvidia stock for over a year, and the stock continues to rise. Even if a structural judgment is ultimately correct, it does not mean that the market will agree with this view at any particular point in time.
What's different about this July 10th article is that its arguments are very specific. He doesn't just say that artificial intelligence is too expensive, but he points out a specific technical decision made at the beginning of the current wave of artificial intelligence, and believes that this decision made the entire industry seem to be on the wrong path. This is not a problem where bulls can easily preempt the past with a financial report that exceeds expectations.
For investors holding shares in the semiconductor or AI computing power cluster industry chain related to artificial intelligence computing power infrastructure, or any technology company whose valuation is based on the assumption that “the current generation of large-scale language models is leading to true general artificial intelligence,” Burry's arguments are worth pondering.
His position may eventually prove wrong. He has also misjudged in the past. But the last time he made such a specific structural critical assertion about a major market, the facts ultimately proved him right.
“Big Short” Bury and Wall Street giants are gambling for the century on “the topic of AI computing power infrastructure reaching the top”
South Korea's huge expansion of production is an important “late cycle symbol” of Bury shorting AI. His complete AI bearish logic also relies on the large-scale declaration of failure of the AI big model technology route, AI commercial returns falling short of capital expenses, extreme valuations and the radicalization of leveraged positions, and future concentrated supply releases. What investors should really track is not the forward investment announcement itself, but rather the time when the fab starts production, the growth rate of storage bit supply, customer inventory, cancellation rate of long-term purchase agreements, HBM and general storage prices, cloud vendor capital expenditure as a share of total revenue, and whether revenue generated around AI computing power infrastructure can continue to grow faster than depreciation and operating costs.
According to some pessimistic supporters of the “AI computing power infrastructure cycle is about to burst” and “the AI bubble is about to burst,” if marginal revenue from AI models declines, computing power resale increases, and storage production capacity is concentrated when demand slows down, Bury's short framework may evolve from premature countertrend trading to the real top script of the AI computing power supercycle.
Bank of America and Nomura, on the other hand, represent the strongest argument against Bury: currently, it is still insufficient supply, not oversupply, that actually constrains the industry. High-bandwidth memory takes up more wafer area and squeezes general dynamic random access memory and flash memory production capacity, and it will take 5 to 10 years to build a new fab to form an effective supply; SK Hynix CEO even expects the worst storage shortage in history in 2027, and customer demand may still exceed supply capacity after 2030.
Bank of America and Nomura's latest research report can be described as a tacit agreement that Meta's sale of idle computing power can also be explained as increasing utilization and return on capital, and does not necessarily mean that demand has peaked; if the rental of computing power reduces the cost per token, the “Jevens paradox” may occur, and instead expand overall computing and storage requirements.
Nomura, a well-known investment institution on Wall Street, released a research report to refute the “semiconductor peak theory,” and the latest research report released by Bank of America (BofA) this week shows that by 2027, against the backdrop of a strong trend where AI inference computing power continues to surge under the big wave of AI agents, global capital expenditure on cloud computing and artificial intelligence related infrastructure will reach 1.5 trillion US dollars, and points out that the current summer correction of AI semiconductors, including memory chip stocks, is a round of healthy reset trajectory, rather than any structural changes at the level of AI computing power requirements.
According to Goldman Sachs, the AI bull market is far from over. Instead, it has moved from the “AI chip purchase frenzy” to the second stage of “large-scale AI factory construction” — that is, the next round of excess alpha revenue will no longer only belong to the list of the strongest leaders in the AI GPU/AI ASIC field, but will also spread systematically to data center high-performance CPUs, DRAM/NAND/HBM storage, AI PCBs, liquid cooling systems, data center optical interconnection systems, ABF carrier boards/glass substrates, MLCC, electronic cloth, and extensive wafer foundry Facility level.
Nomura's key to refuting the “semiconductor peaking theory” is not simply saying that AI chips will rise, but rather pointing out that demand for AI cloud infrastructure is spreading from a shortage of single-point GPUs to a mismatch of systemic components. According to the Nomura Research Framework, AI server revenue is expected to increase by 78% and 76% respectively in 2026 and 2027, and the number of global data center projects will increase from 240 to 280, including about 50 gigawatt-scale projects. The deployment of additional computing power is expected to reach 32 GW in 2027, and 23 GW will already be visible in 2028.
The underlying logic behind Michael Berry's shorting the AI theme was not simply that “artificial intelligence has no value,” but rather a dangerous mismatch between betting on technology routes, return on capital, and market pricing. At the level of the underlying technology path of AI, he believes that the current industry mistakenly treats the “intelligent output” of language generation as reasoning itself. By increasing parameters, training data, and computational volume, it is not necessarily possible to simultaneously establish stable causal reasoning, world models, and long-term planning capabilities.
At the level of the AI economy, such as return on capital and market pricing, the industry invests hundreds of billions of dollars to build chips, data centers, and power systems according to the assumption that “continuing to expand the model can approach general intelligence.” If the increase in marginal capacity gradually slows and monetizable revenue fails to cover depreciation, electricity, storage, and financing costs, then the entire AI infrastructure valuation system will face a decline in return on capital. Barry summarized this risk as a “wrong starting point” and a “parameter trap” of artificial intelligence.
His second level of logic is a typical counterproductive bubble and capital expenditure cycle: rising chip stocks reinforce the narrative of unlimited AI demand, making it easier for hyperscale cloud vendors to finance and expand capital expenses; growing orders also drive up profit expectations and stock prices of chip companies, which in turn stimulates further expansion of the supply chain. This cycle appears to be a verification of demand during the upward period, but in Burry's view, part of the increase actually stemmed from mutual reinforcement of asset prices, financing capacity, and corporate investment decisions, rather than the fact that terminal AI cash flow has been fully realized. He recently shorted Nvidia, Micron, Applied Materials, semiconductor exchange-traded funds, Tesla, and Caterpillar, and actually covered multiple aspects of the AI capital cycle — AI cluster core accelerators (AI GPU/AI ASIC/TPU), storage chips, wafer manufacturing equipment systems, autonomous driving narratives, and capital goods related to data center construction, rather than just targeting an AI-related technology company.