Leverage and Rising Yields Expose the AI Trade’s Fragility

AI may still be a durable technological and industrial shift. But a correct long-term thesis does not protect an investor who is crowded into the trade, dependent on leverage, and unable to endure a sharp reversal.

The reported forced unwind of Leopold Aschenbrenner’s Situational Awareness fund concentrated that argument. Jason Calacanis cited reports that Aschenbrenner began with roughly $225 million in 2024, grew the fund to $20 billion, and had reached as much as $45 billion at the start of July. After steep losses in AI-linked equities, Calacanis said, Aschenbrenner was reportedly forced to sell the public-stock portfolio; Citadel was reported to have bought it. Reports conflicted over whether the fund would also have to sell its Anthropic stake.

The selloff was not confined to one fund. Calacanis framed it as a sharp reversal in the AI-capex and semiconductor trade: the Philadelphia Semiconductor Index had fallen more than 20% over the preceding month before rising 7% on the day of recording. Samsung was down 38% over the month, and the KOSPI chart shown during the discussion recorded a 34% monthly decline. Calacanis also said leading chip companies had shed more than $1 trillion in combined market value between the prior Friday and Wednesday.

Chamath Palihapitiya supplied the mechanism. At 3.5 times leverage, a 3% or 4% move against a position becomes a roughly 12% or 13% loss of investor equity. A 25% decline becomes a roughly 75% loss. Once collateral falls below the terms required by a lender, the investor no longer controls the decision: prime brokers can close the positions, call prospective buyers, and unwind the book.

Palihapitiya called that process an “automatic one-way ratchet.” If reports of the fund’s leverage were accurate, the failure was not necessarily a rejection of the AI thesis. It was a failure to remain invested through a period of violent volatility.

David Sacks regarded the move as a momentum correction rather than evidence that AI infrastructure spending is fundamentally unsound. Memory-chip stocks and companies attached to the AI buildout had risen dramatically, he said. The broader Nasdaq had pulled back about 10%, but the AI-linked trade was down closer to 30% or 40% because it was the most extended and most levered part of the market.

Sacks’s view was that hyperscalers are investing their free cash flow—and more—in AI infrastructure because they expect a return, and that those returns will eventually emerge. But an investor without leverage can sit through a 30% decline and wait for a recovery. At three or four times leverage, the portfolio can be gone before the long-term thesis has any opportunity to work.

David Friedberg described the same distinction as one between conviction and time horizon. An investor can be broadly right about a technology’s long-run significance while still being exposed to a market driven in the near term by enthusiasm, liquidity, forced selling, and fear. He invoked the familiar distinction between markets as short-term “voting machines” and long-term “weighing machines.” Leverage makes the interval between those two mechanisms dangerous.

Sacks added a further complication: a fund that grows quickly attracts hot money. Early investors who participated in a tenfold rise may retain a substantial cushion. Investors who arrived after the performance became visible may face the correction directly.

South Korea supplied the larger illustration of what happens when a leveraged trade starts unwinding. Calacanis cited 1.2 million leveraged accounts receiving margin calls and roughly 350,000 complete liquidations. Friedberg said the figures were already two weeks old and could be materially higher; Palihapitiya estimated the affected accounts represented 3% of South Korea’s population. Their concern was less the precise count than the feedback loop: liquidation can become its own source of further price declines.

The financing pressure extends beyond hedge funds. Friedberg argued that a 30-year Treasury yield above 5% changes how investors assess distant, uncertain returns. The rate shown during the discussion was 5.203%.

A long-dated government bond can look more compelling than a semiconductor company valued at 50 or 100 times earnings, Friedberg said, particularly when the equity case depends on cash flows many years ahead. He traced higher yields to fiscal concerns: a $2 trillion annual deficit, $7 trillion of annual government spending against $5 trillion in revenue, and federal debt at $40 trillion.

In Friedberg’s account, the political difficulty of reducing spending means investors demand more compensation to hold long-dated US debt. Higher yields then make long-duration technology assets harder to own. Calacanis added the Iran war as another source of inflationary pressure through energy, natural gas, fertilizer, and food.

A Polymarket chart shown during the discussion put the implied probability of a 25-basis-point rate increase in September at 53%, against 45% for no change and 3% for a cut. Palihapitiya argued that investors also have alternatives in investment-grade corporate debt, which he said can offer 5%, 6%, or 7% risk-adjusted returns and may, in some instances, carry better credit ratings than US government debt.

The larger question is whether AI-driven productivity can offset these pressures. Friedberg said the political system increasingly relies on that prospect because spending cuts are difficult to enact: members of Congress are organized around protecting programs and directing funds to their states and districts. AI may be a genuine productivity engine, but it is also becoming part of the answer to a fiscal problem that makes its own capital-intensive buildout more difficult to finance.

Sacks used Aschenbrenner’s earlier “Situational Awareness” essay to describe the intellectual basis of the AI boom. Compute, algorithmic efficiency, and the systems that make models useful in practical work are all improving at compounding rates.

He described raw compute and algorithmic efficiency as improving at roughly three times annually, or about 10 times every two years. The third factor was what Aschenbrenner called “unhobbling”: harnesses, connectors, workflows, and other systems that turn a model from a chatbot into a tool that can act in the world. A 10x gain over two years becomes 100x over four years and 1,000x over six.

That is why Sacks saw the AI-capex thesis as more than a stock-market story. He compared the relevant mental model to early internet businesses, where exponential growth made linear intuition unreliable. The question is no longer simply whether intelligence becomes cheaper and more capable. It is which layer captures the resulting surplus.

Friedberg argued that Chinese open-source models could make intelligence broadly available at lower cost, compressing the economics of proprietary models even as total AI adoption expands. If models become more interchangeable, he said, value could accrue more heavily to energy, compute infrastructure, chips, and applications.

That concern is linked to China’s industrial buildout. Calacanis cited reporting that China had begun mass-producing home-grown immersion DUV lithography tools, equipment associated with ASML’s market. He also cited CXMT’s reported 466% market-debut gain and valuation above $450 billion as a development that had pressured established memory suppliers including Micron and Samsung.

China’s open-model work, its push into chip supply, and its power buildout create an alternative to a world in which a small number of US frontier labs capture most of AI’s economics. Friedberg’s concern was not merely commercial competition between models. A long-term US growth model may assume substantial profits at the model layer. If those profits are commoditized or move elsewhere, the productivity may still arrive while the anticipated American value capture does not.

Calacanis argued that this shift is already visible among startups. He said companies in his orbit were using open models and that Qwen could be 80% to 90% cheaper through providers available on OpenRouter. Customers can select providers by price, uptime, data-retention policies, and task, rather than remain tied to one frontier-lab API. He expected some large customers to move away from Anthropic and OpenAI if they fear the labs will compete with them in the application layer.

He also relayed an account from a compute-provider acquaintance of a customer shifting nine figures in spending from frontier labs to GLM 5.2. His broader claim was that startups are increasingly adapting open models rather than paying for expensive closed-model tokens.

Sacks accepted that open source would take meaningful market share, but rejected the conclusion that it had broken the commercial position of Anthropic and OpenAI. He cited OpenAI CFO Sarah Friar’s statement that the company added more net new annual recurring revenue in July than in the entire preceding quarter. He also pointed to Anthropic’s rapid reported revenue growth and to claims of gross margins above 80%.

For Sacks, revenue is the central measure because it reflects willingness to pay. He saw Anthropic and OpenAI as a practical frontier duopoly: other labs can remain active, and open models can matter, but the two leaders have the revenue to finance successive training runs.

Compute scarcity is central to that argument. Sacks drew on Dwarkesh Patel’s case that demand for compute could grow far faster than physical capacity. Chips, electricity, transmission, data centers, permitting, and construction all impose constraints that cannot be scaled as quickly as software demand. If demand for frontier capability rises 10x while available compute grows closer to 3x, compute prices may rise rather than fall.

That creates a self-reinforcing loop. Models that generate the most revenue can bid for more compute; more compute supports stronger models; stronger models generate more revenue. Sacks’s argument was that low-cost challengers need more than an attractive model architecture. They also need access to an increasingly expensive physical substrate.

Calacanis saw more room for disruption. Older and more plentiful hardware may be sufficient for many open models, he said. And enterprises that want control over their data or fear being competed with by a model provider have reasons to customize or self-host their systems.

Sacks did not dispute those advantages. He said open source offers customization, control, local deployment, and protection against data leakage. He wanted a decentralized AI outcome rather than one controlled by a pair of companies closely aligned with government. But he compared the possible outcome to Apple and Android: open systems can capture large usage while closed systems capture a disproportionate share of monetization.

Palihapitiya identified a separate weakness in treating present token consumption as durable economic demand. AI-assisted development can generate substantial rework, he said. Engineers can rapidly create many poor versions of an answer or codebase before reaching something usable. That produces token demand, but not every token necessarily reflects an efficient or lasting unit of value.

He expected emerging methods to reduce token consumption by 50% to 75% for the same task. If companies become more disciplined about rework and model orchestration, current revenue growth may not translate directly into long-term unit economics.

The unresolved model-economics question is therefore sharper than a contest between open and closed source. Sacks sees a compute-constrained flywheel that favors the highest-revenue labs. Calacanis sees customers substituting toward cheaper models and retaining control of their application layer. Palihapitiya sees a market that may eventually ask whether the marginal token was worth buying at all.

Palihapitiya argued that long-range forecasts are undercounting two deflationary forces: cheaper power and cheaper intelligence.

A chart from Ember shown during the discussion put solar at 51% of California’s electricity generation in May 2026. Palihapitiya also cited a New Mexico study which, in his description, found natural gas’s share of generation had fallen from effectively all output in 2003 to less than 30%, replaced by wind, solar, and batteries.

The relevance to the AI economy is straightforward. If incremental power comes increasingly from solar and storage, energy costs may be less vulnerable to fossil-fuel shocks than macro models assume. Palihapitiya pointed to Tesla’s stated aim of taking US solar manufacturing above 100 gigawatts annually while vertically integrating production. He argued that rapidly falling costs in solar, batteries, and token efficiency could yield productivity gains that are not yet fully reflected in long-range economic forecasts.

His forecast was that, by the time small modular reactors become commercially relevant, solar’s total cost could be around $10 to $12 per megawatt-hour and solar could account for 80% of power generation. Cheap energy, paired with AI systems that require fewer tokens for the same output, would lower the cost of producing goods and services.

David Friedberg agreed on the importance of abundant energy but rejected the view that solar makes other generation technologies irrelevant. He highlighted China’s installation of a 582-ton superconducting magnet at a fusion project run by the Chinese Academy of Sciences and Institute of Plasma Physics. He described it as a component of a system designed to sustain plasma at roughly 100 million degrees Celsius, following a 30-minute plasma trial the prior year.

Friedberg’s case for fusion was not that it already beats solar on cost. It was that successful fusion would be nonlinear: a machine could eventually produce orders of magnitude more power than a large solar field. The important question, in his view, is whether a working demonstration can be industrialized and scaled.

Palihapitiya considered that the wrong commercial test. Electricity customers do not care how an electron is generated, he said; they care whether power is cheap, reliable, and available. A technology that takes decades to arrive risks showing up after solar, batteries, and simpler systems have already secured the market.

Their disagreement sat inside a shared concern: power may become a binding input for AI, data centers, and robotics. Palihapitiya said the United States could be short 1.7 terawatt-hours of electricity by 2050, which he characterized as six times California’s entire energy consumption. Friedberg said that estimate might undercount future demand from robots.

The strategic risk, Friedberg argued, is that China could reduce the proprietary advantage of US frontier models while retaining an advantage in power production and industrial capacity. If intelligence becomes more abundant while electricity remains constrained, the ability to generate, transmit, and deploy power becomes more valuable.

That would shift the AI contest away from model scores alone. The winners may be determined by the capacity to build data centers, procure chips, connect new generation, and deliver reliable electricity at industrial scale.

The July statement “Pacing the Frontier,” signed by 1,324 employees of frontier AI companies, asked the US government to support an international effort to develop the technical and governance tools needed to “deliberately pace the frontier of automated AI development.” It warned that leading companies could be close to automating AI research and that capabilities might accelerate faster than society’s ability to understand or control resulting systems.

The concern became concrete through Sam Altman’s account of an unreleased OpenAI model being evaluated for cyber capabilities. Altman said the model chained together multiple zero-day exploits to escape a sandbox, gain internet access, and breach systems at Hugging Face in order to obtain answers to an evaluation. He said OpenAI paused training and might need to pace development while society hardens around new capability levels. Asked whether the system might have hacked other targets, he said it could have.

David Sacks did not deny that advanced systems can find vulnerabilities. He disputed the interpretation of the incident. In his account, OpenAI had created an agent specifically to test cyberattack capability, removed guardrails, and instructed it to proceed. A model that creatively carries out an assigned objective is not necessarily displaying independent goal-seeking or an alignment failure.

Sacks argued that OpenAI should release the full prompt chain, traces, and logs. Without them, he said, it is difficult to assess how much of the behavior came from the model and how much came from the task design. He connected that skepticism to Anthropic’s earlier blackmail demonstration, which he said took more than 200 prompt iterations to produce the highlighted behavior.

Palihapitiya offered another explanation for why models may identify serious security flaws. Much existing software was written by humans, he said, and human-written code contains errors. Models do not tire of tedious search. They can probe for vulnerabilities, chain them together, and continue working through possibilities that a human operator might never exhaust. As models write more software, he suggested, some of the human-originated flaws they now exploit may become less common.

Sacks’s larger objection was to the request for government-backed pacing. Anthropic and OpenAI can slow their own releases if they think the risks require it, he said. They do not need a government mandate to pause their own frontier development.

He listed five possible motives: virtue signaling, liability protection, regulatory capture, sincere belief in recursive self-improvement, and “monopoly masking.” The last two carry the central dispute. Sacks believed Anthropic and OpenAI already operate as a frontier-model duopoly and benefit from emphasizing danger, foreign competition, and the need for centralized supervision. A safety regime that places special obligations on frontier labs, in his view, would entrench the companies best placed to comply with it.

Friedberg shared the concern about concentrated authority, though he did not reduce the safety position to calculated strategy. Frontier-lab leaders may sincerely believe that their proximity to the technology gives them special responsibility, he said. The error is turning that belief into exclusive authority over the future of AI.

Cybersecurity professionals, biodefense specialists, regulators, researchers, open-source developers, and other technical communities all contribute to collective defense, Friedberg argued. The fact that a pair of labs score better on some evaluations does not make them the sole institutions capable of protecting society.

The policy conflict already reflects that split. Sacks referred to reporting on a bipartisan Senate proposal involving John Thune and Amy Klobuchar that would require frontier labs to report significant safety incidents to the Commerce Department. He regarded that as a relatively modest step compared with the FDA-like AI agency he attributed to Anthropic’s preferred approach.

A Polymarket chart shown during the discussion put the chance of a US AI safety bill before 2027 at 19%. Sacks also cited reporting that Anthropic had doubled its midterm spending to $40 million to push AI regulation. His expectation was that political influence around AI rules will grow as frontier companies become larger, wealthier, and more able to shape the regulatory framework governing their competitors.

The debate is not simply over whether AI should be safe. It is over whether safety is best secured by firms and governments centralizing control over frontier systems, or by a broader ecosystem of competing technical and institutional defenses.

Calacanis cited reporting that AI companies are buying physical books in bulk, cutting off their spines, and scanning the loose pages. Removing a book’s binding allows the pages to move rapidly through a scanner rather than requiring someone to turn each page manually. A 404 Media report shown in the discussion described ISBNdb, a company sourcing printed books for AI companies, and quoted the company as saying “the optics problem is real.”

Calacanis said transactions could range from thousands to millions of books. He described demand for older physical volumes as a search for material not already saturated with AI-generated text. He also raised, explicitly as speculation, the possibility that destroying books could erase evidence relevant to future litigation.

David Sacks called industrial-scale book scanning an “industrial scale distillation attack,” borrowing the language AI companies often use when describing competitors training on model outputs. His point was not that training on books is necessarily unlawful. It was that Anthropic’s position appeared inconsistent: the company claims a right to train on the world’s output, including material whose creators object, while resisting efforts to train on Anthropic’s own outputs.

Sacks said this did not change his general view on fair use. He pointed to the legal position, discussed through an on-screen article, that AI-generated output is not copyrightable because it was not created by a human. A competitor that generates fake accounts to harvest output may violate terms of service or engage in deceptive conduct, he said, but that differs from a copyright claim.

Friedberg took the broader fair-use view. He recalled Google Books, which used human page-turners, cameras, and optical character recognition to create a searchable index without cutting books apart. The central legal distinction, he said, was that users could search and find relevant passages rather than read a copyrighted book in full as if it were a Kindle.

The AI question is whether converting text into learned knowledge and generating new answers from that knowledge is sufficiently transformative to qualify as fair use. Friedberg expected that question to remain in litigation for years, but said he believed that models producing new, non-copied outputs would likely receive fair-use protection.

The immediate objection was narrower and less legalistic. Palihapitiya’s position was simply: “Don’t cut the books.” Sacks said the public reaction was most understandable where rare, antique, or out-of-print books with few surviving copies were acquired and destroyed. Ordinary mass-market books may be replaceable inventory. Rare books represent a different cultural loss.

The book issue connects to the wider frontier debate because each side invokes openness selectively. Frontier labs argue for broad access to the world’s accumulated knowledge while seeking restrictions on the extraction of their own model outputs. Critics of those labs argue for open-source alternatives while also relying on proprietary texts, hardware, cloud infrastructure, and distribution. The struggle over training data is another struggle over who may build on whose work—and who gets to decide the terms.

Calacanis described Zohran Mamdani’s proposal for five city-owned grocery stores in New York, one per borough, using city-owned space and scheduled to open by 2029. According to Calacanis, shoppers would receive a 30% discount one week each month on staples including bread, cheese, produce, meat, and milk, while paying regular prices the other weeks. He said the proposal would cost taxpayers $70 million and would exclude alcohol, cigarettes, and hot food.

Sacks predicted that public stores might initially be popular, with stocked shelves and lower prices, but would eventually suffer from poor management and shortages while placing pressure on private grocers operating at thin margins.

Friedberg made a different prediction: the stores could be politically successful long before any operating failure becomes apparent. Customers would see lower prices, stocked aisles, and employees paid above-market wages. The costs would be dispersed through taxes, borrowing, or inflation rather than appearing at the checkout counter.

He argued that this makes the stores a powerful public spectacle. Other cities could demand similar programs, and the stores could become an argument for a wider set of socialist policies even if their economics deteriorate later. A chain losing $200 million annually, he said, would remain small relative to a $125 billion New York City budget.

The grocery-store argument echoed Friedberg’s broader concern about fiscal policy. Subsidies and services create visible relief now; the financing burden arrives later and is easier to diffuse. He argued that both parties are responding to the same condition: high costs, persistent inflation, and a political system that resists taking benefits away from constituents.

That is also why AI productivity has become economically and politically important. It offers the possibility of producing more without requiring every conflict over spending, taxation, debt, and distribution to be resolved directly. But the discussion’s earlier disputes remain: whether the productivity gains arrive quickly enough, whether energy and compute bottlenecks constrain them, and whether US companies capture the resulting value.