AI Stock Unwind: Leverage, Compute Scarcity and Bitcoin

Last updated: August 1, 2026, 3:30 p.m. Eastern Time

The most useful way to understand the violent July reversal in artificial-intelligence stocks is not to choose between two dramatic stories—“the AI boom is over” or “nothing fundamental changed.” The evidence points to a more complicated conclusion. A crowded and leveraged trade broke at extraordinary speed, forced sellers amplified the decline, and the damage spread through U.S. technology shares and Asian semiconductor markets. At the same time, the latest results from Amazon, Microsoft, Alphabet and Meta show that demand for cloud capacity, AI infrastructure and enterprise services remains substantial. The investment thesis survived, but the financing structure and the prices attached to it did not escape unscathed.

That distinction matters because the market is now trying to price three different questions at once. First, how much of July’s decline was mechanical selling rather than a change in long-term earnings power? Second, can hyperscalers earn an adequate return on capital when annual spending plans have reached levels that would have seemed implausible only a few years ago? Third, does a world of faster technological disruption, less predictable central-bank communication and more programmable finance strengthen the case for Bitcoin—or merely create another narrative around an asset that remains exceptionally volatile?

Those questions were at the center of an August 1 episode of The Pomp Podcast, hosted by Anthony Pompliano and featuring veteran macro investor Jordi Visser. The source video was titled “Bitcoin Is The Best Hedge Fund That’s Ever Existed.” Visser argued that the liquidation of Leopold Aschenbrenner’s Situational Awareness portfolio resembled earlier leverage events, that compute scarcity remains the decisive economic bottleneck in AI, and that AI agents and tokenization could accelerate adoption of Bitcoin and other digital assets. His framework is provocative and, in parts, useful. It also contains claims that need qualification.

The early answer is this: July’s AI stock unwind was a warning about concentration, leverage and liquidity—not a clean referendum on AI demand. Compute scarcity is supported by current cloud growth and contracted demand, but scarcity does not guarantee attractive shareholder returns because the cost of supplying that compute is also surging. Bitcoin can be treated as a scarce digital asset and a high-risk portfolio allocation, but calling it a hedge fund confuses a volatile asset with an actively managed investment partnership. Tokenization is advancing, yet it does not automatically create wealth, eliminate economic borders or make central banks unnecessary.

The July AI unwind was a market-structure event before it was a technology verdict

Markets often tell a coherent story after the fact. Prices rise because a technology is transformative; prices fall because investors finally recognize that the technology was overhyped. That narrative habit is comforting, but it can misdescribe what happens when leverage, common positioning and thin liquidity interact. The sharpest moves are frequently caused not by a sudden discovery about long-term value, but by the urgent need to reduce exposure.

The collapse in Situational Awareness’s public-equity book is the clearest example. The fund, established by former OpenAI researcher Leopold Aschenbrenner, was built around a forceful thesis: the race toward increasingly capable AI would require enormous amounts of chips, memory, power, data-center equipment and cloud capacity. That thesis produced extraordinary gains while AI infrastructure shares rose. It also encouraged concentration in securities whose prices were increasingly dependent on the same assumptions.

According to the investor letter described by Reuters, the portfolio lost 67% in July, removed leverage and sold most of its public holdings. The reported year-to-date gain of about 80% demonstrates the strange arithmetic of compounding. A portfolio that rises 400% grows from 100 to 500. A subsequent 67% decline reduces it to 165—still 65% above the starting point, but with most of the peak value destroyed. A 70% loss requires a 233% gain merely to recover. Large percentage gains do not neutralize risk; they often disguise how much risk has accumulated.

The critical detail is that Citadel reportedly purchased most of the public-stock portfolio. That is not the same as a government rescue, and it does not establish that Situational Awareness was systemically important. It does show that an orderly block transaction can prevent a distressed holder from selling every position into the open market. A buyer with capital, trading infrastructure and the ability to evaluate a complex book can absorb risk that would otherwise be transmitted through repeated market orders.

Visser’s description of this as a “cleansing event” captures part of the mechanism. When a leveraged portfolio reaches a point where financing terms, margin requirements or liquidity needs force action, the fund’s investment horizon collapses. A position may still look attractive over three years, but that becomes irrelevant if cash is needed today. The buyer is no longer negotiating against the seller’s estimate of intrinsic value; the buyer is negotiating against the seller’s deadline.

There is, however, no basis for treating the unwind as proof that all weak holders have been removed. The wider AI trade involved hedge funds, systematic strategies, retail traders, leveraged exchange-traded products, corporate treasury decisions and index investors. One portfolio sale can reduce a specific overhang without eliminating crowded positioning across the market. Goldman Sachs described the cumulative selling of technology stocks in its data as historically large while still characterizing the move as a possible reset rather than abandonment of the AI theme. That is a useful distinction: a position can become less crowded without becoming cheap, and a theme can remain structurally important while its most aggressive expression fails.

July also showed why the phrase “fundamentals are unchanged” is not enough. Even when future cash-flow expectations do not immediately change, the discount rate, required risk premium and probability assigned to adverse outcomes can change sharply. A company expected to earn the same amount may deserve a lower price if investors demand a higher return, financing costs rise or the business becomes more difficult to model. In fast-moving technology, uncertainty itself has a price.

What happened inside Situational Awareness

Aschenbrenner became prominent through his 2024 essay series, Situational Awareness: The Decade Ahead. The work argued that rapid scaling in computing power and algorithmic progress made highly capable AI plausible on an unusually short timetable. It connected model development to national security, industrial capacity and the need for a vast buildout of electricity and computing infrastructure. Whatever one thinks of its timetable, the essay correctly directed attention toward physical bottlenecks that conventional software analysis often ignored.

The investment fund applied that worldview to public and private markets. Its reported holdings included companies connected to chips, memory, data centers, power and AI infrastructure. Reuters reported that the public-equity book was valued at roughly $16 billion before the transaction with Citadel. Other reports placed the combined public and private portfolio at a higher level, with a substantial private holding in Anthropic. These figures should not be blended casually: gross exposure, net asset value, private-company marks and the market value of listed securities are different measurements.

The investor letter, as reported, acknowledged that the fund came close to permanent capital impairment. That wording matters. Temporary mark-to-market losses become permanent when the fund must sell, when financing is withdrawn or when the underlying company’s economics deteriorate. A manager who can hold through volatility has optionality. A manager who has borrowed against the position has sold part of that optionality to the lender.

Concentration magnified the problem. A portfolio of companies exposed to the same capital-spending cycle can look diversified by ticker while remaining economically concentrated. A memory-chip manufacturer, a data-center developer, a networking supplier and a power-equipment company may have different products, but all can fall together if investors reduce their estimate of AI infrastructure demand or decide that financing costs make the buildout less valuable. Correlation tends to rise precisely when diversification is needed most.

Public disclosure can add another layer. Large U.S. equity holdings may become visible through regulatory filings, although those filings are delayed and incomplete representations of an active portfolio. Once market participants believe they know the positions of a distressed holder, they can reduce their own exposure, hedge related securities or wait for better prices. That does not prove manipulation. It describes rational behavior when a forced seller may exist.

The transcript raised the possibility that Citadel’s public comments about a potential Fed rate increase contributed to the crisis. There is no verified evidence that a policy forecast was designed to damage the fund. Rate-hike probabilities were already meaningful, inflation remained above target and the Federal Reserve’s new communication approach had deliberately preserved uncertainty. A market participant expressing a view about monetary policy is not evidence of a coordinated attack. The responsible conclusion is that a hawkish policy scenario may have accelerated losses in a vulnerable portfolio, while the intent alleged by online speculation remains unsupported.

Situational Awareness’s decision to eliminate leverage is therefore more consequential than the manager’s continued confidence in AI. It recognizes that a correct long-term thesis cannot rescue an unsustainable financing structure. Investors evaluating the episode should separate three questions: Was the AI infrastructure thesis broadly right? Were the selected securities purchased at prices that offered a margin of safety? Was the portfolio constructed so that it could survive being early, wrong or temporarily illiquid? A fund can be right on the first question and fail on the other two.

How a forced liquidation becomes a “speed crash”

Visser uses the term “speed crash” to describe the modern pattern of gradual appreciation followed by sudden collapse. The metaphor is apt because leverage creates asymmetry. Gains can accumulate slowly as collateral values rise and financing remains available. Losses accelerate when falling prices reduce collateral, trigger margin calls and force sales that push prices lower.

The process usually follows a recognizable sequence:

1. A popular theme produces strong returns. Success attracts capital, encourages larger positions and makes historical volatility look deceptively low.
2. Managers increase gross exposure. Borrowing or derivatives allow the portfolio to own more risk than its equity capital would otherwise support.
3. Different portfolios converge. Managers buy the same winners because the securities have strong earnings momentum, attractive narratives or favorable factor characteristics.
4. A catalyst changes the price path. The catalyst may be earnings, policy, a competitive threat, a funding concern or simply the exhaustion of buyers.
5. Risk models react. Higher volatility increases measured risk. Funds reduce positions to remain within limits even if their fundamental view is unchanged.
6. Prime brokers demand protection. Lenders raise margin requirements, restrict financing or request additional collateral.
7. Liquidity deteriorates. Other investors anticipate sales and become less willing to buy at previous prices.
8. Correlations rise. Securities that seemed different trade as one basket because the same holders are selling them.

This is why the first decline can be modest while the final phase is violent. The supply of shares for sale becomes insensitive to price. A forced seller does not stop because a stock is down 20%; the seller stops when the required amount of risk has been removed.

It also explains why a rebound can be equally sharp. Once the forced flow ends, normal buyers face less competition from distressed supply. Short sellers cover, market makers rebuild positions and investors who still believe the fundamental thesis return. The rebound does not prove the decline was irrational, just as the decline did not prove the businesses were broken.

Modern execution systems make this process faster, but AI should not be treated as the sole cause. Algorithmic trading, portfolio optimization, volatility targeting and electronic market making existed long before the current generation of AI agents. What AI can do is lower the cost of analysis and execution, broaden access to sophisticated tools and accelerate the speed at which similar information is incorporated into portfolios. If many users ask comparable systems to optimize around comparable data, technological sophistication can produce convergence rather than diversity.

The key market-structure risk is not that an AI agent becomes emotional. It is that many systems respond rationally to the same constraint. A volatility target, margin rule or loss limit can generate synchronized selling without panic in the human sense. Automation removes hesitation. That can improve discipline in normal conditions and intensify feedback loops in abnormal ones.

Lessons from LTCM, the 2007 quant meltdown and Archegos

Historical comparisons are useful only when the differences are made explicit. The July 2026 event shares features with Long-Term Capital Management, the August 2007 quant meltdown and Archegos Capital Management, but it is not a duplicate of any of them.

The 2007 quant episode is especially relevant. Research by Amir Khandani and Andrew Lo, available through the National Bureau of Economic Research, described how the unwinding of factor-driven portfolios began before the most dramatic losses and intensified as liquidity providers reduced risk. The event demonstrated that strategies built by different firms could behave like one large portfolio because they used similar signals and owned similar securities.

Archegos illustrates the danger of fragmented prime-broker information. The family office used derivatives with several banks, allowing it to build concentrated exposure without each counterparty initially seeing the whole picture. When Archegos could not meet margin calls, banks sold the underlying shares. Some moved faster than others, and the slower institutions suffered enormous losses. A Credit Suisse report filed with the Securities and Exchange Commission described gross exposure of roughly $120 billion and remaining equity of only $9 billion to $10 billion when the crisis reached its decisive stage.

Situational Awareness appears different in scale, instruments and counterparty impact. The public information does not show a hidden network of swaps that threatened multiple banks. The portfolio sale to Citadel may have reduced market impact precisely because a capable buyer absorbed positions before a chaotic open-market liquidation. Calling it “another LTCM” would therefore overstate what is known.

The common lesson is more modest and more durable: liquidity is a portfolio characteristic, not a permanent property of a security. A stock that trades billions of dollars on an ordinary day can become difficult to sell in size when several large holders need to exit together. The relevant question is not average daily volume in calm conditions. It is how much can be sold during stress without moving the price enough to create additional margin pressure.

Why South Korea became the most extreme expression of the trade

South Korea’s market turned the global AI reversal into a retail-leverage crisis. The KOSPI had been one of the strongest major equity markets, supported by enthusiasm for memory chips, semiconductor equipment and the country’s central role in the AI supply chain. That strength drew investors into leveraged products and margin-financed accounts.

When the trend reversed, the scale of the borrowing mattered. Reuters reported that margin-loan balances reached a record 38.63 trillion won on June 24. Goldman Sachs estimated that more than 1.2 million leveraged retail accounts had received margin calls by July 13, with hundreds of thousands fully liquidated. Those figures turned a decline in semiconductor shares into a household financial event.

The KOSPI fell 10.84% on July 28, its worst day in months, as Samsung Electronics and SK Hynix each dropped more than 14%, according to Reuters. The following session brought another large decline. The concentration of the index magnified the move because Samsung and SK Hynix represented an exceptionally large share of benchmark weight.

The most important corrective to the “collapse” narrative is the starting point. The Korean market had risen dramatically before the selloff. A 25% or 30% decline after a huge advance can erase trillions of dollars in market value while leaving long-term holders with gains. That does not make the losses harmless. It shows why percentage changes must be anchored to a time period.

South Korea also demonstrates why a market can experience financial stress while the underlying industry reports strong demand. On August 1, Reuters reported that July exports rose 62.8% from a year earlier, with semiconductor exports up 179% and computer exports up 404%. The figures were driven by AI-related demand and higher memory prices. The equity market’s fall was therefore not a simple reflection of collapsing current orders.

Instead, investors were reassessing how much future growth had already been priced in, whether Chinese competitors could pressure margins and how much leverage had to be removed. A company can report record profits and still fall if expectations were higher. A cyclical semiconductor company can face both scarcity today and oversupply risk later. Memory markets have historically moved between shortage and glut because high prices encourage investment, while new capacity arrives with a delay.

The Korean experience should also caution U.S. investors against assuming that retail leverage is only an overseas issue. The products differ, but single-stock options, leveraged ETFs, margin loans and zero-day index options can create similar nonlinear exposure. The U.S. household sector may have greater aggregate wealth, as Visser noted, but aggregate wealth does not protect individual accounts from concentrated leverage.

Japan and the global transmission of an AI-factor reversal

Japan’s market was less dependent on retail margin than South Korea’s, but it was deeply exposed to the same global repricing. Semiconductor-equipment companies, technology conglomerates and exporters had benefited from the AI investment cycle and a favorable currency environment. When investors questioned the durability of AI spending, those shares became sources of liquidity.

On July 17, Japan’s Nikkei entered correction territory, falling more than 10% from its June 25 record close as the global technology selloff intensified, according to Reuters. A later 2.73% decline on July 24 showed that the pressure had not disappeared even as individual sessions stabilized.

Japan matters to U.S. markets for several reasons. Its companies supply critical equipment and materials to the semiconductor industry. Japanese institutions are major owners of global bonds and equities. Currency moves can change the value of overseas assets and the economics of hedging. A stronger yen may pressure exporters while encouraging Japanese investors to repatriate capital; a weaker yen can support profits but raise import costs.

The global nature of the selloff also exposes the limitations of country-level explanations. Chinese progress in memory and lithography was cited as a threat to Korean and Japanese incumbents. U.S. hyperscaler spending shaped demand for Asian chips. Federal Reserve uncertainty affected global discount rates. The same trade connected data centers in the United States, memory fabrication in Korea, equipment suppliers in Japan and energy infrastructure across multiple regions.

This is why the phrase “AI bubble” can be too imprecise. There is no single AI asset. Model developers, cloud providers, chip designers, memory producers, foundries, equipment companies, power suppliers and software vendors have different margins, capital needs and competitive risks. A correction in one layer may benefit another. Lower chip prices could hurt producers while improving economics for model developers. More efficient models could reduce compute per task while expanding total usage enough to increase aggregate demand.

The market’s mistake in euphoric periods is to price every layer as a winner. Its mistake in panics is to sell every layer as if the economics were identical. The analytical task is to identify where scarcity creates durable bargaining power and where high prices will attract enough supply to destroy that advantage.

Market concentration turned an industry correction into an index event

The AI unwind was amplified by the way modern indexes and portfolios concentrate exposure. Market-capitalization-weighted benchmarks give the largest weights to the companies whose values have already risen the most. That is not inherently irrational; it allows investors to hold the market without making active judgments. But when one investment theme drives several of the largest companies and their suppliers, a thematic reversal becomes an index-level event.

Concentration operates through more than index weights. Momentum strategies favor securities with strong recent returns. Quality strategies may favor profitable technology companies. Growth strategies favor long-duration earnings. Retail traders may choose the same names because they are familiar. Hedge funds may hold the shares long while shorting weaker companies. Different mandates therefore converge on a common set of exposures for different reasons.

During the advance, the flows reinforce one another. Rising prices increase index weights, attract momentum capital and improve collateral values. Companies can issue securities at favorable prices and fund more investment. Analysts raise estimates as demand strengthens. The market creates a feedback loop between price, financing and fundamentals.

During the reversal, the loop changes direction. Volatility-targeting funds reduce risk. Leveraged investors meet margin calls. Active managers protect gains. Index investors are not necessarily forced to sell simply because prices fall, but redemptions from technology-focused funds create additional supply. A security can therefore decline even when passive benchmark holders remain in place.

July’s divergence between equal-weighted and technology-heavy benchmarks showed the difference between the market and its most crowded leadership. MarketWatch reported that the equal-weighted S&P 500 outperformed the Nasdaq-100 by an unusually wide margin during the month. That pattern suggests rotation rather than a uniform collapse in economic expectations.

Rotation can be healthy because it broadens market participation. It can also conceal stress. A capitalization-weighted index may appear stable if a few cloud companies rise on earnings while dozens of AI suppliers fall. Conversely, an index can fall because of its largest weights even when the median company is stable.

Investors evaluating “the market” should therefore examine several lenses: capitalization-weighted indexes, equal-weighted indexes, sector performance, factor returns and individual-company breadth. No single measure fully describes the event.

Concentration also complicates diversification through funds. An investor may own an S&P 500 fund, a technology fund, a semiconductor fund and a growth fund while believing these are four allocations. In practice, the largest holdings and economic drivers can overlap substantially. The portfolio owns the same risk through several wrappers.

The Korean market demonstrated an even more direct form of concentration because Samsung Electronics and SK Hynix represented a very large share of the KOSPI. When both fell more than 14% in one session, the benchmark could not avoid a historic decline. The index move then triggered risk controls and margin processes tied to the benchmark, feeding the cycle.

Concentration is not automatically a reason to avoid an asset. The largest companies may be large because they produce exceptional cash flows and possess durable advantages. The risk arises when portfolio construction assumes that recent correlations and liquidity will remain stable. A concentrated winner can continue winning; it must still be sized so that a reversal does not determine the investor’s solvency.

The strongest skeptical case against the AI infrastructure boom

A balanced analysis should not treat strong cloud earnings as the final answer. The skeptical case begins with the possibility that today’s demand includes strategic over-ordering, duplicated capacity and financing relationships that make the ecosystem look more independent than it is.

Frontier-model developers need enormous compute commitments to train and serve models. Cloud companies invest in those developers, provide capacity and recognize cloud revenue as the developers use contracted services. The relationships can be commercially rational, but they complicate analysis. An investor should ask how much demand comes from end customers generating positive returns and how much comes from capital raised inside the AI ecosystem.

This is not an accusation of improper accounting. It is a question about the durability of the customer’s funding. If a model developer depends on repeated equity financing to purchase compute, cloud revenue is ultimately supported by investor capital until the developer produces enough operating cash flow. The chain can grow rapidly while capital is abundant and slow sharply when financing becomes expensive.

The second skeptical point is depreciation. AI hardware loses economic value quickly as new generations improve performance. A company may operate the equipment for many years, but the competitive revenue it can earn may decline sooner. Accounting estimates of useful life affect reported profit. Cash returns depend on actual utilization and pricing.

The third point is customer bargaining power. The largest buyers of compute are technically sophisticated and can shift workloads, design custom chips or negotiate long-term pricing. A shortage gives suppliers leverage today, but the customers are actively investing to reduce dependence. The more money a bottleneck extracts, the stronger the incentive to bypass it.

The fourth point is energy economics. Data centers may secure power through long-term contracts, dedicated generation or grid investment. Those arrangements transfer costs but do not remove them. If electricity prices rise or communities require infrastructure contributions, the return on data-center capital can fall.

The fifth point is utilization. A data center announced in a press release is not necessarily operating at full economic capacity. Construction delays, chip shortages, networking issues and customer timing can leave assets underused. Investors need disclosure about deployed capacity and revenue, not only planned gigawatts.

The sixth point is software competition. Open-source and lower-cost models can reduce the price customers are willing to pay. Even if total token demand rises, revenue per token may fall. Cloud providers can benefit from volume, while independent model developers face margin pressure. The value may migrate from models to applications, data or distribution.

The seventh point is macro sensitivity. AI infrastructure has a long duration. Much of the expected return arrives years after the cash is spent. Higher real interest rates reduce present value and increase financing cost. A project can be technologically successful and financially disappointing if the cost of capital changes.

The eighth point is industrial policy. Governments view AI capacity as strategic and may subsidize domestic production. Subsidies can accelerate supply beyond what private returns justify. They can also fragment the market through export controls and local-content rules, raising costs and reducing economies of scale.

None of these arguments requires AI adoption to fail. The railway, telecommunications and internet booms created valuable infrastructure while many investors lost money. Society can benefit from overbuilding because future users receive cheap capacity. Shareholders who financed the overbuilding may not receive the social return.

The skeptical conclusion is therefore not “AI is fake.” It is that the current infrastructure race can be economically transformative and still produce weak returns for some suppliers. The more certain the market becomes about the technology, the more important price and capital discipline become.

What happened after the worst of the selling

By the end of July, several developments reduced immediate panic. Citadel’s transaction gave the market a clearer path for the Situational Awareness portfolio. Microsoft’s earnings demonstrated strong cloud demand and an enormous contracted backlog. Amazon followed with accelerating AWS growth and higher investment plans. Korean semiconductor shares stabilized after their most violent sessions, while export data confirmed robust current demand.

Those developments support the view that forced selling had pushed some prices beyond the change in near-term fundamentals. They do not restore the previous valuation regime. Investors now have evidence that the demand side is strong and that the capital cost is rising. The market must decide which companies can convert one into returns after paying for the other.

The rebound also creates a behavioral risk. Investors may interpret stabilization as proof that leverage was harmless and rebuild exposure too quickly. A market can experience several liquidation waves as different holders reach limits at different times. Private-asset marks and quarterly redemptions can delay the transmission of stress.

At the research cutoff on August 1, no complete public map of hedge-fund exposure existed. It was therefore too early to declare the event finished. The more defensible statement was that a major known overhang had been transferred, the fundamental demand data remained strong and uncertainty about financing and valuation persisted.

Does the compute-scarcity thesis survive the selloff?

Visser’s strongest argument is that the physical requirements of AI remain underestimated. Software can be copied at low marginal cost, but AI inference and training depend on semiconductors, memory, networking, data-center construction, cooling and electricity. Those inputs cannot be created instantly. Permitting, grid interconnection, transformer production and advanced-chip fabrication operate on timelines measured in years.

The latest earnings reports support the claim that demand remains strong. They do not prove that every dollar of capital expenditure will earn an attractive return.

Amazon’s second-quarter release reported net sales of $200.6 billion, operating income of $27.5 billion and AWS sales growth of 37%, the fastest in 18 quarters. The company raised its investment plans as management said demand continued to exceed available capacity. This is direct evidence against the idea that AI infrastructure has already become universally overbuilt.

Microsoft’s fiscal fourth-quarter results were similarly strong. Microsoft Cloud revenue rose 27% to $59.3 billion, and commercial remaining performance obligations increased 84% to $678 billion. The company said growth excluding OpenAI commitments was still substantial. Backlog is not revenue, and contracts can contain conditions, but the scale of commitments demonstrates that enterprise and model-company demand is not merely a social-media narrative.

Alphabet reported that Google Cloud revenue rose 82% to $24.8 billion, led by enterprise AI solutions, AI infrastructure and core cloud services. Its growth rate supports Visser’s view that demand has surprised even major suppliers. Yet Alphabet’s capital raising and spending also show the financial burden of staying competitive.

Meta’s results provide the skeptical counterexample. Revenue increased 28%, but costs and expenses rose 55%, operating income declined and free cash flow fell to less than billion for the quarter. Meta raised the lower end of its full-year capital-expenditure forecast to $130 billion. The company may ultimately earn excellent returns through advertising, personal agents, wearables or enterprise services. Current results show that the transition can pressure cash generation long before new business models mature.

The conclusion is not that compute is scarce or abundant in an absolute sense. The relevant question is scarce at what price and for which workload? A shortage can coexist with poor returns if companies pay too much for equipment, if hardware depreciates faster than expected or if customers gain bargaining power. Conversely, falling unit costs can expand usage so rapidly that total revenue rises. The economics resemble transportation and telecommunications: cheaper capacity can destroy margins in one segment while creating an enormous market in another.

Backlog is evidence of demand, not a guarantee of economic value

The interview referenced combined hyperscaler backlogs approaching $2 trillion. The broad point is directionally plausible, but the figures should not be added without caution. Companies define backlog and remaining performance obligations differently. Contract duration, cancellation rights, pricing adjustments, customer concentration and revenue-recognition rules vary. A dollar of cloud RPO is not automatically comparable with a dollar of another company’s contracted orders.

Backlog can still be useful. It provides evidence that spending is supported by customer commitments rather than built entirely on speculation. It may improve confidence that facilities will be utilized. It can also allow suppliers to finance expansion against future revenue.

But four risks remain.

First, customer concentration. A meaningful portion of demand may come from a small number of frontier-model developers. If one customer restructures, delays a project or renegotiates pricing, the effect on the supplier can be larger than the headline backlog suggests.

Second, capital intensity. Revenue growth can look impressive while free cash flow deteriorates because the company must purchase chips and build data centers before revenue is recognized. Investors must compare incremental operating profit with incremental invested capital, not just celebrate sales growth.

Third, technological depreciation. AI accelerators can become economically obsolete before they stop functioning. A data center designed for one generation of hardware may require expensive upgrades. Extending accounting lives reduces annual depreciation expense but does not guarantee that assets will remain competitive.

Fourth, price competition. Scarcity supports pricing until enough capacity arrives or customers become more efficient. Open-source models, specialized chips and algorithmic improvements can reduce the amount of expensive frontier compute needed for some tasks. The result may be a shift from scarcity rent to commodity economics.

Visser’s “tokens per watt” framework is valuable because it connects model quality to energy efficiency. Customers will care not only about the intelligence of a model, but about the economic output produced per unit of power and hardware. A model that completes a task with fewer tokens, lower latency and less energy can be more valuable even if its benchmark score is similar.

Still, efficiency does not necessarily reduce aggregate compute demand. The Jevons paradox—where improved efficiency lowers the cost of using a resource and increases total consumption—may apply. Cheaper inference could make it economical to deploy agents in customer service, coding, design, logistics, finance and personal devices. Total tokens consumed can rise even as tokens per task fall.

That is the core bull case for infrastructure: the addressable market expands faster than efficiency improves. The bear case is that capital investment races ahead of monetization, competition compresses prices and hardware becomes obsolete before earning its cost of capital. Both can be true at different points in the cycle.

Why “AI will destroy all public companies” is too broad

Visser’s most dramatic corporate claim was that AI would “destroy” the growth certainty of public companies. Interpreted narrowly, the claim has merit. AI lowers the cost of software development, research, marketing, customer support and business formation. New competitors can emerge faster, and established companies may have less confidence that current margins will persist for a decade. Greater uncertainty can justify lower valuation multiples.

Taken literally, however, the claim is unsustainable. AI does not eliminate physical assets, distribution networks, regulatory licenses, brands, customer relationships, proprietary data or economies of scale. It can strengthen incumbents that possess these advantages. Large companies can buy technology, train models on unique data and spread infrastructure costs across enormous revenue bases.

The effect will vary by industry. Software businesses with high margins and low switching costs may face direct pressure from AI-native products. Regulated utilities may use AI to improve operations without losing their franchise. Manufacturers can gain productivity while remaining protected by engineering, capital and supply-chain constraints. Financial institutions can automate analysis, but trust, balance-sheet capacity and regulation remain barriers to entry.

Recent research on enterprise adoption also suggests a gradual process rather than universal overnight disruption. A 2026 study of S&P 500 companies using regulatory filings found that deep integration had expanded, especially among technology firms, but remained far from universal. That is consistent with a J-curve: costs appear before productivity benefits, and the transition can temporarily depress profitability.

The more defensible investment implication is that duration deserves a larger uncertainty discount. Investors should be cautious about paying high multiples for cash flows expected far in the future when AI can change market structure quickly. That does not mean every multiple must fall. A company that owns scarce infrastructure or gains market share through AI may deserve a premium. The key is whether the moat becomes stronger or weaker as intelligence becomes cheaper.

This also explains why “scarcity” became Visser’s preferred category. Power equipment, advanced memory, grid connections and certain manufacturing capabilities cannot be reproduced as quickly as software. Yet scarcity is not permanent. High returns attract capital, governments subsidize capacity and customers redesign products to use less of the expensive input. The investor’s challenge is to distinguish a bottleneck with durable pricing power from a temporary shortage near the top of a cycle.

Kevin Warsh’s Federal Reserve adds uncertainty by design

The Federal Reserve’s July decision became another stress point because markets entered the meeting without the level of guidance they had come to expect. Under Chairman Kevin Warsh, the central bank has shortened its policy communication and reduced explicit signals about the next move.

On July 29, the Federal Open Market Committee voted 9–3 to maintain the federal-funds target range at 3.5% to 3.75%. Three members preferred a quarter-point increase. The official statement emphasized the dual mandate and the policy of maintaining ample reserves but offered little forward guidance.

Warsh’s approach is not simply indecision. He has argued that excessive guidance can encourage markets to outsource judgment to the central bank, suppress useful volatility and create a false sense of certainty. The Fed has also launched task forces to review its analytical tools, communication and operating framework, as described in a July 9 announcement.

There is a legitimate case for flexibility. Economic data can change, forecasts are uncertain and promises about future policy may become constraints. A central bank that precommits too strongly can lose credibility when conditions require a different action.

There is also a cost. Less guidance increases the distribution of possible outcomes before each meeting. Investors demand compensation for that uncertainty, particularly in long-term bonds. After the July decision, the yield curve steepened as long-term yields rose. Markets appeared unconvinced that holding rates steady was fully consistent with the Fed’s stated determination to control inflation.

For leveraged portfolios, this communication regime matters even when the policy rate does not change. A wider range of expected outcomes raises volatility in Treasuries, currencies and equities. Higher bond yields reduce the present value of long-duration growth stocks. More volatile collateral can prompt lenders to tighten terms. The Fed can therefore contribute to deleveraging through uncertainty without raising rates.

Visser approved of the decision to wait, arguing that markets had already tightened financial conditions through higher yields. That is a reasonable interpretation. The policy rate was below some market yields, inflation had shown modest improvement and the Fed could observe additional data. The opposing view is that allowing long-term yields to rise because investors doubt the inflation response is not a benign substitute for deliberate tightening. Market tightening caused by credibility concerns can be more damaging than a clearly explained policy action.

The transcript also connected Warsh’s desire for a smaller Fed balance sheet with tokenization and the release of “dormant” asset value. That chain is speculative. Tokenization can improve settlement and collateral mobility, but it does not create net financial wealth. Making an asset easier to trade can raise liquidity and potentially reduce a liquidity discount; it can also increase volatility and leverage. The size of the Fed’s balance sheet is determined by monetary-policy implementation, currency demand, reserve needs and financial-stability considerations—not simply by whether private assets become tokenized.

Would a smaller Fed balance sheet reduce the K-shaped economy?

The interview argued that quantitative easing and the expansion of the Fed’s balance sheet contributed to a K-shaped economy by supporting asset prices, benefiting households that already owned stocks and real estate. This criticism has a factual basis: asset purchases lower term premiums and can raise the value of financial assets. Wealth gains are uneven because ownership is uneven.

But the distributional effect of monetary policy cannot be reduced to asset prices. The Fed’s emergency actions during the financial crisis and pandemic also supported employment, credit markets and economic activity. A deeper recession would have harmed lower-income households through job losses. The relevant counterfactual is not a world where asset prices stay lower while employment remains unchanged.

A smaller balance sheet may reduce some market distortions, but it could also raise term premiums and borrowing costs. Mortgage rates, corporate financing and government interest expense may increase. Whether Main Street benefits depends on inflation, wages, employment, housing supply and fiscal policy—not merely the quantity of securities held by the central bank.

The most credible version of Warsh’s argument is institutional. The Fed should not become the default buyer whenever markets fall. Private investors should bear normal investment losses, and central-bank support should be reserved for dysfunction that threatens the transmission of monetary policy or financial stability. That principle can reduce moral hazard.

The challenge is defining “normal.” A market can move sharply without becoming dysfunctional. Conversely, a modest price move can hide stress in funding markets. The Fed must distinguish solvency problems from liquidity problems and must communicate why intervention is or is not justified.

For investors, the implication is that the post-2008 assumption of a predictable “Fed put” may be less reliable. That supports Visser’s expectation of more frequent speed crashes. It does not necessarily imply lower long-term returns, but it favors strategies that can survive volatility without forced selling.

Is Bitcoin really “the best hedge fund ever”?

The title of the source video is memorable because Bitcoin’s historical returns resemble the type of outcome an extraordinarily successful investment fund would seek. But the analogy breaks down under examination.

A hedge fund is an organization. It has a manager, mandate, custody arrangements, risk controls, fees, investors and a process for allocating capital. Bitcoin is a digital asset governed by a protocol and a distributed network. It generates no corporate cash flow, does not rebalance itself and does not hedge risk in the ordinary sense. Its price reflects supply, demand, liquidity, regulation, macro conditions and beliefs about future adoption.

Bitcoin’s strongest investment characteristic is verifiable scarcity. Its maximum supply is defined by the protocol, and ownership can be transferred without relying on a central issuer. That gives it a plausible role as a non-sovereign store-of-value candidate. The word “candidate” matters because a store of value must be evaluated through behavior, not design alone.

At approximately 3:30 p.m. Eastern Time on August 1, Bitcoin traded near $62,500, with a market value around $1.25 trillion. It had experienced another large decline from prior highs. BlackRock’s analysis of historical volatility notes that Bitcoin has suffered several drawdowns exceeding 50%, with the largest averaging roughly 80%. Academic research has found Bitcoin volatility far higher than major currencies.

That history does not disprove long-term appreciation. It does disqualify Bitcoin from being described as a reliable short-term hedge against ordinary portfolio losses. A hedge should offset a defined risk with some consistency. Bitcoin has sometimes risen during monetary expansion and currency concerns, but it has also fallen during periods of tightening, deleveraging and risk aversion. It often behaves like a high-beta liquidity asset.

There are at least four different meanings hidden inside the word “hedge”:

• Inflation hedge: an asset expected to preserve purchasing power when consumer prices rise.
• Currency-debasement hedge: an asset expected to benefit from growth in fiat-money supply or declining confidence in sovereign liabilities.
• Tail-risk hedge: an asset expected to rise during market crises.
• Long-horizon scarcity allocation: an asset held because its supply cannot be increased in response to demand.

Bitcoin’s evidence is strongest for the fourth category and mixed for the first three. It may perform well over long periods when adoption grows faster than supply, but that is an investment thesis, not a mechanical hedge.

The “best hedge fund” phrase is therefore better understood as a rhetorical comparison to compounding. Bitcoin has survived repeated drawdowns and recovered to new highs in previous cycles. A fund that produced similar long-term returns would be celebrated. Yet survivorship should not be confused with risk control. Bitcoin holders endured volatility that many funds could not tolerate because funds face redemptions, leverage limits and fiduciary obligations.

What Michael Saylor and Strategy actually demonstrate

Visser invoked Michael Saylor’s decision to place Bitcoin at the center of MicroStrategy—now Strategy—as evidence that businesses eventually seek a store of value when cash is eroded by low interest rates and competitive pressure. The history is important, but the lesson is not one-directional.

Strategy transformed from an enterprise-software company with a large cash balance into the world’s largest corporate Bitcoin holder. The company used equity, convertible debt and preferred securities to acquire Bitcoin. Its public ledger provides transaction-level details, and SEC filings document the scale and financing of the holdings.

The strategy created extraordinary shareholder gains when the company’s equity traded at a premium to the value of its Bitcoin. Issuing stock above net asset value allowed the company to acquire more Bitcoin per share under favorable conditions. The model depended on capital-market access, investor demand and a persistent premium.

By 2026, the risks were equally visible. Strategy reported a large quarterly loss as Bitcoin prices fell and had to manage preferred dividends, debt and liquidity. Its holdings remained enormous, but the equity price reflected both the value of Bitcoin and the complexity of the capital structure. A corporate treasury strategy can increase exposure to a scarce asset while also increasing financing risk.

Saylor’s case therefore proves that a company can use its balance sheet to become a leveraged Bitcoin vehicle. It does not prove that every operating business should do so. A company needs cash for payroll, taxes, inventory, acquisitions and downturns. Holding a highly volatile asset can create a mismatch between liabilities due in dollars and assets priced by a speculative market.

The relevant comparison is opportunity cost. A mature business with excess cash and limited reinvestment opportunities may evaluate Bitcoin, bonds, buybacks, dividends or acquisitions. A growing business with high-return projects may destroy value by diverting capital into an unrelated asset. The correct decision depends on the company’s liabilities, risk tolerance, governance and access to financing.

It is also wrong to treat cash debasement as a simple, continuous process. Cash loses purchasing power when inflation is positive, but it provides nominal stability and optionality. Short-term Treasury instruments can generate yield. Bitcoin offers scarcity but no contractual income. The two assets solve different problems.

AI agents could expand digital-asset use—but not for the reason enthusiasts assume

One of Visser’s most interesting forecasts is that AI agents will evaluate Bitcoin more dispassionately than human investors. An agent might examine historical returns, scarcity and correlations, then assign a probability-weighted allocation rather than rejecting the asset because it feels unconventional.

That is possible, but it assumes the agent’s objective function, constraints and data lead to that result. A fiduciary agent may penalize Bitcoin for volatility, drawdowns, uncertain valuation and regulatory risk. A tax-aware agent may reduce trading. A risk-parity agent may allocate very little. An aggressive long-horizon agent may allocate more. AI does not eliminate judgment; it encodes judgment in goals, data and constraints.

The more immediate connection between AI and crypto is payments. Software agents need authorization, identity, spending limits, dispute resolution and settlement. Blockchain networks and stablecoins can provide programmable transfer, global availability and machine-readable transaction records. Traditional payment networks are building agentic tools as well, so crypto is not the only solution.

Early evidence suggests agent payments remain small relative to global finance. Industry research cited by CoinDesk estimated tens of millions of dollars in blockchain settlement by AI agents across a very large number of low-value transactions. The number demonstrates experimentation, not mass adoption.

Stablecoins may be more natural for agents than volatile assets because the unit of account remains stable. An agent purchasing compute, data or software services generally wants predictable costs. Bitcoin may function as reserve collateral or long-term savings, while stablecoins handle transactions. Ethereum and other programmable networks may provide settlement and smart-contract functionality. These roles should not be conflated.

Security is the central obstacle. An autonomous agent with a wallet can make mistakes at machine speed. It can be manipulated by malicious instructions, compromised credentials or false data. Reversible card payments and bank controls exist partly because errors and fraud occur. Blockchain finality is useful when settlement certainty matters and dangerous when authorization fails.

For agentic finance to scale, the system needs bounded autonomy: transaction limits, approved counterparties, multi-party authorization, auditable intent, revocation and insurance. The future may combine cryptographic settlement with regulated identity and traditional dispute mechanisms rather than replacing one system with another.

Tokenization is real, but the grandest claims remain unproven

Tokenization means representing ownership or claims on a shared digital ledger so that transfer, settlement and contractual rules can be executed programmatically. It can apply to money-market funds, bonds, private credit, real estate, bank deposits and other assets.

The institutional case has strengthened. The International Monetary Fund has described how tokenization can combine the transfer of an asset and payment on a shared ledger, reducing reconciliation and settlement delays. BlackRock Chairman Larry Fink has argued that tokenization could modernize market plumbing and broaden access. Major banks and asset managers are developing tokenized funds and transfer-agency systems.

The benefits are plausible:

• Faster settlement and reduced counterparty exposure.
• Fractional ownership of assets with high minimum investments.
• Programmable compliance and automated corporate actions.
• Twenty-four-hour transfer within approved networks.
• Better use of collateral across financial institutions.
• Potential access to private-market exposures through regulated products.

But tokenization does not remove the underlying legal and economic problems. A token representing real estate still depends on property law, title, maintenance and local regulation. A tokenized private-company share still depends on valuation, disclosure and shareholder rights. Fractionalization can improve access while distributing illiquid or overpriced assets to less sophisticated investors.

Liquidity cannot be created by software alone. A market needs buyers, sellers, transparent information and confidence in the legal claim. Trading a token around the clock does not guarantee a fair price. In stressed conditions, continuous markets can transmit volatility rather than reduce it.

Tokenization also does not “release money supply” in the sense of creating net purchasing power. It can increase collateral velocity, lower settlement frictions and make assets easier to borrow against. Those changes may increase effective liquidity and leverage. Whether that is beneficial depends on safeguards. A system that makes every asset instantly financeable can improve capital efficiency and make margin spirals faster.

The likely path is gradual coexistence. Traditional records and tokenized ledgers will operate together for years. Regulated institutions will adopt the technology first where reconciliation costs are high and legal ownership is clear. The revolutionary language may be ahead of the operational reality, but the infrastructure shift is no longer hypothetical.

Can tokenization shift wealth from public markets to small businesses?

The interview proposed that AI-native businesses with very few employees could capture a larger share of global wealth and that tokenized products could give investors access to private companies. This is one of the more compelling long-term possibilities.

AI lowers the fixed cost of creating a business. A small team can produce software, marketing, customer support and analysis that once required a much larger organization. If revenue per employee rises, founders may retain more ownership and reach profitability without repeated venture-capital rounds. That could distribute business ownership more broadly.

There are counterforces. AI infrastructure has enormous fixed costs, and access may be controlled by a small number of cloud and model providers. Distribution platforms can capture much of the value created by small businesses. Successful founders may still concentrate wealth, and the majority of new businesses may fail.

Tokenization could broaden investment access, but it can also weaken investor protection if products are sold before disclosure and valuation standards mature. Public markets impose reporting, governance and liquidity requirements for a reason. Private markets offer potential returns partly because information is limited and capital is locked up.

The future may not be a simple transfer from public companies to private founders. Large public platforms could become the infrastructure on which millions of small AI-native businesses operate. Value would be divided among model providers, cloud companies, payment networks, marketplaces and entrepreneurs. The bargaining power of each layer would determine who captures the surplus.

For workers, the transition is equally ambiguous. One person may be able to operate several businesses, but competition among AI-enabled entrepreneurs could compress prices. Greater productivity does not automatically produce greater median income. Policy, education, market concentration and ownership will influence distribution.

What the latest earnings say about the “AI bubble” debate

A bubble is not defined merely by high prices or transformational technology. It involves expectations and financing that become detached from plausible future cash flows. Transformative technologies can produce bubbles because their long-term importance makes short-term valuation difficult.

The latest earnings provide evidence for both camps.

The bullish evidence is concrete: cloud revenue growth accelerated; contractual commitments expanded; enterprise adoption increased; memory demand strengthened; and the largest companies continued to invest because they reported insufficient capacity. These are not the conditions of a theme supported only by imagination.

The skeptical evidence is equally concrete: capital spending consumed free cash flow; some stocks had priced in extraordinary growth; debt and equity issuance expanded; competitive threats from Chinese models and open source increased; and the market was highly concentrated in a narrow set of winners. A real technology can still be a bad investment at the wrong price.

Investors should therefore separate three bubbles that are often treated as one:

• Capability expectations: Are models improving fast enough to justify claims about economic transformation?
• Infrastructure investment: Will demand use the data centers and chips being built?
• Security valuation: Does the price of a specific stock offer a reasonable return after accounting for spending and competition?

A person can be bullish on AI capability, bullish on total compute demand and bearish on a particular stock. That combination is not contradictory. It may be the most rational response to a technology whose benefits are widespread but whose profits could be captured unevenly.

Claim by claim: What the interview gets right, overstates or leaves unresolved

The source conversation moves quickly across hedge-fund mechanics, semiconductor demand, monetary policy, Bitcoin and tokenization. That creates useful connections, but it also makes it easy to treat a forecast as if it were an observed fact. A claim-by-claim review helps preserve the strongest insights without accepting the entire narrative.

The first claim is the easiest to accept in a limited form. Citadel’s purchase appears to have reduced the need for open-market selling. Prices rebounded in several AI-linked securities after the immediate pressure eased. Yet the word “cleansing” can encourage false confidence because it implies the source of instability has been removed. The better description is that one known leverage problem was resolved through a transfer of risk.

The compute claim has stronger current evidence than many bubble skeptics acknowledge. The reported growth rates at AWS, Azure and Google Cloud are too large to dismiss as accounting noise. The crucial caveat is that demand is being met through one of the largest capital-investment programs in corporate history. The relevant shareholder question is not whether customers want AI; it is whether suppliers can charge enough, for long enough, to recover their investment and earn an excess return.

The claim about public-company destruction is best translated into a valuation argument. When technological change accelerates, the confidence interval around distant cash flows widens. A lower multiple may be rational even if near-term earnings rise. But lower certainty is not identical to lower value. A company with a stronger balance sheet, proprietary data and distribution can use AI to take share from weaker competitors.

The crypto claim should be divided into transaction demand and investment demand. Agents may create transaction demand for programmable dollars or tokenized deposits because machines operate continuously and can settle small payments globally. Investment demand for Bitcoin requires a separate decision about risk, expected return and custody. An agent that follows conservative institutional rules may be less willing—not more willing—to hold an asset with repeated 50% drawdowns.

The tokenization claim contains a common category error. Liquidity and wealth are related but not identical. Making an asset easier to divide and trade can reduce transaction costs and may increase its price. It does not change the underlying cash flow, productive capacity or legal liability. An empty building does not become more productive because ownership is represented by one million tokens instead of one deed.

Finally, the Bitcoin comparison is useful as a challenge to conventional thinking. An asset with a fixed supply and global network can produce returns that traditional valuation methods struggle to explain. The comparison becomes misleading when it substitutes historical appreciation for the functions of portfolio management. A hedge fund can hold cash, reduce risk, short securities and return capital. Bitcoin can only be held, transferred or sold.

The physical economics of compute scarcity

Compute is often discussed as if it were a single commodity. In reality, the AI supply chain contains several bottlenecks with different economics. A shortage in advanced accelerators can coexist with excess capacity in older chips. Memory can become the constraint after processing improves. A completed data-center building can remain unusable because power is unavailable. Understanding the layers helps explain why a broad AI selloff can create both genuine bargains and value traps.

Advanced processors and manufacturing capacity

Frontier training depends on advanced accelerators manufactured through a highly concentrated supply chain. Designing a competitive chip requires specialized intellectual property and engineering. Manufacturing it requires leading-edge fabrication, advanced packaging and access to sophisticated equipment. Capacity cannot be expanded by simply ordering more machines; factories take years to build, qualify and operate at acceptable yields.

That supports high prices and strong margins during shortages. It also invites substitution. Cloud providers design custom chips, model developers optimize software for alternative hardware and customers shift less demanding workloads to cheaper processors. The most valuable supplier may be the one that controls a bottleneck no customer can redesign around.

High-bandwidth memory

AI accelerators need large quantities of fast memory to move data efficiently. The surge in Korean semiconductor exports and the importance of SK Hynix and Samsung to the KOSPI illustrate how memory became a macroeconomic variable. Memory producers benefit when demand exceeds supply, but the industry remains cyclical. Capacity additions, process improvements and customer inventory can turn shortages into gluts.

Investors should therefore distinguish spot scarcity from structural scarcity. A structural bottleneck has persistent barriers, long replacement times and limited substitution. A cyclical shortage produces exceptional profits that attract enough investment to normalize returns. The market frequently overvalues the current margin at the peak and undervalues the balance sheet near the trough.

Networking and interconnection

Large AI clusters depend on networking equipment that can move data between processors with low latency. The value of a cluster is not simply the number of chips; it is the efficiency with which the system uses them. Networking failures can leave expensive hardware idle. This makes switches, optical components and interconnect technology critical, but it also exposes suppliers to rapid product cycles.

Electricity, transformers and grid access

Power may be the longest-lived constraint. A chip can be purchased within months, but a large grid connection may require years of planning, transmission upgrades and regulatory approval. Transformers and turbines have their own lead times. Data centers compete with households and factories for power, creating political pressure over rates and reliability.

The economic value of a megawatt depends on location, uptime, cooling, network connectivity and the workload it supports. This is why announcements measured only in gigawatts can mislead. Contracted power is not the same as an operating data center, and an operating data center is not the same as profitable utilization.

Cooling, water and construction

Higher-density computing produces more heat. Cooling systems, water availability and local environmental rules can constrain deployment. Construction labor and permitting add further delays. These frictions support Visser’s view that physical infrastructure cannot scale at software speed.

They also create political and social risk. Communities may resist projects that consume large amounts of electricity or water while providing relatively few permanent jobs. Governments may impose conditions, raise taxes or prioritize grid access for other industries. A project can be technologically feasible and financially unattractive after local costs are included.

Model efficiency and workload economics

Every physical bottleneck interacts with software efficiency. Better algorithms can reduce the compute needed for a given task. Smaller models can handle routine work while frontier models are reserved for difficult problems. Caching, retrieval and specialized inference can lower cost.

The demand response determines the market outcome. If efficiency reduces cost by 50% and usage rises fivefold, total compute demand increases. If enterprises discover that many agentic applications do not generate enough value, efficiency may reduce total spending. Investors need evidence about workload economics: how much customers save, what revenue they create and whether the agent performs reliably enough to replace human labor.

This is why a blanket “compute shortage” thesis is incomplete. The most attractive investments may be in constraints whose supply cannot respond quickly and whose customers can earn enough to pay higher prices. A shortage that exists only because buyers are subsidizing uneconomic experimentation can disappear when budgets tighten.

Why leverage may fall without disappearing

Visser predicted that prime brokers would be less willing to extend leverage after repeated speed crashes. History supports a temporary tightening after losses. Lenders review collateral, raise haircuts, reduce concentration limits and demand more transparency. Managers who survived may voluntarily lower gross exposure.

But leverage is unlikely to disappear because it serves legitimate functions. Market-neutral funds borrow to offset long and short positions. Dealers finance inventories. Arbitrage strategies need balance-sheet capacity to align prices across markets. Corporations borrow to invest. The question is not whether leverage exists; it is whether the system prices and monitors it correctly.

Three forces could reduce leverage in AI-linked equities.

Higher volatility. Risk models translate volatility into position limits. If daily price swings remain large, the same dollar position consumes more risk budget.

Greater concentration charges. Prime brokers may require more collateral for portfolios whose holdings are economically correlated, even if the securities span several industries and countries.

Less confidence in exit liquidity. The Situational Awareness sale showed that a multibillion-dollar book may require a negotiated buyer. Lenders may discount the amount that can be realized quickly.

Other forces could rebuild leverage. Strong returns attract capital; volatility eventually declines; new derivative structures provide exposure; and competition among prime brokers can weaken discipline. The cycle often begins again under a new name.

The regulatory question is whether authorities need more visibility into concentrated exposures across counterparties. Archegos demonstrated the danger when each bank saw only part of the risk. Situational Awareness appears to have been resolved without large bank losses, but the scale of the portfolio raises similar questions about aggregate exposure and liquidity.

Improved reporting could help, but transparency can have costs. Publishing positions in real time may make funds vulnerable to front-running and reduce their willingness to provide liquidity. Regulators need enough information to see the system without forcing every strategy into public view.

Risk management lessons for funds, companies and individual investors

The July episode is not only a hedge-fund story. The same principles apply to corporate treasuries and individual portfolios.

Liquidity should be modeled under stress

Average trading volume overstates the amount that can be sold during a common exit. Stress tests should assume wider spreads, lower depth and rising correlation. Private assets should not be counted as immediately available cash merely because they carry a high valuation.

Financing maturity should match asset duration

A three-year thesis financed with daily margin is a structural mismatch. The investor may be right eventually and insolvent first. Longer-term capital reduces forced-sale risk but usually costs more.

Position limits should reflect common drivers

Owning ten AI-infrastructure stocks is not the same as owning ten independent risks. Exposure should be grouped by economic factor: hyperscaler spending, memory pricing, power availability, interest rates and customer concentration.

Profit taking is not proof of weak conviction

Reducing a position after a large gain can protect the ability to hold the remaining exposure. A manager who refuses to rebalance because the thesis is strong may allow the market to determine the portfolio’s risk.

Cash has option value

Cash and short-term government securities may lose purchasing power over long periods, but they allow an investor to meet obligations and buy during stress. A portfolio invested entirely in high-volatility assets has no internal source of liquidity.

Corporate strategy should not depend on permanent capital-market access

Companies financing AI infrastructure or Bitcoin holdings through continuous issuance should test what happens if their equity trades below asset value, credit spreads widen or markets close. A model that works only when new capital is cheap is not resilient.

Narrative diversification is not financial diversification

AI, Bitcoin and tokenization can appear to be separate themes while sharing sensitivity to liquidity, real yields and risk appetite. Their prices may fall together when financing conditions tighten. Diversification should be evaluated empirically and under stress, not inferred from different stories.

The larger lesson is that risk management is not a prediction that the thesis will fail. It is an acknowledgment that price, timing and financing can move independently from the thesis. The investor who survives retains the right to be correct later.

A practical framework for evaluating AI-linked investments

This article does not provide personalized investment advice, but the July event suggests a useful analytical framework.

1. Identify the economic exposure, not just the label

“AI stock” can describe a cloud platform, a chip designer, a memory producer, a utility, a software vendor or a speculative data-center developer. Determine what drives revenue, who pays, how prices are set and what capital must be invested.

2. Measure cash conversion

Revenue growth matters less if capital expenditures, working capital and stock compensation consume the cash. Compare operating cash flow, capital expenditures and free cash flow over several periods. For infrastructure businesses, evaluate return on invested capital after realistic depreciation.

3. Stress the funding model

Ask whether the company or fund can survive higher rates, lower collateral values and delayed projects. Debt maturity, covenant terms, margin requirements and customer prepayments can matter more than long-term demand during a crisis.

4. Examine customer concentration

A supplier serving a handful of hyperscalers may have excellent near-term demand and weak bargaining power. Backlog quality depends on counterparties and contract terms.

5. Separate scarcity from cyclicality

A shortage raises prices and profits, encouraging new capacity. Estimate how long supply takes to respond and whether the bottleneck is protected by technology, regulation or geography.

6. Evaluate valuation under multiple outcomes

Do not use one forecast. Model strong adoption, moderate adoption and a spending pause. A security that requires the most optimistic scenario to justify its price offers little margin for error.

7. Respect position size

A high-conviction idea can be held at a size that allows survival. The Situational Awareness episode is a reminder that concentration and leverage turn analytical error into existential risk.

What a 25-year-old should take from the transition

Pompliano asked how a person beginning a career could benefit from tokenization and AI. The most durable answer is not to predict the winning token or single stock. It is to build skills at the intersection of technology, domain knowledge and trust.

AI reduces the value of generic output. It increases the value of problem selection, proprietary data, distribution, judgment and accountability. A young professional should learn to use models, but should also develop expertise that allows them to detect when the model is wrong.

Promising areas include data-center engineering, power systems, semiconductor supply chains, cybersecurity, financial compliance, AI evaluation, product design and industry-specific automation. Tokenization creates demand for lawyers, accountants, custody specialists, smart-contract auditors and market-structure experts—not only developers.

Entrepreneurship may become more accessible because small teams can automate functions. The opportunity is greatest where AI solves a measurable business problem and customers already have budgets. Products that merely demonstrate impressive technology without a distribution strategy may struggle.

Career resilience also requires ownership. That can mean equity in a business, diversified long-term investments or intellectual property. It does not require maximum exposure to volatile assets. The goal is to participate in productivity gains without making personal solvency dependent on one forecast.

Scenarios for the next 6 to 12 months

The scenarios are not mutually exclusive. A scarcity squeeze can occur inside a broader capex disappointment if specific components remain constrained while investors lower valuations. A second liquidation wave can be followed by a healthy reset. The point is to define evidence that would change the conclusion rather than attach certainty to one narrative.

What to watch before the Fed’s next decision

The next phase of the market will be shaped by both earnings and policy. Warsh’s reduced forward guidance makes incoming data more important, but it also means each data release may produce larger changes in rate expectations.

Investors should watch core inflation, wage growth, unemployment claims, oil prices and long-term inflation expectations. The split 9–3 vote shows that a meaningful group favored immediate tightening. If inflation reaccelerates, the hurdle for a rate increase may be lower than under a communication regime that prepares markets weeks in advance.

Long-term yields may be as important as the policy rate. Hyperscalers can finance spending more easily than smaller companies, but a higher cost of capital affects every valuation. It also changes the relative attractiveness of bonds versus growth stocks and crypto.

The Fed’s balance-sheet review deserves attention, but investors should not assume rapid quantitative tightening. Warsh has framed reform as a multi-year institutional project. Shrinking the balance sheet while cutting rates is operationally possible because the two tools affect different parts of financial conditions, but the combination would need careful calibration.

What to watch in AI earnings and credit markets

The next earnings cycle should be judged on utilization and cash returns, not only headline spending. Key indicators include cloud growth, remaining performance obligations, data-center depreciation, free cash flow, customer concentration and management commentary on capacity.

Credit-default-swap spreads and bond yields for hyperscalers also matter. Wider spreads do not mean default is likely; the companies retain some of the strongest balance sheets in the world. They do reveal the marginal cost of financing and the amount of risk the market is being asked to absorb.

Equity issuance by major technology companies and private model developers can create competition for capital. Even a market with trillions of dollars in savings needs time to digest unusually large financings. When supply of securities increases faster than demand, prices adjust.

The private market is another source of uncertainty. Anthropic, OpenAI and other model developers can be marked at high valuations without daily trading. Those marks may be more stable than public stocks, but they are not necessarily less risky. A private holding cannot always be sold when a fund needs liquidity.

What to watch in Bitcoin and crypto

Bitcoin’s response to the AI unwind and Fed uncertainty will help clarify its current market role. If it trades primarily with technology and liquidity-sensitive assets, the “digital gold” narrative remains secondary to risk appetite. If it begins to hold value during equity deleveraging and rising policy uncertainty, the hedge argument strengthens.

ETF flows are an important bridge between crypto and traditional finance. BlackRock’s iShares Bitcoin Trust provides regulated exchange-traded exposure, but the product does not remove Bitcoin’s price risk. It changes custody and access.

Stablecoin regulation, tokenized funds and agentic-payment pilots may matter more for day-to-day adoption than Bitcoin price forecasts. The infrastructure that allows software agents to transact securely could grow even during a crypto bear market.

Investors should also distinguish network use from token price. A blockchain can process more transactions while its native token underperforms if fees fall or value accrues elsewhere. Adoption is not automatically captured by every digital asset.

Frequently asked questions

Did the Situational Awareness hedge fund collapse?

The fund suffered a 67% portfolio decline in July, removed leverage and sold most of its public-equity holdings to Citadel, according to reports. It was still described as being up roughly 80% for 2026 and retained private investments. “Collapse” can imply closure or total loss, neither of which had been confirmed at the research cutoff.

Was Citadel’s purchase a bailout?

It was a private transaction in which Citadel reportedly bought most of the public-stock portfolio. It may have reduced the risk of disorderly selling, but there was no reported government rescue. The price, exact terms and Citadel’s expected profit were not fully public.

Did Citadel or Ken Griffin cause the fund’s losses?

There is no verified evidence that Citadel intentionally caused the losses. The portfolio was already exposed to a broad AI-share reversal and leverage pressure. Speculation about deliberate policy signaling should not be presented as fact.

Does the unwind mean the AI boom is over?

No definitive conclusion can be drawn from one liquidation. Current hyperscaler earnings show strong cloud and AI demand. The event does show that valuations, leverage and crowding had become vulnerable.

What is compute scarcity?

Compute scarcity refers to demand for processing capacity exceeding the available supply of chips, memory, networking, data centers and electricity at an acceptable price. It can create pricing power, but it can also trigger investment that eventually increases supply.

Why can AI companies fall when demand is strong?

Stock prices reflect expectations. A company can grow rapidly and still decline if investors expected faster growth, if capital spending rises, if financing becomes more expensive or if the valuation already assumed exceptional success.

What happened in South Korea’s stock market?

After a large AI-driven rally, the KOSPI suffered record-setting declines as semiconductor stocks fell and leveraged retail accounts faced margin calls. The market lost trillions of dollars in value from its peak, although the preceding gains were also unusually large.

What did the Federal Reserve decide in July 2026?

The FOMC held the federal-funds target range at 3.5% to 3.75% by a 9–3 vote. Three members preferred a quarter-point increase. Chairman Kevin Warsh provided limited guidance about the next meeting.

Why did long-term Treasury yields rise after no rate change?

Investors may have demanded more compensation for inflation and policy uncertainty. Holding the short-term rate steady does not prevent long-term yields from rising if markets doubt the inflation outlook or expect future tightening.

Is Bitcoin a hedge fund?

No. Bitcoin is a digital asset. The phrase compares its historical compounding with successful fund returns, but Bitcoin has no manager, active strategy or risk controls.

Is Bitcoin a reliable hedge?

Bitcoin has not been a consistent short-term hedge against stock losses, inflation or crises. Its strongest case is as a scarce, long-horizon digital asset, but its volatility and drawdowns remain substantial.

Will AI agents automatically buy Bitcoin?

No. Their allocations will depend on objectives, risk limits, regulation and data. Some systems may allocate to Bitcoin; others may reject it because of volatility or uncertain valuation.

Why might AI agents use stablecoins?

Stablecoins can provide programmable, internet-native settlement without the price volatility of Bitcoin. They may be useful for purchasing compute, data and services, although security and authorization remain major challenges.

What is tokenization?

Tokenization represents ownership or financial claims on a digital ledger. It can speed settlement and enable fractional access, but it does not remove legal, valuation or liquidity risks.

Can tokenization make the Fed unnecessary?

No evidence supports that conclusion. Tokenization may improve market plumbing, but the Fed’s roles in monetary policy, bank reserves, payments and financial stability remain distinct.

What is the main lesson for investors?

A powerful long-term theme does not justify unlimited leverage or concentration. Survival depends on position size, liquidity, financing and the ability to remain invested through adverse moves.

Final assessment

The July 2026 AI stock unwind should be remembered as a collision between a credible technological transformation and an unstable investment structure. The demand for AI services did not vanish. Cloud revenue, contracted commitments and semiconductor exports remained strong. What changed was the market’s willingness to finance the theme at any price and with any amount of leverage.

Leopold Aschenbrenner’s fund embodied the contradiction. Its thesis about the importance of compute may remain broadly correct. Its public-equity portfolio was still forced through a devastating drawdown. The episode confirms that insight and risk management are separate skills.

Jordi Visser’s argument that markets will experience more parabolas and speed crashes deserves serious attention. Faster information, common models and automated risk controls can compress market cycles. Reduced Federal Reserve guidance may add volatility. Investors and lenders may respond by using less leverage, but that adjustment will not be smooth.

His compute-scarcity case is supported by current evidence, especially the growth reported by Amazon, Microsoft and Alphabet. The strongest rebuttal is not that demand is fictional. It is that suppliers may spend so much to meet demand that shareholder returns disappoint. The next stage of the AI trade will be decided by cash conversion and returns on capital, not by the number of times executives mention AI.

The Bitcoin argument is more conditional. Bitcoin’s scarcity and institutional access have strengthened its status as a distinct asset. Its volatility prevents it from being treated as a reliable hedge or cash equivalent. It may benefit from a world of monetary uncertainty and machine-native finance, but that outcome depends on adoption, regulation and portfolio behavior that cannot be assumed.

Tokenization and agentic payments are moving from theory to infrastructure. They can reduce settlement friction and allow software to participate in commerce. They can also accelerate errors, leverage and financial contagion. The technology’s success will depend on legal claims, authorization and accountability as much as code.

The most defensible conclusion is neither maximalist nor dismissive. AI is changing business, markets and capital allocation. Compute remains scarce in important areas. Bitcoin remains a volatile scarcity asset rather than a proven universal hedge. Tokenization is a significant financial-infrastructure development rather than a magic source of wealth. And leverage remains the mechanism that can turn a good idea into a forced sale.

Sources

• Source video: Anthony Pompliano, The Pomp Podcast, “Bitcoin Is The Best Hedge Fund That’s Ever Existed,” featuring Jordi Visser, published August 1, 2026.
• Reuters: Situational Awareness portfolio sinks 67% in July
• Reuters: Citadel buys most of Situational Awareness’s stock holdings
• Reuters: Citadel and the Situational Awareness transaction
• MarketWatch: Equal-weighted S&P 500 versus the Nasdaq-100 in July
• Goldman Sachs: Hedge-fund positioning and the AI reset
• Reuters: South Korea’s KOSPI rout
• Reuters: South Korean margin-loan risks
• Reuters: South Korea’s July exports and AI demand
• Reuters: Japan’s Nikkei and AI-spending concerns
• Federal Reserve: July 29, 2026 FOMC statement
• Federal Reserve: Review task forces and policy framework
• Amazon: Second-quarter 2026 results
• Microsoft: Fiscal fourth-quarter 2026 results
• Alphabet: Second-quarter 2026 results
• Meta: Second-quarter 2026 results
• Leopold Aschenbrenner: Situational Awareness: The Decade Ahead
• NBER: What Happened to the Quants in August 2007?
• SEC filing: Credit Suisse report on Archegos
• IMF: Tokenization and financial architecture
• IMF Note: Tokenized Finance
• BlackRock: Larry Fink’s annual chairman’s letter
• iShares: Bitcoin volatility and historical drawdowns
• Empirical Economics: Bitcoin volatility and its role as a medium of exchange
• Strategy: Bitcoin transaction ledger
• SEC filing: Strategy first-quarter 2026 results
• BlackRock: iShares Bitcoin Trust

Affiliate disclosure: Businessfinance.news may earn a commission from qualifying purchases or registrations made through affiliate links. This does not influence our editorial analysis, and no affiliate relationship changes the facts, risks or conclusions presented in this article.