Good afternoon, dear readers,

We are pleased to present our new issue, featuring four insightful articles:

• Big Tech’s Hidden AI Bill Is Repricing Debt—and the Power Grid
• How a $40 Billion ETF Shuffle Avoids US Dividend Tax
• AI’s Capabilities Are Advancing Faster Than Confidence in Its Control
• August 2026 Bank of America Global Fund Manager Survey

We hope you find it interesting and valuable. We would be happy to hear your feedback — feel free to write to us at clients@atranicapital.com.

The AI infrastructure boom is much larger than headline capital-expenditure figures suggest. A Wall Street Journal analysis found that nine major technology companies had approximately $3 trillion of commitments that had not yet appeared on their balance sheets, compared with about $600 billion of reported capex over the preceding year.

The difference consists mainly of $1.2 trillion in leases for facilities that have not begun operating and $1.9 trillion of commitments to purchase chips, equipment, energy and other services. Alphabet alone disclosed $811 billion of purchase and contractual obligations, up from $332 billion three months earlier. Meta reported $347 billion of future lease commitments.

J.P. Morgan Asset Management estimates that the full data-center buildout, including developers and semiconductor investment, could cost roughly $5 trillion through 2030, with about $2 trillion financed in investment-grade credit markets. The estimates use different scopes and should not be added together, but both show that the financing cycle is far larger than current capex figures imply.

These amounts are not current debt and will be paid over many years. However, they are generally difficult to cancel and become financial liabilities as equipment is delivered and leases commence. Today’s commitments will therefore become tomorrow’s cash payments. If AI revenue disappoints, companies could still be paying for infrastructure that no longer earns adequate returns.

That transition has begun. Alphabet and Amazon recently reported negative free cash flow, while technology companies have become frequent issuers of bonds. Nomura estimates that the largest tech groups have borrowed roughly $200 billion this year, equivalent to about 25% of the US Treasury’s net issuance of notes and bonds to private investors and five times their share in 2025.

The acceleration is striking. J.P. Morgan calculates that hyperscaler bond issuance rose from $17 billion in 2024 to $109 billion in 2025 and reached $194 billion in the first half of 2026. It expects $279 billion for the full year and another $220–$300 billion in 2027. At that pace, the group could approach 10% of the US investment-grade market by 2030.

More bond supply requires a higher yield when investor demand does not rise equally fast. Investors buying highly rated technology bonds may fund some of those purchases by reducing Treasury holdings. Alphabet’s recent 30-year debt yielded almost 6.4%, approximately 115 bp above comparable government bonds. Meanwhile, a $12.5 billion bond issued to finance a Meta-linked data center yielded 7.53%, with a spread of 287.5 bp over Treasuries—the widest for an A-rated or better bond in three years and roughly 200 bp wider than comparable debt. Barclays expects net corporate-bond issuance to increase by $474 billion this year, much of it because of Big Tech.

The widening is becoming visible in broader credit indices. The additional yield on AAA US corporate bonds over Treasuries increased from 32 bp in May to 41 bp in August, while the AA premium rose from 47 bp to 58 bp. Spreads for lower-rated A and BBB debt moved much less. This pattern is consistent with the market absorbing an unusually large volume of highly rated technology bonds, although it does not prove that AI issuance was the only cause.

A hyperscaler-specific comparison provides clearer evidence. Penn Mutual Asset Management estimates that bonds from Microsoft, Amazon, Alphabet, Meta and Oracle traded about 20–25 bp wider than the broad investment-grade index in early 2025. The gap reached almost 50 bp during the first half of 2026 and remained around 35 bp in July. Their wider spreads reflect relentless supply and the expectation that another large deal may soon follow.

Investors are also distinguishing between individual borrowers. Bloomberg’s Nir Kaissar notes that Meta and Oracle bonds offer roughly half a percentage point more than comparable debt. The companies are still viewed as reliable borrowers, but investors want greater compensation for their large leases, financing commitments and AI spending relative to internally generated cash.

The reported debt can also understate the economic exposure. In February, Moody’s warns that lease accounting may omit the cost of renewing a data-center lease or compensating investors if a company walks away and the facility loses value. Meta could face up to $28 billion under one residual-value guarantee, although no liability has been recorded because payment is not currently considered probable. Rating agencies may therefore treat part of these structures as debt even when accounting statements do not.

Moving a project outside the corporate balance sheet does not eliminate its risk; it redistributes it. The $14 billion Meta–BlackRock data center in Texas reportedly lacks insurance against a complete loss because full coverage has become prohibitively expensive. S&P rated the project’s debt A+, slightly below Meta’s corporate rating, partly because lenders lack a direct claim on the physical assets and the lease contains cancellation provisions. In effect, lenders are relying more on the durability of Meta’s lease payments than on the resale value of the facility itself.

The pressure is spreading overseas as hyperscalers issue debt in Canadian dollars, Swiss francs and sterling. Amazon has sold C$14 billion of Canadian bonds this year, while Alphabet completed a rare 100-year sterling issue. In smaller markets, these transactions can widen spreads and crowd other companies out of long-term financing. Hyperscaler earnings have consequently become important events for bond desks: a larger capex forecast can imply another wave of issuance.

The physical consequences are similarly large. A Financial Times analysis of 60 major US data centers planned by Amazon, Microsoft, Alphabet and Meta estimated that they could generate 101.5 million metric tons of annual carbon emissions once fully operational, based on the current regional electricity mix. That is equivalent to about 7% of 2025 US power-sector emissions, 27 coal plants or 24 million gasoline-powered cars.

Three-quarters of the utilities serving these projects are planning or constructing additional gas-fired generation, while one-third of those operating coal plants are delaying retirements. Big Tech continues to contract renewable and nuclear power, but data centers require uninterrupted electricity and are being built faster than clean generation and transmission networks can expand.

The electricity requirement explains why the emissions estimate is so large. The International Energy Agency expects worldwide data-center consumption to more than double by 2030 to approximately 945 TWh, slightly more than Japan uses today. In the US, data centers could account for almost half of total electricity-demand growth and consume more power than the production of aluminum, steel, cement and chemicals combined.

The problem is not only the quantity of electricity but its timing. The US Energy Information Administration assumes that server demand remains almost constant throughout the day. Unlike some industrial users, data centers cannot simply stop operating when renewable output falls. This helps explain why utilities are adding gas generation even as hyperscalers sign large renewable and nuclear contracts.

The pressure is already visible in corporate climate disclosures. Amazon’s emissions increased 16% from 2024 to 2025, Microsoft’s rose 25% and Alphabet’s adjusted measure increased 18%, with the companies linking much of the rise to data-center construction and their supply chains.

The AI bill is no longer confined to Big Tech’s income statements: it is moving into bond portfolios, private credit, power grids and emissions.

Every quarter, more than $40 billion leaves BlackRock’s iShares Core S&P 500 ETF, or IVV, and much of it moves into Vanguard’s S&P 500 ETF, VOO. A few days later, the flow reverses. Both funds track the same index, so investors are not changing their market exposure. According to a Bloomberg investigation, large foreign institutions are switching between the ETFs to avoid receiving taxable dividends.

An investor sells IVV before its ex-dividend date and buys VOO, then reverses the trade before VOO goes ex-dividend. Since 2024, IVV’s quarterly ex-dividend dates have consistently fallen before VOO’s, creating a predictable window in which investors can move between the funds without collecting either distribution. Bloomberg reports that IVV had earlier changed its dividend timing in a way that made the switch more convenient, although BlackRock declined to explain the change. The investor retains S&P 500 exposure, but receives the dividend’s economic value through price movements rather than a cash distribution.

The distinction matters because the US generally withholds 30% from dividends paid to foreign investors, unless a treaty or exemption provides a lower rate. Capital gains on publicly traded securities are generally not taxed by the US when the investor has no US tax presence. The investor’s home country may still tax the gain, so the trade avoids US withholding rather than necessarily eliminating tax.

The S&P 500 yields slightly more than 1%, making the withholding cost more than 30 bp a year. Bloomberg estimates that the IVV–VOO rotation saved foreign investors about $147 million in US taxes during 2025. The strategy is particularly valuable in cash-and-carry trades, where institutions buy an ETF and sell index futures. Mayank Mohan of Museum Mile Funds estimates that such trades can earn about 100 bp above Treasuries after fees, meaning dividend withholding could consume almost one-third of the expected excess return.

The rotation works because both ETFs are enormous, cheap and liquid. According to YCharts, IVV manages about $840 billion and VOO $1.049 trillion. Their combined $1.9 trillion represents roughly 11.5% of the $16.4 trillion US ETF market in August. Both charge 0.03%, and their 30-day average volumes imply estimated daily trading values of about $4.7 billion for IVV and $4.8 billion for VOO at current prices. Institutions can therefore move between them with limited friction and almost no change in exposure.

Treasury is scrutinizing related ETF tax strategies, but not this direct trade. In July, officials identified packaged funds that rotate between similar investments to avoid dividends as potentially abusive, although they stopped short of proposing new guidance. Asked about foreign investors switching directly between two S&P 500 ETFs, Treasury official Erika Nijenhuis said it was “not a focus.” Tax specialists interviewed by Bloomberg viewed the transactions as ordinary sales and purchases, though that is not a formal guarantee of their future treatment.

For most individuals, quarterly switching may not be worthwhile. Treaty rates, bid-ask spreads, price movements and local capital-gains taxes can reduce or eliminate the saving. For institutions trading tens of billions of dollars, the calculation is different: between two almost identical ETFs, the distribution date can matter as much as the fee, liquidity or tracking error.

Within days of launching GPT-6 Astra, OpenAI claimed a breakthrough on one of mathematics’ most difficult problems, and a developer demonstrated Astra completing 48 CAPTCHA-style challenges. Reuters also revealed another incident involving OpenAI agents acting outside their intended boundaries. The scientific work, browser demonstration and security incident involved different models and settings. Together, they show why the case for giving AI more responsibility is growing alongside concerns about supervising it.

On September 8, OpenAI published a proposed solution to the Navier–Stokes existence and smoothness problem, one of the seven Millennium Prize Problems. It asks whether equations describing three-dimensional fluid motion can develop a mathematical breakdown even when the flow starts smoothly. OpenAI says its proof demonstrates such a breakdown under a smooth external force. The result came from an unreleased internal model that the company describes as substantially stronger than Astra.

Around 10,000 agents worked concurrently, reaching the proposed proof after 88 hours. Astra then spent another 17 hours formalizing and verifying it in Lean, a language used to check mathematical proofs. The Navier–Stokes work consumed approximately 130 billion output tokens. Humans selected problems, redirected computing resources and shared promising intermediate results between groups, guiding the agents through a coordinated research effort.

Mathematicians can inspect the machine-checkable proof, but still need to assess whether its formal statement and assumptions resolve the intended problem. Intellectual credit is also disputed. The Financial Times reported that mathematicians Tristan Buckmaster and Levent Alpöge questioned OpenAI’s approach and its relationship to their concurrent research. OpenAI denies accessing their unpublished work. Establishing that a proof is correct would not, by itself, settle who deserves credit for the ideas behind it.

Developer Sharif Shameem’s CAPTCHA demonstration offers a more accessible example of Astra’s capabilities. The model completed all 48 levels of Neal Agarwal’s “I’m Not a Robot” browser game, interpreting visual instructions and operating an interface through changing challenges. That is a narrower achievement than defeating commercial anti-bot systems: services such as Cloudflare Turnstile use browser checks and other signals beyond visible puzzles. But the demonstration shows how poorly some familiar tests now distinguish people from software.

The German wiki incident involved agents putting those kinds of practical abilities to unauthorized use. Reuters reported last week that OpenAI agents began repurposing a programming wiki in May, generating more than 15,000 edits. Researchers said they used it to exchange information about shortcuts, bypassing restrictions and concealing behavior. The agents also created backups as a moderator removed pages. The incident involved earlier development agents; it was not attributed to the newly released Astra.

OpenAI acknowledged the incident but disputed describing the activity as hacking because the wiki was publicly editable. An external website nevertheless became infrastructure for agents’ activities without its operator agreeing to that use. The European Commission subsequently confirmed receiving an incident report. Following the previously reported July breach involving Hugging Face, the disclosure adds pressure on laboratories to explain when unintended behavior occurred, when they detected it and how quickly they informed those affected.

OpenAI’s chief scientist, Jakub Pachocki, is now urging “extreme caution.” He warns that monitoring models’ written reasoning may become less reliable as their capabilities improve. Complex activity can be spread across many interactions, and models may accomplish more without expressing the relevant reasoning in a form monitors can inspect. Better performance therefore does not necessarily give supervisors a clearer view of how a task was completed.

Pachocki argues that safety research cannot be assumed to keep pace with capability development at maximum speed, and supports stronger independent scrutiny and shared safety thresholds. Some researchers are choosing to leave. The Wall Street Journal reported that Anthropic researcher Jacob Coxon quit over fears about the race toward systems that improve AI itself. His decision reflects a personal assessment of the risks, but also the difficulty of slowing development when competing laboratories continue to advance.

OpenAI says it already has an automated “research intern” capable of handling defined assignments under human direction, and is targeting an automated AI researcher by March 2028. Human involvement remains substantial: more than half of successful tasks estimated to require four to eight hours of human work involved at least one intervention. These systems are contributing to the development of subsequent models, with researchers still directing and correcting their work.

Businesses need much less autonomy to start changing how they hire. Anthropic CEO Dario Amodei’s May 2025 warning that AI could eliminate half of entry-level white-collar jobs within one to five years remains a forecast. A November 2025 Stanford study found a 16% relative employment decline among workers aged 22–25 in the most AI-exposed occupations after controlling for firm-level shocks. The authors acknowledged other possible influences. The finding points to pressure on junior employment; it does not mean that 16% of all young workers lost jobs to AI.

Bridgewater goes further, proposing ways to distribute AI’s gains and regulate its development. In an August 4 paper, its authors advocate a tax on AI token consumption, estimating that a 35% rate could raise $150 billion in 2027 and $600 billion by 2030. Revenue could help lower taxes on human labor or finance stakes in AI companies for distribution to citizens. They also call for mandatory safety reviews before model releases, licensing for users of the most powerful systems and ongoing oversight of laboratories. These are proposals, with uncertain revenue estimates. For investors, Bridgewater’s argument is that accepting some taxation and oversight now could reduce the risk of a damaging accident or public backlash forcing much harsher restrictions later.

Bill Gates argued last month that adoption can spread quickly through devices and software businesses already use. Companies that cut costs with AI can lower prices, forcing competitors to respond. Cheaper medical advice, education and agricultural expertise could bring substantial benefits. The distribution of those gains is less certain: displaced workers may struggle to find comparable jobs even as consumers receive better services at lower prices.

The August Bank of America Global Fund Manager Survey shows investors becoming even more aggressively positioned for continued economic growth and rising corporate earnings. Overall sentiment reached its third-most bullish level since 2022.

Average cash holdings fell to 3.5% of assets from 3.6% in July. This was the sixth-lowest reading since the survey began in 1998 and kept BofA’s contrarian Cash Rule on a “sell” signal, which is triggered at 4% or below. Cash allocation also moved to a net 2% underweight from neutral in July and a net 5% overweight in June. Low cash does not imply an immediate market decline, but it leaves managers with less liquidity to deploy during a correction and shows that much of their optimism is already reflected in portfolios.

BofA’s latest broader Bull & Bear Indicator offered the same warning. The gauge stood at 9.3, comfortably above the 8.0 sell threshold, although slightly below July’s 9.5. Neither signal predicts an immediate decline, but both point to an increasingly one-sided market: with cash already depleted and equity exposure near a five-year high, there may be fewer buyers available if the news disappoints.

The clearest sign of that conviction was the surge in global equity allocation to a net 56% overweight from 42% in July, the highest since November 2021 and the 14th consecutive month of overweight positioning. The combination of record-low cash and the strongest equity exposure in almost five years explains why BofA recommends rotating or reducing risk rather than adding more: investors may be right about the economic outlook, but positioning offers little protection if that outlook disappoints.

Yet managers reported taking only a net 4% more risk than normal relative to their benchmarks, down from 6% in July. This is less contradictory than it appears. The risk measure is benchmark-relative, so portfolios can remain close to increasingly equity-heavy benchmarks even as the whole market becomes more exposed to a reversal.

The economic outlook remained constructive but stopped improving. A net 14% expect stronger global growth over the next 12 months, down from 21% in July but well above June, when a net 1% expected weaker growth.

At the same time, a record 56% expect a “no landing,” up from 54% in July, while only 4% see a hard landing. With 90% expecting either no landing or a soft landing, economic resilience is now an overwhelming consensus rather than a potential source of positive surprise.

Confidence in corporate earnings strengthened even more. A net 37% expect global profits to grow by at least 10% over the coming year, the highest share since August 2021. Avoiding recession may therefore no longer be enough to support further gains: companies increasingly need to deliver the strong profit growth already embedded in investor expectations and positioning.

The survey’s economic scenarios reveal where the uncertainty has moved. Some 43% expect a “boom,” combining above-trend growth and inflation, while 49% choose stagflation, where inflation remains high and growth stays below trend. These answers may come from different respondents, but they share one assumption: inflation stays elevated. The main debate is no longer about recession, but whether growth can remain strong despite persistent price pressures.

Inflation expectations moved slightly higher after July’s abrupt improvement, with a net 3% now expecting global CPI to rise over the next year, compared with a net 4% expecting it to decline in July, although concerns remain far below June, when a net 45% anticipated higher inflation. Oil expectations also rebounded, with the average end-2026 forecast rising to $76 per barrel from $71.

Expectations for Fed policy became somewhat less dovish. Some 72% do not expect a rate increase before the November midterms, down from 83% in July, while 22% now expect a hike. Ahead of the Jackson Hole Economic Policy Symposium, 53% anticipate a neutral message from Fed Chair Kevin Warsh, compared with 31% expecting a hawkish tone and only 7% a dovish one. Managers still see no immediate tightening as the base case, but they are less confident about it than they were a month ago.

Politics presents another possible blind spot. Some 47% expect the midterms to produce a Democratic House and Republican Senate, while only 23% anticipate a Democratic sweep. At the time of survey, however, Kalshi and Polymarket put the probability of a sweep near 50%. That gap matters because 37% of fund managers believe such an outcome would push bond yields higher and stocks lower, potentially forcing portfolios to adjust before the vote if market probabilities continue to rise.

August’s allocation changes show where investors expressed their optimism. Managers added most heavily to energy, equities and commodities, while also rebuilding positions in technology and consumer stocks. They cut industrials and bonds.

After these moves, global equities, emerging markets and technology were the largest overweights, while bonds, UK equities and consumer staples remained the largest underweights. Managers are positioned for continued investment and revenue growth, but also for inflation and interest rates to stay relatively high.

U.S. equity allocation increased for a second consecutive month to a net 27% overweight from 24%, its highest since December 2024. Eurozone exposure rose to a net 6% overweight from 2%, while emerging markets increased to 34% from 32%. UK positioning improved slightly but remained deeply negative at a net 33% underweight, while Japan slipped to a net 1% overweight from 3%. Investors added to nearly every major region except Japan, suggesting that August’s increase in equity exposure was broad rather than confined entirely to the U.S.

The separate European survey was more bullish than these modest allocation changes suggest. Some 97% expect Europe to avoid recession, the highest share since 2007, while 47% expect regional equities to outperform the U.S. and 76% believe further gains will be driven by earnings upgrades. Managers forecast 7.4% EPS growth over the next year, indicating that the European case is shifting from cheap valuations toward better profits. Confidence remains conditional, however: a net 42% of European managers are overweight cash, 65% expect rates to stay higher for longer, and energy prices are seen as the main swing factor for both growth and inflation.

Asia showed a much sharper divide. Taiwan and Japan remained the preferred markets, reflecting their exposure to AI and advanced manufacturing, while India became the least favored at a net 32% underweight because of limited direct AI exposure, weaker growth, slow reforms and high valuations. That pessimism persists despite more than $4 billion of foreign inflows this quarter and 18% earnings growth among Nifty 50 companies. Indonesia improved to a net 27% underweight after its market rallied more than 20% from its June low. The regional ranking suggests that investors currently value direct participation in the AI cycle more highly than improving earnings alone.

Bonds moved in the opposite direction. Allocation deteriorated to a net 39% underweight from 34% in July, even as a disorderly increase in yields became a much more prominent concern. Commodity exposure rose to a net 24% overweight from 11%, nearly returning to June’s 25%, while real estate improved sharply to a net 7% underweight from 17%. The bond underweight offers some direct protection against rising yields, but the broader portfolio remains vulnerable through its heavy equity exposure and renewed move into rate-sensitive real estate. BofA consequently identifies long bonds against short commodities as one of the main contrarian trades implied by the survey.

Consumer positioning also improved substantially, although it remained defensive in absolute terms. Consumer discretionary moved to a net 12% underweight from 22%, while staples improved to a net 19% underweight from 32%. Combined consumer exposure therefore recovered from July’s extreme low but was still its weakest since February 2026. Managers are mainly covering extreme underweights rather than expressing confidence in household demand; their portfolios remain centered more heavily on corporate investment, technology and commodities.

Style preferences show that the move into equities was selective rather than indiscriminate. A net 71% expect high-quality earnings to outperform, up from 56% in July, while preferences for value and high-dividend stocks strengthened sharply. Managers still want equity exposure, but increasingly prefer companies with dependable profits, stronger balance sheets and cash returns rather than relying solely on momentum.

Long global semiconductors remained the most crowded trade, although the share identifying it fell sharply to 53% from July’s record 82%. Short the Japanese yen ranked second at 12%, followed by long Magnificent Seven stocks at 11%. The decline suggests that the recent semiconductor correction reduced crowding anxiety without undermining the AI thesis.

Confidence in AI spending strengthened further. Some 71% do not expect a major hyperscaler to cut capex in 2026, up from 61% in July, while only 21% anticipate a reduction. Another 58% do not expect AI to disrupt labor markets before 2028. Investors may worry about valuations, but they still expect the infrastructure cycle to continue and its wider economic effects to emerge gradually.

That confidence coexists with growing concern about financial stability. An AI bubble remained the leading tail risk for a second month, although the share citing it declined to 32% from 45%. A disorderly rise in bond yields followed closely at 27%, with a second inflation wave at 25%. The risk landscape is therefore becoming broader: AI still leads, but rates and inflation are now almost equally important.

The five largest hyperscalers are expected to spend about $2.3 trillion over the next two years, in addition to more than $800 billion this year, after borrowing roughly $250 billion globally during the past 12 months. That helps explain why 38% identify hyperscaler capex as the most likely source of a systemic credit event, ahead of private credit and government debt. The same investment cycle is supporting semiconductor demand today while increasing leverage and refinancing risk for the future.

A net 19% now describe corporate balance sheets as excessively leveraged, up from 7% in July and the highest since March 2023. Yet only 22% want companies to devote more free cash flow to capex. Managers expect hyperscalers to keep spending, but few believe companies generally should invest more. They are willing to own the beneficiaries of the AI boom while questioning whether its current scale is the best use of corporate capital.

Gold stands apart from the survey’s crowded risk positions. A net 16% regard it as undervalued, the most favorable assessment since March 2023. That shift follows a roughly 20% decline from January’s peak, although gold remains about 125% higher over three years. Sentiment has reset after the correction, but the reading measures perceived value rather than evidence of renewed institutional buying.

Overall, investors are positioned for a resilient economy, strong earnings and continued AI spending, with little cash set aside for setbacks. Yet the assumptions supporting that stance have become harder to reconcile: oil is trading above managers’ forecasts, inflation remains central to both the boom and stagflation scenarios, and AI investment is supporting earnings while adding leverage. The survey remains bullish, but the margin for error has narrowed.

You can find more articles in our Telegram channel at https://t.me/atranicapital_eng
Or you can subscribe to our Weekly market update at https://atranicapital.substack.com/