# Picks thoughts

> Longer, dated essays about forces that cut across many companies. Claims cite numbered sources as [1], [2] and so on, listed at the end of each essay. Newest first. Page: https://picks.kamadorueda.com/thoughts/

Picks is a personal record of stock picks. It gives information only, not financial, investment, tax or legal advice. Prices, fair values and ratings can be delayed or wrong; check the original source. Full disclaimer: /disclaimer/.

## What a 5% Treasury yield does to the picks

Date: 2026-10-11. Page: https://picks.kamadorueda.com/thoughts/interest-rates-and-the-picks/

The US 10-year Treasury yield rose from 4.44% at the end of June 2026 to 5.22% on October 8, and the Fed raised its rate in September. Most of the rise comes from the term premium and a higher expected Fed path, and inflation expectations stayed near 2.3%. Energy from the Iran war causes most of the gap between headline and core inflation, so a peace deal can reverse the Fed part of the rise. The term premium has other causes, so I expect yields to stay high into 2027. The picks that lose most have high debt, debt that resets soon, or demand that depends on loans. Few picks gain.

### Rates went up for two reasons

On September 16, 2026, the Federal Reserve raised its target range by 0.25 percentage points, to 3.75% to 4.00%. The vote was 12 to 0 [1]. The Fed cut three times in late 2025, to 3.50% to 3.75%, and this was its first rise after those cuts [2]. The statement says that "inflation remains elevated" [1]. The consumer price index rose about 3.4% in the 12 months to August 2026 (my arithmetic on the index) [3]. Core PCE inflation, which leaves out food and energy, was 3.0% in August [34]. The median Fed official now expects a rate of 4.1% at the end of 2026 and at the end of 2027. In June, the medians were 3.8% and 3.6% [5]. That median implies one more rise this year, and the minutes say that "most participants" expected another rise "by year end" [4].

Long yields rose more than the Fed rate. The 10-year yield closed at 5.31% on October 5, its highest close since May 2002 [6]. The 10-year breakeven inflation rate stayed near 2.3% [7]. The real yield on 10-year inflation-protected Treasuries went from 2.20% on June 30 to 2.87% on October 8 [8]. Investors ask for a higher real return, and their inflation expectations stayed almost the same.

**How US rates moved**

| Measure | October 31, 2025 | June 30, 2026 | October 8, 2026 |
| --- | --- | --- | --- |
| Fed funds target, upper bound | 4.00% [2] | 3.75% [2] | 4.00% [2] |
| 2-year Treasury yield | 3.60% [13] | 4.14% [13] | 4.75% [13] |
| 10-year Treasury yield | 4.11% [6] | 4.44% [6] | 5.22% [6] |
| 10-year real yield (TIPS) | 1.81% [8] | 2.20% [8] | 2.87% [8] |
| 10-year breakeven inflation | 2.30% [7] | 2.24% [7] | 2.35% [7] |
| 10-year term premium (NY Fed model) | 0.56% [9] | 0.51% [9] | 0.92% [9] |
| 30-year mortgage rate | 6.17% (October 30) [14] | 6.43% (July 2) [14] | 7.40% [14] |
| High-yield bond spread | 2.94% [15] | 2.75% [15] | 3.15% [15] |

Note: The mortgage rate is Freddie Mac's weekly survey, so its dates are the nearest survey dates. The spread is the extra yield of US high-yield bonds over Treasuries.

The New York Fed's term structure model divides the rise into two parts. From June 30 to October 8, 2026, its fitted 10-year yield rose 0.74 points, from 4.48% to 5.22%. The term premium, the extra yield that investors ask to hold a long bond, gave 0.41 points. A higher expected path for short rates gave 0.33 points (my arithmetic on the model) [9].

The Fed minutes list the reasons that market commentary gave for the higher term premium. These are geopolitical events, uncertainty about the Treasury's buyback program, and "competition for capital from heavy private debt issuance to finance" AI infrastructure [4]. At the press conference, Chair Kevin Warsh said that "the so-called hyperscalers are out in the market raising funding" [10]. The government also borrows more. CBO puts the fiscal 2026 deficit at $2.0 trillion and net interest at $1.1 trillion, 11% more than in 2025, according to a press report [11]. In August, Treasury kept its auction sizes the same and said that it expects to keep them "for at least the next several quarters" [12].

### Stocks now yield about the same as bonds

The S&P 500 traded at 19.1 times its expected earnings for the next 12 months on October 9, 2026. On June 30 the figure was 20.4 [16]. Its earnings yield is about 5.2% (my arithmetic), almost the same as the 10-year yield [6][16]. The index rose 3.5% after June 30 because earnings estimates rose 10.4% in the same period [16]. The Fed minutes say that this year's stock gains came from earnings, while price-to-earnings multiples fell [4].

Credit markets stay calm. The spread on high-yield bonds rose from 2.75% on June 30 to 3.15% on October 8, and it is still low [15]. Companies can borrow. They pay more because the base rate is higher.

The picks fell while the index rose. From August 28 to October 9, 2026, the 10-year yield went from 4.73% to above 5.2% [6]. In the same weeks, the S&P 500 rose 1.3%, and the median pick fell about 3.7% (my arithmetic on the daily closes of the 48 picks with a US price) [17]. Each company also had its own news, so rates explain only part of each move.

### Debt is the first channel

The picks with high debt and weak interest cover fell most. Campbell's operating profit covered its interest only about 2.6 times in fiscal 2026 [18]. On September 23, Moody's cut Campbell's to Baa3, the lowest investment grade, with a negative outlook [19]. Wynn's operating income covered its interest only about 1.9 times in the first half of 2026. About $1.35 billion of its debt is due in 2027, and in September it replaced 5.25% notes with 6.875% notes [20]. Clorox's net debt is about 4.2 times its fiscal 2026 EBITDA, and $1.09 billion of commercial paper pays floating rates [21]. At Sucro, interest of US$23.3 million used almost all of the US$24.0 million of adjusted EBITDA in the 12 months to June 30, 2026 [22]. PG&E must refinance $2.15 billion of 4.25% convertible notes due December 1, 2027 [23].

Some picks reset their debt soon. Yum must repay or refinance at least $500 million of its June 2027 notes 91 days before they mature. If it does not, its term loan and its revolver also come due [24]. Pool Corporation borrows at variable rates. Two swaps fix the rate on $300 million, and both end on February 26, 2027 [25]. Only 54% of RELX's borrowings have a fixed rate, and their average remaining life is 3.7 years [26].

**Picks where debt makes higher rates hurt**

| Company | Debt fact | Price change, August 28 to October 9, 2026 |
| --- | --- | --- |
| Wynn (WYNN) | interest cover about 1.9 times, $1.35B due in 2027 [20] | -21.1% [17] |
| PG&E (PCG) | $2.15B of convertible notes due December 2027 [23] | -20.9% [17] |
| Clorox (CLX) | net debt about 4.2 times EBITDA, $1.09B floating [21] | -18.3% [17] |
| Campbell's (CPB) | interest cover about 2.6 times, Baa3 with a negative outlook [18][19] | -17.9% [17] |
| Pool (POOL) | variable-rate borrowing, swaps on $300M end February 2027 [25] | -15.2% [17] |
| Yum (YUM) | at least $500M must be refinanced about 91 days before June 2027 [24] | -5.9% [17] |
| RELX (RELX) | 54% fixed, average life 3.7 years [26] | -1.1% [17] |
| Sucro (SUGR) | interest used almost all of EBITDA [22] | no US price |

Note: Price changes are my arithmetic on daily closes. Other news also moved each stock in these weeks.

Long fixed debt protects a company for some years. Brookfield's corporate term debt of $14.4 billion has an average term to maturity of 15 years [27]. For a company like this, higher rates raise the interest cost only when old debt matures.

Companies: [CPB](https://picks.kamadorueda.com/company/CPB.md), [WYNN](https://picks.kamadorueda.com/company/WYNN.md), [CLX](https://picks.kamadorueda.com/company/CLX.md), [SUGR](https://picks.kamadorueda.com/company/SUGR.md), [PCG](https://picks.kamadorueda.com/company/PCG.md), [YUM](https://picks.kamadorueda.com/company/YUM.md), [POOL](https://picks.kamadorueda.com/company/POOL.md), [RELX](https://picks.kamadorueda.com/company/RELX.md), [BN](https://picks.kamadorueda.com/company/BN.md)

### Loans, demand and valuation

Higher rates also reduce demand that depends on loans. The 30-year mortgage rate rose to 7.40% on October 8, from 6.43% on July 2 [14]. ICE's mortgage technology revenue grew 5% in the second quarter of 2026, to $557 million, because more loans started [28]. Of that revenue, origination technology gave $197 million and servicing software gave $226 million [28]. Servicing software earns money on loans that already exist, so it is less sensitive to rates. Pool sells to owners who often pay for a new pool with a loan (my reading).

Steady payers lose value when bond yields rise, because investors compare their cash with the yield on bonds. McDonald's fell 11.2% and PepsiCo 10.7% from August 28 to October 9 [17]. Procter & Gamble rose 5.2% in the same weeks [17], so the effect is different for each company.

Few picks gain. Schwab earns interest on client cash. Its own model says that a rise of 1 point in rates adds 3.6% to its net interest revenue over 12 months, and a fall of 1 point removes 1.6% [29]. Its stock fell 12.2% in these weeks [17]. The gain is small, and higher long yields also increase the unrealized loss on its securities, which was $10.6 billion at June 30 and reduces its capital ratio [29]. Copart earned $182 million of net interest income in fiscal 2026 on about $4.5 billion of cash and securities [30]. The largest gains in these weeks came from AI companies: Marvell rose 27.1%, Meta 24.3% and Synopsys 15.2% [17]. Their own earnings news was stronger than the effect of rates (my reading).

Companies: [ICE](https://picks.kamadorueda.com/company/ICE.md), [POOL](https://picks.kamadorueda.com/company/POOL.md), [MCD](https://picks.kamadorueda.com/company/MCD.md), [PEP](https://picks.kamadorueda.com/company/PEP.md), [PG](https://picks.kamadorueda.com/company/PG.md), [SCHW](https://picks.kamadorueda.com/company/SCHW.md), [CPRT](https://picks.kamadorueda.com/company/CPRT.md), [MRVL](https://picks.kamadorueda.com/company/MRVL.md), [META](https://picks.kamadorueda.com/company/META.md), [SNPS](https://picks.kamadorueda.com/company/SNPS.md)

### How long the high rates can last

Some people say that the high rates will not last, because the war with Iran causes most of the inflation. The war started on February 28, 2026, when the United States and Israel struck Iran. Oil shipments through the Strait of Hormuz fell, and a ceasefire in April did not hold. In October, CFR calls the conflict "worsening" [35].

The data supports part of this view. A barrel of WTI oil cost $64.51 on average in February 2026 and $97.31 in September [36]. Energy prices in the CPI rose about 16% in the 12 months to August. Headline CPI rose about 3.4%, and CPI without food and energy rose about 2.4% (my arithmetic on the indexes) [3][32][33]. Energy causes most of the gap between these two numbers. The Fed agrees that it cannot control oil prices. Chair Warsh said: "We cannot affect any individual price, whether it be oil prices" [10]. In September, Governor Waller said that he does not see energy prices and tariffs "as a significant source of ongoing inflation pressure" [38].

Other data shows inflation outside energy. The core PCE index rose 3.0% in the 12 months to August [34]. The Minneapolis Fed estimates that core PCE inflation would be about 1 percentage point above the 2% target even without tariffs [39]. Governor Cook said that "new and more broadly based price pressures" can come when inflation in the AI sector becomes slower [40]. The minutes say that inflation "could prove to be more persistent than the staff anticipated", and the staff expect inflation to reach 2% only in 2029 [4].

The war affects the two parts of the 10-year rise differently. A higher expected path for short rates gave 0.33 points of the rise after June 30 [9]. This part follows the Fed, and the Fed follows inflation. If a peace deal lowers oil prices, headline inflation falls, and the Fed can stop its rises. I expect most of this part to reverse in that case (my estimate). The term premium gave 0.41 points [9]. Geopolitics is one of its causes, but the minutes also name AI borrowing, and the deficit is large [4][11]. A peace deal does not change these. Investors also expect the war to raise inflation only for a short time. The market's expected inflation rate for the five years that start in 2031 was 2.32% on October 9, near its 2.25% average of October 2025 [37].

If all of the expected-rate part reversed and the term premium stayed, the 10-year yield would be about 4.9% (my arithmetic on the model) [9]. That is still above the 4.1% to 4.4% of late 2025 and early 2026 [6]. For the picks, a peace deal would help the steady payers and the borrowers only a little. Debt that resets in 2027 still costs more than the old debt it replaces, as Wynn's new 6.875% notes show [20].

Companies: [WYNN](https://picks.kamadorueda.com/company/WYNN.md)

### What to watch

The Fed meets on October 27 and 28, 2026 [4]. On October 2, fed funds futures gave a 79.5% chance of no change in October and a 66.2% chance of a rise in December, according to a press report of CME FedWatch [31]. The September consumer price report comes before that meeting. Treasury gives its next refunding statement on November 4 [12]. Talks with Iran and the flow of oil through Hormuz decide the energy part of inflation [35]. A larger auction size would push the term premium up (my reading).

The weak borrowers report soon. Wynn and Clorox report in early November, and Campbell's in early December. Their dates are estimates on this site. Pool's swaps end on February 26, 2027 [25]. Yum's 91-day test on its June 2027 notes falls in about March 2027 (my arithmetic) [24]. Each refinancing shows the new cost of debt for these companies.

The AI buildout is the link to my earlier thought on who pays for it. If the hyperscalers borrow less, one force behind the term premium becomes weaker [4][10].

Companies: [WYNN](https://picks.kamadorueda.com/company/WYNN.md), [CLX](https://picks.kamadorueda.com/company/CLX.md), [CPB](https://picks.kamadorueda.com/company/CPB.md), [POOL](https://picks.kamadorueda.com/company/POOL.md), [YUM](https://picks.kamadorueda.com/company/YUM.md)

### My current view

I expect the 10-year yield to stay near or above 5% into 2027. Inflation is above the Fed's goal, the Fed expects to raise its rate again, the deficit is large, and AI companies borrow large amounts. Each of these supports a higher term premium. A peace deal with Iran would lower oil prices and the Fed part of the rise, but I expect the term premium to stay. A slowdown in AI borrowing would change this view more.

At these rates, the main risk is to the borrowers: Campbell's, Wynn, Clorox, Sucro and PG&E, and Pool and Yum because their debt resets soon. The second risk is to demand that depends on loans, such as new pools and the origination part of ICE. The steady payers, such as McDonald's and PepsiCo, lose some value but keep their cash flow. Companies with net cash, such as Copart, feel the least.

This is a mental model to test. It is information only and gives no advice to buy or sell.

### Sources

1. [Federal Reserve: FOMC statement (2026-09-16)](https://www.federalreserve.gov/newsevents/pressreleases/monetary20260916a.htm)
2. [FRED: federal funds target range, upper limit (DFEDTARU)](https://fred.stlouisfed.org/series/DFEDTARU)
3. [FRED: consumer price index for all urban consumers (CPIAUCSL)](https://fred.stlouisfed.org/series/CPIAUCSL)
4. [Federal Reserve: minutes of the FOMC meeting of September 15-16, 2026 (released 2026-10-07)](https://www.federalreserve.gov/monetarypolicy/fomcminutes20260916.htm)
5. [Federal Reserve: summary of economic projections (2026-09-16)](https://www.federalreserve.gov/monetarypolicy/fomcprojtabl20260916.htm)
6. [FRED: 10-year Treasury constant maturity yield (DGS10)](https://fred.stlouisfed.org/series/DGS10)
7. [FRED: 10-year breakeven inflation rate (T10YIE)](https://fred.stlouisfed.org/series/T10YIE)
8. [FRED: 10-year inflation-indexed Treasury yield (DFII10)](https://fred.stlouisfed.org/series/DFII10)
9. [Federal Reserve Bank of New York: ACM term premium data](https://www.newyorkfed.org/medialibrary/media/research/data_indicators/ACMTermPremium.xls)
10. [Federal Reserve: transcript of the Chair's press conference (2026-09-16)](https://www.federalreserve.gov/mediacenter/files/FOMCpresconf20260916.pdf)
11. [The Fiscal Times: deficit rose to $2 trillion for 2026 fiscal year, CBO (2026-10-08)](https://www.thefiscaltimes.com/2026/10/08/Deficit-Rose-2-Trillion-2026-Fiscal-Year-CBO)
12. [US Treasury: quarterly refunding statement (2026-08-05)](https://home.treasury.gov/news/press-releases/sb0590)
13. [FRED: 2-year Treasury constant maturity yield (DGS2)](https://fred.stlouisfed.org/series/DGS2)
14. [FRED: Freddie Mac 30-year fixed mortgage rate (MORTGAGE30US)](https://fred.stlouisfed.org/series/MORTGAGE30US)
15. [FRED: ICE BofA US high yield option-adjusted spread (BAMLH0A0HYM2)](https://fred.stlouisfed.org/series/BAMLH0A0HYM2)
16. [FactSet: Earnings Insight (2026-10-09)](https://advantage.factset.com/hubfs/Website/Resources%20Section/Research%20Desk/Earnings%20Insight/EarningsInsight_100926.pdf)
17. [Yahoo Finance: daily closing prices](https://finance.yahoo.com/)
18. [Campbell's 10-K for fiscal 2026](https://www.sec.gov/Archives/edgar/data/16732/000001673226000026/cpb-20260802.htm)
19. [Investing.com: Moody's cuts Campbell's rating on weak earnings, high leverage (2026-09-23)](https://www.investing.com/news/stock-market-news/moodys-cuts-campbells-rating-on-weak-earnings-high-leverage-93CH-4913757)
20. [Wynn Resorts 10-Q for the second quarter of 2026](https://www.sec.gov/Archives/edgar/data/1174922/000117492226000055/wynn-20260630.htm)
21. [Clorox 10-K for fiscal 2026](https://www.sec.gov/Archives/edgar/data/21076/000002107626000034/clx-20260630.htm)
22. [Sucro: Q2 2026 financial statements](https://sucro.us/wp-content/uploads/2026/08/Sucro-Limited-Q2-2026-Financial-Statements-Signed.pdf)
23. [PG&E 10-Q for the second quarter of 2026](https://www.sec.gov/Archives/edgar/data/1004980/000100498026000048/pcg-20260630.htm)
24. [Yum! Brands 10-Q for the second quarter of 2026](https://www.sec.gov/Archives/edgar/data/1041061/000104106126000160/yum-20260630.htm)
25. [Pool Corporation 10-Q for the second quarter of 2026](https://www.sec.gov/Archives/edgar/data/945841/000119312526322532/pool-20260630.htm)
26. [RELX: 2026 interim results](https://www.sec.gov/Archives/edgar/data/929869/000092986926000054/relx-20260723xex99d1.htm)
27. [Brookfield Corporation: Q2 2026 interim report](https://www.sec.gov/Archives/edgar/data/1001085/000100108526000021/bn-20260630.htm)
28. [Intercontinental Exchange 10-Q for the second quarter of 2026](https://www.sec.gov/Archives/edgar/data/1571949/000157194926000011/ice-20260630.htm)
29. [Charles Schwab 10-Q for the second quarter of 2026](https://www.sec.gov/Archives/edgar/data/316709/000031670926000031/schw-20260630.htm)
30. [Copart 10-K for fiscal 2026](https://www.sec.gov/Archives/edgar/data/900075/000119312526405731/cprt-20260731.htm)
31. [Fox Business: September 2026 jobs report and CME FedWatch (2026-10-02)](https://www.foxbusiness.com/economy/us-jobs-report-september-2026)
32. [FRED: consumer price index less food and energy (CPILFESL)](https://fred.stlouisfed.org/series/CPILFESL)
33. [FRED: consumer price index for energy (CPIENGSL)](https://fred.stlouisfed.org/series/CPIENGSL)
34. [BEA: personal income and outlays, August 2026 (2026-09-30)](https://www.bea.gov/news/2026/personal-income-and-outlays-august-2026)
35. [CFR Global Conflict Tracker: confrontation between the United States and Iran](https://www.cfr.org/global-conflict-tracker/conflict/confrontation-between-united-states-and-iran)
36. [FRED: WTI crude oil price (DCOILWTICO)](https://fred.stlouisfed.org/series/DCOILWTICO)
37. [FRED: 5-year, 5-year forward inflation expectation rate (T5YIFR)](https://fred.stlouisfed.org/series/T5YIFR)
38. [Federal Reserve: speech by Governor Waller (2026-09-03)](https://www.federalreserve.gov/newsevents/speech/waller20260903a.htm)
39. [Minneapolis Fed: the pass-through of tariffs to consumer prices has arrived (2026-08-28)](https://www.minneapolisfed.org/article/2026/initially-delayed-the-pass-through-of-tariffs-to-consumer-prices-has-arrived)
40. [Federal Reserve: speech by Governor Cook (2026-09-28)](https://www.federalreserve.gov/newsevents/speech/cook20260928a.htm)

## Who pays for the AI buildout when tokens become cheap

Date: 2026-10-07. Page: https://picks.kamadorueda.com/thoughts/ai-commoditization-and-the-buildout/

If tokens become priced like a commodity, I expect the losses to land on Oracle, the neoclouds and their lenders. I expect Microsoft, Alphabet, Amazon and Meta to absorb the change with little damage. The contracts and the debt support this view. But the trigger has not happened yet, and lab revenue and GPU rents both increased in 2026. I give four scenarios with my odds, the effect of each on the companies on Picks, and a model of the return on one GPU.

### Five links from token prices to investors

My view needs five links to hold. Two links have evidence today. Three links have no evidence yet. Microsoft, Alphabet, Amazon, Meta and Oracle spend about $200B a quarter on chips and buildings, and some of them already spend more cash than they make [1][2][3][4][5]. Figures marked "my estimate" or "my arithmetic" are my own calculations on company filings.

Link 1, partly evidenced: token prices fall toward cost. The cost of a fixed level of capability falls about 5 to 10 times each year [6]. In 2026 the best open-weight models, most of them Chinese, were about four months behind the best closed models on average [7]. Ramp sees company card spending. It reports that the effective price its customers paid per million tokens fell 41%, from a peak of $1.15 in March 2026 to $0.68 in September [8]. The frontier is the exception. OpenAI priced GPT-5.5 at double the price of its predecessor in April 2026, according to pricing trackers [9]. Frontier models also still made up 45% of the tokens in Ramp's September data [8].

Link 2, not evidenced yet: cheaper tokens decrease the labs' margins. The FT reported in September that Anthropic's gross margin is above 80% before the revenue it shares with partners, and that Anthropic expects an adjusted operating profit this quarter [10]. A leaked draft of Anthropic's prospectus shows revenue of $11.5B in the June quarter, up from $4.73B in March [11]. OpenAI is weaker. Leaked statements show a $20.9B operating loss on $13.1B of revenue in 2025 [12]. Its gross margin was about 43% (my arithmetic on the leaked figures).

Link 3, not tested: lower margins make the labs' commitments hard to pay. The same leaked draft shows $518B of planned spending on cloud services and data centers [11]. That is about 8 times the $65B annual run rate that the press reported for July [13].

Link 4, concentration evidenced: trouble at the labs goes into the clouds' order books, which total about $2.5 trillion (table below). Microsoft said that OpenAI was about 45% of its order book at the end of 2025 [14]. The Information reported in May that the two labs are about half of the combined order books of Amazon, Microsoft, Google and Oracle [15]. Oracle's order book changes into revenue most slowly. Oracle expects only about 13% of it within twelve months, against about 30% at Microsoft [5][16].

**The clouds' order books (remaining performance obligations) at their latest quarter**

| Company | Order book | Lab exposure | When it turns into revenue |
| --- | --- | --- | --- |
| Microsoft | $678B, June 30, 2026 [16] | OpenAI about 45% at December 31, 2025 [14] | about 30% within 12 months [16] |
| Oracle | $664B, August 31, 2026 [5] | OpenAI about half, an analyst estimate in press reports [20] | about 13% within 12 months, 37% in months 13 to 36 [5] |
| Google Cloud | $514B, June 30, 2026 [21] | not disclosed | just over 50% within 24 months [21] |
| AWS | $496B, June 30, 2026 [22] | in 2026, the OpenAI deal grew by $100B over 8 years and the Anthropic deal by more than $100B over 10 years [22] | average remaining contract life of 6.4 years [22] |
| CoreWeave | $104B, June 30, 2026 [23] | largest customer 36% of June-quarter revenue [23] | not disclosed |

Note: Order books are contracted revenue that the company has not yet recognized. The OpenAI shares are the latest that the companies or the press gave, and their dates are different.

Link 5, structure evidenced: trouble in the order books goes to data center owners and lenders. At the end of August, Oracle had $288B of signed data center leases that have not started, with terms of 15 to 19 years [5]. Its main OpenAI contract is reported at about $300B over about five years from 2027 [17]. The Bank of England expects private credit to supply about $700B of data center spending in 2026 to 2028 [18].

Compute is still scarce, and this fact goes against the whole chain. H100 rents increased from just under $2 an hour in January 2026 to near $3 by August, according to the Silicon Data index. The same report gives B200 rents of $5.50 to $5.80 [19]. While rents stay high and the labs pay, the clouds get paid at any token price. Today the chain breaks at links 2 and 3.

Companies: [MSFT](https://picks.kamadorueda.com/company/MSFT.md), [GOOGL](https://picks.kamadorueda.com/company/GOOGL.md), [AMZN](https://picks.kamadorueda.com/company/AMZN.md), [META](https://picks.kamadorueda.com/company/META.md)

### Clouds already sell commodities at a profit

A commodity can be a very good business for a cloud. AWS uses mostly commodity parts and open-source software. Its operating margin was 37.0% in 2024 and 35.4% in 2025 (my arithmetic on Amazon's annual report) [24].

S3 is the clearest example. It sells disk space by the gigabyte. S3 Standard has a list price of $0.023 per GB a month for the first 50 TB in Amazon's Virginia region, and $0.021 above 500 TB [25]. At the first price, that is $276 per terabyte a year. Data sent out to the internet costs $0.09 per GB more [26]. A large hard drive cost about $28 to $41 per terabyte at retail in the second half of 2026, according to a price tracker [27]. S3 keeps each object in a minimum of three data centers [28], so each stored terabyte needs more disks, servers, power and staff. Even so, the hardware for it probably costs a few dollars a month, against a list price of $23 (my rough estimate; Amazon does not report storage profit). Amazon says that the price is about 85% lower than at launch in 2006, and that S3 now holds more than 500 trillion objects [29]. After twenty years of price cuts, S3 is still part of a business with a 35% margin.

Servers and databases work in the same way. An EC2 server is a commodity CPU that customers rent by the hour. Amazon depreciates it over five years, after it decreased the life of some servers from six years in 2025 [24]. The managed layer on top earns a premium. A managed PostgreSQL database (RDS db.r6g.xlarge) costs about $0.45 an hour on demand. The same machine as a plain EC2 server costs $0.202 [30][31]. The premium of 2.2 times pays for backups, patches and failover.

The clouds also sell open-weight models as a managed product. Amazon Bedrock, Google Cloud's Vertex AI Model Garden and Microsoft Foundry host DeepSeek, Alibaba's Qwen, Meta's Llama, Mistral, Kimi, MiniMax and OpenAI's gpt-oss, priced per million tokens. Their prices are close to the model makers' own prices, and frequently higher. In prices published between late September and early October 2026, DeepSeek V3.2 costs $0.62 per million input tokens and $1.85 per million output tokens on Bedrock, $0.56 and $1.68 on Google, and $0.62 and $1.85 on Microsoft [32][33][34]. DeepSeek's own API charges $0.30 and $1.20 at peak hours for its newer V4.1 Flash model, and half that off peak [35]. Bedrock lists MiniMax M2.5 at $0.30 and $1.20 [32], about the price that MiniMax asked for a similar model in February [36]. Buyers pay more for one bill, for data that stays in their cloud account and for security reviews that are already done (my reading).

A commodity pays its owner when the asset lasts long, stays busy, sells to many small customers and has a managed layer on top. Storage and CPUs have all four. GPUs have two. Use is high today, and the open-model prices show that the managed layer exists. But a GPU loses most of its value in three to five years [37][38]. Also, much of its revenue comes from a few large labs that negotiate hard. A GPU sold as a managed model to thousands of companies is similar to S3. A GPU rented bare to one lab with borrowed money is a bet on rents.

**What clouds charge for commodity products, against the cost of the hardware**

| Product | Price | Cost or comparison | Price over cost |
| --- | --- | --- | --- |
| Managed database (RDS vs EC2, r6g.xlarge, on demand) | about $0.45 an hour [30] | $0.202 an hour for the plain server [31] | 2.2 times |
| Object storage (S3 Standard, first 50 TB) | $23 per TB a month [25] | a few dollars per TB a month for the hardware, from drives at $28 to $41 per TB [27] (my estimate) | several times (my estimate) |
| GPU rent, H100 | about $3 an hour, August 2026 [19] | about $1.17 an hour, owned at hyperscaler scale [39] | about 2.6 times |
| GPU rent, B200 vs owned GB200 | $5.50 to $5.80 an hour, August 2026 [19] | about $1.86 an hour, owned at hyperscaler scale [39] | about 3 times |
| Open model (MiniMax on Bedrock) | about $0.40 per million tokens, with 8 input tokens for each output token [32] | $0.06 to $0.19 per million tokens on an owned GB200 at full load [39] | 2 to 7 times |

Note: Ratios are my arithmetic. SemiAnalysis's cost figures are for a MiniMax M2.7 model in July 2026 terms. Bedrock's price is for MiniMax M2.5. Costs do not include the clouds' staff and software. The storage cost is a rough estimate.

Companies: [AMZN](https://picks.kamadorueda.com/company/AMZN.md), [GOOGL](https://picks.kamadorueda.com/company/GOOGL.md), [MSFT](https://picks.kamadorueda.com/company/MSFT.md), [BABA](https://picks.kamadorueda.com/company/BABA.md)

### What one GPU earns

A rented GPU pays for itself only if rents and use stay high for most of its life. I made a simple model of one GB200-class GPU on my own assumptions, which the table lists. The base case uses a rent of $3.50 an hour and 85% use over five years. That rent is below today's spot rents, because long contracts sell for less [40]. In the base case, the GPU returns about 18.6% a year before tax and pays back in about 3.1 years. A hyperscaler's cost of capital is about 8% to 10% (my assumption).

**One GB200-class GPU: base case assumptions and results (my model)**

| Item | Base case | Basis |
| --- | --- | --- |
| Equipment cost | $55,000 per GPU | between rack prices of about $3M (GB200) and $5M (GB300) for 72 GPUs [40][47] |
| Power draw | 1.9 kW per GPU | Wing VC [40] |
| Rent | $3.50 per GPU-hour | my assumption, below spot rents of $5.50 to $5.80 for B200 [19] |
| Use | 85% of hours | my assumption |
| Useful life | 5 years, no resale value | my assumption |
| Electricity | $0.08 per kWh with a 1.3 overhead factor: $1,575 a year | my assumption |
| Data center space | $196 per kW a month: $4,469 a year | Wing VC [40] |
| Other costs (staff, network, software) | 4% of equipment cost: $2,200 a year | my assumption |
| Revenue | $26,061 a year, $0.47 per dollar of equipment | result |
| Cash margin before tax | $17,817 a year (68% of revenue) | result |
| Operating profit after depreciation | $6,817 a year (26% of revenue) | result |
| Yearly return (IRR, before tax) | 18.6% | result |
| Payback | 3.1 years | result |

Note: The model assumes that idle GPUs use 40% of full power. Running costs are about $1.11 per hour of use, or $2.58 with depreciation.

Rent and use have the largest effect. The chart shows the yearly return at each rent for four levels of use. At 85% use, the GPU earns more than a 9% cost of capital at about $3.01 an hour. At 70% use it needs about $3.62, and at 60% use about $4.21. Morgan Stanley warns that returns "evaporate rapidly" below about 70% use or below about $1 per million tokens, according to a press summary [41].

**Yearly return on one GB200-class GPU by rent, at four levels of use (my model)**

Line chart, Yearly return (IRR, before tax) by Rent per GPU-hour.

| Rent per GPU-hour | 60% use | 70% use | 85% use | 95% use |
| --- | --- | --- | --- | --- |
| $2.5/h | -20.9% | -12.5% | -1.9% | 4.3% |
| $2.75/h | -15.4% | -7.1% | 3.6% | 10.1% |
| $3/h | -10.5% | -2.1% | 8.9% | 15.5% |
| $3.25/h | -6% | 2.5% | 13.8% | 20.8% |
| $3.5/h | -1.8% | 6.9% | 18.6% | 25.8% |
| $3.75/h | 2.2% | 11.1% | 23.1% | 30.6% |
| $4/h | 6% | 15.1% | 27.6% | 35.4% |
| $4.25/h | 9.6% | 19% | 31.9% | 40% |
| $4.5/h | 13.1% | 22.7% | 36.1% | 44.5% |
| $4.75/h | 16.5% | 26.4% | 40.2% | 49% |
| $5/h | 19.8% | 30% | 44.3% | 53.3% |
| $5.25/h | 23% | 33.5% | 48.3% | 57.7% |
| $5.5/h | 26.1% | 36.9% | 52.2% | 61.9% |
| $5.75/h | 29.2% | 40.3% | 56.1% | 66.2% |
| $6/h | 32.2% | 43.7% | 59.9% | 70.3% |

Reference line: Cost of capital (9%) at 9%.

Note: The base case is the same as in the table above: $55,000 of equipment, a five-year life, no resale value, power at $0.08 per kWh and space at $196 per kW a month [40], before tax. Rents stay flat for five years.

Time is the other danger. The second table changes one input at a time. If rent falls 10% a year, the return falls from 18.6% to 7.0%, below the cost of capital. If rent falls 20% a year, the return is negative. AWS decreased on-demand H100 prices by 44% in June 2025 [42]. A three-year life loses money, and a six-year life earns 23.1%.

**Yearly return when one input changes from the base case (my model)**

| Change from the base case | Yearly return |
| --- | --- |
| Rent flat at $3.50 (base case) | 18.6% |
| Rent falls 10% a year | 7.0% |
| Rent falls 20% a year | -7.3% |
| Rent falls 30% a year | -25.4% |
| Useful life of 3 years | -1.4% |
| Useful life of 4 years | 11.2% |
| Useful life of 5 years (base case) | 18.6% |
| Useful life of 6 years | 23.1% |
| Equipment costs $45,000 | 29.3% |
| Equipment costs $55,000 (base case) | 18.6% |
| Equipment costs $70,000 | 7.3% |
| Equipment costs $80,000 | 1.7% |

Note: All other inputs are as in the base case: $3.50 an hour, 85% use, five years and $55,000. When rents fall, an owner stops renting when cash costs are more than revenue.

Reported numbers already show the strain. AWS earned $1.25 of revenue per dollar of net property and equipment in 2023 and $0.68 in 2025, as AI equipment entered the base. Its pre-tax operating return on average net property fell from 43% in 2024 to 30% in 2025. Both figures are my arithmetic on Amazon's annual report [24]. Press reports in October 2025, based on leaked documents, put Oracle's gross margin on renting Nvidia servers at about 14% [43]. Oracle replied that such contracts reach 30% to 40% over their life [44]. CoreWeave is a pure renter. In the June 2026 quarter it had revenue of $2.58B, an operating loss of $49M and interest expense of $640M, and its delayed-draw loans have effective rates of 9% to 15% [23]. Its adjusted EBITDA margin was about 59% [45]. The gap between that margin and the operating loss shows how much of a renter's cash goes to depreciation and interest.

A hyperscaler has more control of use and life than a pure renter. It can fill idle GPUs with its own work and move older GPUs to cheaper inference. It can sign contracts that run "for the entire useful life of the GPU", as Microsoft says it does for its largest customers [46]. It can also sell the managed layer on the same chips. A renter that borrows at 9% or more while rents fall 20% a year has none of these tools.

Companies: [MSFT](https://picks.kamadorueda.com/company/MSFT.md), [AMZN](https://picks.kamadorueda.com/company/AMZN.md)

### Four scenarios and their effect on each company

Most companies on Picks pay for AI data centers or sell something to them. I give each scenario below my own odds, to the end of 2028. Each scenario names the companies on Picks that gain or lose, and says if the loss is to the balance sheet or to earnings power.

**The four scenarios at a glance (my odds, to the end of 2028)**

| Scenario | My odds | Token prices | GPU rents | Who takes real losses |
| --- | --- | --- | --- | --- |
| A: frontier prices stay high | 30% | flat or rising for the best models | near today's level | almost no one |
| B: prices fall slowly | 40% | falling, with volume growing faster | falling about 10% a year | neocloud shareholders |
| C: prices collapse | 20% | falling fast; labs renegotiate | falling 20% to 30% a year | Oracle, neoclouds, private credit, holders of lab stakes |
| D: demand stops growing | 10% | less important than volume | falling because capacity is idle | chip makers, owners of idle buildings and their lenders |

Twenty-two companies on the list have little direct AI exposure: UNH, POOL, AZO, CPRT, PG, CLX, BTI, WYNN, WM, HSY, CPB, PEP, COST, RBGLY, MCD, YUM, YUMC, OTIS, HONA, NVO, MRX and SUGR. In all scenarios, the economy and interest rates move them more than token prices do.

Companies: [UNH](https://picks.kamadorueda.com/company/UNH.md), [POOL](https://picks.kamadorueda.com/company/POOL.md), [AZO](https://picks.kamadorueda.com/company/AZO.md), [CPRT](https://picks.kamadorueda.com/company/CPRT.md), [PG](https://picks.kamadorueda.com/company/PG.md), [CLX](https://picks.kamadorueda.com/company/CLX.md), [BTI](https://picks.kamadorueda.com/company/BTI.md), [WYNN](https://picks.kamadorueda.com/company/WYNN.md), [WM](https://picks.kamadorueda.com/company/WM.md), [HSY](https://picks.kamadorueda.com/company/HSY.md), [CPB](https://picks.kamadorueda.com/company/CPB.md), [PEP](https://picks.kamadorueda.com/company/PEP.md), [COST](https://picks.kamadorueda.com/company/COST.md), [RBGLY](https://picks.kamadorueda.com/company/RBGLY.md), [MCD](https://picks.kamadorueda.com/company/MCD.md), [YUM](https://picks.kamadorueda.com/company/YUM.md), [YUMC](https://picks.kamadorueda.com/company/YUMC.md), [OTIS](https://picks.kamadorueda.com/company/OTIS.md), [HONA](https://picks.kamadorueda.com/company/HONA.md), [NVO](https://picks.kamadorueda.com/company/NVO.md), [MRX](https://picks.kamadorueda.com/company/MRX.md), [SUGR](https://picks.kamadorueda.com/company/SUGR.md)

### Scenario A: frontier prices stay high (30%)

What must be true: Anthropic's audited gross margin is near the reported 80% [10]. Frontier prices stay flat or increase. GPU rents stay near today's level [19].

The spenders gain in earnings and in balance sheet. MSFT, GOOGL and AMZN change their order books into revenue at today's prices. Amazon and Alphabet continue to book gains on their stakes in Anthropic and OpenAI. Investors judge META's spending by its ads and its new paid products. BABA gains less, because in China open models already set the price, and its own Qwen is one of them.

The chip chain keeps its pricing power. NVDA keeps its gross margin. AVGO continues to grow custom AI chips, which include Google's TPUs for Anthropic. Its AI chip revenue was $16.7B in the quarter to August 2, up 221%, with $21.7B guided for the next quarter [48]. TSM continues to increase prices. High-performance computing, which includes AI chips, was 66% of its revenue in the June quarter, according to press reports [49]. MRVL's custom chip programs and QCOM's new inference chips for AWS get time to grow. Qualcomm's filings put the AWS deal at up to $60B of purchases over ten years [50]. SNPS sells design software to all of them.

Power companies sign more long contracts. CEG signed a 20-year, 690 MW agreement with Amazon at Calvert Cliffs on September 30 [51]. On October 6 it signed a 20-year agreement with Google for 890 MW of new nuclear capacity [52]. PCG connects more of its data center pipeline. VNOM gains little, because its royalties follow oil more than the gas for new power plants (my judgment).

Brookfield and lenders do well. BN's asset manager raises more AI money, CoreWeave refinances and Oracle stays investment grade.

Software and data companies keep their AI fear, and no price war starts. Expensive frontier tokens make it costly for a competitor to copy NOW, SAP, INTU, CSU, RELX, EXPGY, BR or SCHW (my judgment). SPGI continues to rate a large quantity of new debt. Its ratings revenue increased 17% in the June quarter [53]. ICE handles more bond trading and data on the same new debt (my judgment). APP runs its own ad models, so token prices have almost no effect on it.

Companies: [MSFT](https://picks.kamadorueda.com/company/MSFT.md), [GOOGL](https://picks.kamadorueda.com/company/GOOGL.md), [AMZN](https://picks.kamadorueda.com/company/AMZN.md), [META](https://picks.kamadorueda.com/company/META.md), [BABA](https://picks.kamadorueda.com/company/BABA.md), [NVDA](https://picks.kamadorueda.com/company/NVDA.md), [AVGO](https://picks.kamadorueda.com/company/AVGO.md), [TSM](https://picks.kamadorueda.com/company/TSM.md), [MRVL](https://picks.kamadorueda.com/company/MRVL.md), [QCOM](https://picks.kamadorueda.com/company/QCOM.md), [SNPS](https://picks.kamadorueda.com/company/SNPS.md), [CEG](https://picks.kamadorueda.com/company/CEG.md), [PCG](https://picks.kamadorueda.com/company/PCG.md), [VNOM](https://picks.kamadorueda.com/company/VNOM.md), [BN](https://picks.kamadorueda.com/company/BN.md), [NOW](https://picks.kamadorueda.com/company/NOW.md), [SAP](https://picks.kamadorueda.com/company/SAP.md), [INTU](https://picks.kamadorueda.com/company/INTU.md), [CSU](https://picks.kamadorueda.com/company/CSU.md), [RELX](https://picks.kamadorueda.com/company/RELX.md), [EXPGY](https://picks.kamadorueda.com/company/EXPGY.md), [BR](https://picks.kamadorueda.com/company/BR.md), [SCHW](https://picks.kamadorueda.com/company/SCHW.md), [SPGI](https://picks.kamadorueda.com/company/SPGI.md), [ICE](https://picks.kamadorueda.com/company/ICE.md), [APP](https://picks.kamadorueda.com/company/APP.md)

### Scenario B: prices fall slowly and volume grows faster (40%)

What must be true: token prices continue to fall for most tasks, and volume grows faster than prices fall. In the year to May 2026, Google's monthly tokens increased about 7 times [54]. Rents fall about 10% a year. Lab margins decrease, and the labs still pay.

The spenders lose some earnings power, and their balance sheets stay strong. New GPUs earn returns near those of a utility, 7.0% in the model at $3.50 with a 10% decrease each year. Depreciation increases, and order books change into revenue more slowly. AMZN and GOOGL are the least affected, because they sell managed open models and use their own chips, Trainium and TPU. MSFT has the most risk from one customer, because OpenAI is about 45% of its order book [14]. META gains, because cheaper models decrease the cost of its own products. BABA gets more cloud volume. Alibaba Cloud's revenue grew 45% in the June quarter, while capital spending increased 75% [55]. Alibaba also gives away the Qwen weights, and they help push prices down for all companies, Alibaba included.

The chip chain sells more units at a lower price per unit of compute. NVDA's gross margin is the main risk. AVGO, MRVL and QCOM gain as buyers move inference to cheaper custom chips. TSM makes almost all of these chips, so it gains whichever company wins, and SNPS gains from more custom designs (my judgment).

Power continues to grow, a little more slowly. CEG's and PCG's signed contracts stay in force, and VNOM sees no change.

Brookfield and lenders survive. Oracle stays one step above junk [20]. Neocloud shares do badly, and their debt is paid. BN's fees continue to grow at a slower rate.

Software and data companies take the most damage in B. Cheap tokens make AI features cheap for incumbents to add and cheap for competitors to copy. Some companies hold records or data that a model cannot copy. They can use cheap models to decrease their own costs and sell AI on top. ServiceNow does this, and its AI products passed $1B of annual contract value in the June quarter [56]. This group is NOW, SAP and CSU (systems that run daily operations), RELX (legal and science content), EXPGY (credit files), and SPGI and ICE (benchmarks and market data). INTU faces free tax help from AI tools. BR and SCHW face AI assistants that make clients less loyal. APP's ad engine costs less to run.

Companies: [MSFT](https://picks.kamadorueda.com/company/MSFT.md), [GOOGL](https://picks.kamadorueda.com/company/GOOGL.md), [AMZN](https://picks.kamadorueda.com/company/AMZN.md), [META](https://picks.kamadorueda.com/company/META.md), [BABA](https://picks.kamadorueda.com/company/BABA.md), [NVDA](https://picks.kamadorueda.com/company/NVDA.md), [AVGO](https://picks.kamadorueda.com/company/AVGO.md), [TSM](https://picks.kamadorueda.com/company/TSM.md), [MRVL](https://picks.kamadorueda.com/company/MRVL.md), [QCOM](https://picks.kamadorueda.com/company/QCOM.md), [SNPS](https://picks.kamadorueda.com/company/SNPS.md), [CEG](https://picks.kamadorueda.com/company/CEG.md), [PCG](https://picks.kamadorueda.com/company/PCG.md), [VNOM](https://picks.kamadorueda.com/company/VNOM.md), [BN](https://picks.kamadorueda.com/company/BN.md), [NOW](https://picks.kamadorueda.com/company/NOW.md), [SAP](https://picks.kamadorueda.com/company/SAP.md), [INTU](https://picks.kamadorueda.com/company/INTU.md), [CSU](https://picks.kamadorueda.com/company/CSU.md), [RELX](https://picks.kamadorueda.com/company/RELX.md), [EXPGY](https://picks.kamadorueda.com/company/EXPGY.md), [BR](https://picks.kamadorueda.com/company/BR.md), [SCHW](https://picks.kamadorueda.com/company/SCHW.md), [SPGI](https://picks.kamadorueda.com/company/SPGI.md), [ICE](https://picks.kamadorueda.com/company/ICE.md), [APP](https://picks.kamadorueda.com/company/APP.md)

### Scenario C: prices collapse and labs renegotiate (20%)

What must be true: open models close the four-month gap [7]. Buyers move work to the cheapest model that is sufficient. Lab margins fall toward OpenAI's level. Rents fall 20% to 30% a year as supply grows to meet demand. OpenAI, or both labs, renegotiate their contracts.

The spenders lose balance sheet through lab stakes and earnings through depreciation. AMZN and GOOGL write down their Anthropic and OpenAI stakes. A 50% cut would remove about 11% and 10% of their book equity (my estimate, details below). MSFT loses value on its OpenAI stake and on part of the 45% of its order book that comes from OpenAI. META has the smallest loss, because it has no lab customers. It still owes rent on the Hyperion campus [57]. BABA is one of the causes of this scenario. Cheaper Chinese tokens win share, while its own return on capital falls.

The chip chain loses revenue, and some chip companies also lose balance sheet. NVDA's revenue falls, as it did by 28% in two quarters in 2022 [58]. Its equity investments and its guarantee of OpenAI leases cause losses. AVGO depends on Anthropic and supports a $35B TPU loan [59]. TSM gets fewer orders. At June 30 it held about NT$2.1 trillion more cash than debt (my arithmetic on its report [60]), so its loss stays in earnings and its solvency is safe. MRVL's and QCOM's new programs are late, and QCOM's AI revenue can fail to start.

Power companies lose growth, and their contracts stay in force. CEG's growth target and valuation fall. Its 20-year contracts with Microsoft, Meta, Amazon and Google hold, because those buyers have ratings of AA- or better [61]. PCG's pipeline becomes smaller without losses, and VNOM loses a little gas demand (my judgment).

Brookfield and lenders take most of the losses. Rating agencies cut Oracle to junk, and Oracle writes down equipment. Neocloud shares lose almost all their value, and their lenders get back part of their money. Private credit takes losses. BN's fee growth slows, with little loss of its own capital. Most of the money at risk belongs to its fund investors, and the compute deals are take-or-pay with investment-grade tenants (see below). SPGI and ICE lose issuance and trading from AI debt. Downgrades and distressed trading replace part of it (my judgment).

Software and data companies see the disruption fear come true where a product is a thin layer on top of a model. In my view, RELX, INTU, EXPGY and BR are the most exposed. RELX fell 14% in one day in February when Anthropic launched a legal tool [62]. NOW, SAP and CSU are less exposed, because customers cannot easily remove the systems that they run on. SCHW faces cheap AI advice. APP gains, because its own models cost less to run.

Companies: [MSFT](https://picks.kamadorueda.com/company/MSFT.md), [GOOGL](https://picks.kamadorueda.com/company/GOOGL.md), [AMZN](https://picks.kamadorueda.com/company/AMZN.md), [META](https://picks.kamadorueda.com/company/META.md), [BABA](https://picks.kamadorueda.com/company/BABA.md), [NVDA](https://picks.kamadorueda.com/company/NVDA.md), [AVGO](https://picks.kamadorueda.com/company/AVGO.md), [TSM](https://picks.kamadorueda.com/company/TSM.md), [MRVL](https://picks.kamadorueda.com/company/MRVL.md), [QCOM](https://picks.kamadorueda.com/company/QCOM.md), [SNPS](https://picks.kamadorueda.com/company/SNPS.md), [CEG](https://picks.kamadorueda.com/company/CEG.md), [PCG](https://picks.kamadorueda.com/company/PCG.md), [VNOM](https://picks.kamadorueda.com/company/VNOM.md), [BN](https://picks.kamadorueda.com/company/BN.md), [NOW](https://picks.kamadorueda.com/company/NOW.md), [SAP](https://picks.kamadorueda.com/company/SAP.md), [INTU](https://picks.kamadorueda.com/company/INTU.md), [CSU](https://picks.kamadorueda.com/company/CSU.md), [RELX](https://picks.kamadorueda.com/company/RELX.md), [EXPGY](https://picks.kamadorueda.com/company/EXPGY.md), [BR](https://picks.kamadorueda.com/company/BR.md), [SCHW](https://picks.kamadorueda.com/company/SCHW.md), [SPGI](https://picks.kamadorueda.com/company/SPGI.md), [ICE](https://picks.kamadorueda.com/company/ICE.md), [APP](https://picks.kamadorueda.com/company/APP.md)

### Scenario D: demand stops growing (10%)

What must be true: adoption stops spreading. Only 17% to 20% of US firms used AI in any business function between December 2025 and May 2026, according to the Census Bureau [63]. In this scenario, that share stops increasing.

The spenders cut capital spending. MSFT, GOOGL, AMZN, META and BABA lose growth and write down idle capacity. Their bondholders gain as free cash flow recovers.

The chip chain has the largest loss. NVDA, AVGO, MRVL and QCOM lose orders. TSM fills fewer advanced wafers, and SNPS sees fewer new designs.

Power companies lose their growth. CEG and PCG grow more slowly, and VNOM loses a source of gas demand.

Brookfield and lenders hold contracts on idle buildings. BN's AI fund stops raising money, and its valuation falls. Lenders' losses are similar to those in C, and they arrive more slowly.

Software and data companies gain against the spenders, because the AI fear becomes smaller for NOW, SAP, INTU, CSU, RELX, EXPGY, BR and SCHW. SPGI and ICE lose the increase in AI debt issuance, and APP sees little change. If a stall also slowed the economy, interest rates would have a larger effect than AI on SCHW and ICE (my judgment).

Companies: [MSFT](https://picks.kamadorueda.com/company/MSFT.md), [GOOGL](https://picks.kamadorueda.com/company/GOOGL.md), [AMZN](https://picks.kamadorueda.com/company/AMZN.md), [META](https://picks.kamadorueda.com/company/META.md), [BABA](https://picks.kamadorueda.com/company/BABA.md), [NVDA](https://picks.kamadorueda.com/company/NVDA.md), [AVGO](https://picks.kamadorueda.com/company/AVGO.md), [TSM](https://picks.kamadorueda.com/company/TSM.md), [MRVL](https://picks.kamadorueda.com/company/MRVL.md), [QCOM](https://picks.kamadorueda.com/company/QCOM.md), [SNPS](https://picks.kamadorueda.com/company/SNPS.md), [CEG](https://picks.kamadorueda.com/company/CEG.md), [PCG](https://picks.kamadorueda.com/company/PCG.md), [VNOM](https://picks.kamadorueda.com/company/VNOM.md), [BN](https://picks.kamadorueda.com/company/BN.md), [NOW](https://picks.kamadorueda.com/company/NOW.md), [SAP](https://picks.kamadorueda.com/company/SAP.md), [INTU](https://picks.kamadorueda.com/company/INTU.md), [CSU](https://picks.kamadorueda.com/company/CSU.md), [RELX](https://picks.kamadorueda.com/company/RELX.md), [EXPGY](https://picks.kamadorueda.com/company/EXPGY.md), [BR](https://picks.kamadorueda.com/company/BR.md), [SCHW](https://picks.kamadorueda.com/company/SCHW.md), [SPGI](https://picks.kamadorueda.com/company/SPGI.md), [ICE](https://picks.kamadorueda.com/company/ICE.md), [APP](https://picks.kamadorueda.com/company/APP.md)

### How large the losses can be

Hyperscaler shareholders: the loss is mostly to earnings power, with one large balance-sheet channel. At mid-2026, the four companies held about $1.31 trillion of net property and equipment against about $1.90 trillion of equity [16][21][22][64]. If server lives decrease to four years, depreciation increases by about $36B a year, about 7% of their combined operating income (my estimate). If the lab contracts fall by half, the operating income of each company falls by about 4% to 5% a year (my estimate). The lab stakes are the fastest channel. At June 30, Amazon had $122.3B of private stakes, mainly in Anthropic and OpenAI, up from $16.2B six months before [22]. Alphabet had $124.3B of non-marketable equity [21]. A 50% write-down would decrease book equity by about 11% at Amazon and 10% at Alphabet, before tax (my estimate). Microsoft uses the equity method for OpenAI [16], so its book loss is small. But if OpenAI's value fell by 50%, Microsoft's 27% stake would lose about $67B of market value. That stake had a value of about $135B in October 2025 [65] (my arithmetic).

Hyperscaler bondholders: the risk is to bond prices. In May 2026, S&P rated Microsoft AAA, Alphabet AA+, Amazon AA and Meta AA- [61]. If EBITDA fell 30% and debt doubled, their debt would go from 0.2 to 0.8 times EBITDA at mid-2026 to only about 0.7 to 2.2 times (my estimates from filings). The risk comes from supply. Debt paid for 32% of their capital spending in the twelve months to mid-2026, against 9% in fiscal 2024 [66]. Spreads have increased only about 10 basis points [67].

Oracle: the loss is to both balance sheet and earnings, and it is the largest relative loss among large companies. In fiscal 2026, to May 31, Oracle had about $130B of debt against about $28B of EBITDA, about 4.6 times (my estimate from its filings [68]). In the quarter to August 31, it spent $28.5B on capital projects against revenue of $19.3B, and it paid for the difference partly with $19.9B of new shares and $11.4B of customer prepayments [5]. In July, S&P cut Oracle to BBB-, one step above junk, and said that it expects a free cash flow deficit of almost $42B in fiscal 2027 [20]. If OpenAI paid for only half of a reported $60B a year [17], Oracle would lose about $30B of yearly revenue. That is 45% of its fiscal 2026 revenue of $67.4B [68]. With fixed leases and depreciation, operating income could fall to near zero, and debt could increase to 9 to 17 times EBITDA (my illustrative estimates). About $117B of its bonds are in the main US corporate index, so a cut to junk would force many funds to sell [69].

Chip makers: the loss is first to earnings power, then through a balance-sheet channel. Nvidia earned a 75.0% gross margin in the July 2026 quarter [70]. If revenue fell 30% and the gross margin fell to about two thirds, operating income would fall by about 43% (my estimate). Nvidia also holds $99B of equity investments plus $25B committed, against $229.0B of equity, and it has guaranteed up to $105B of 20-year leases for an OpenAI campus in Ohio [70]. A 50% write-down of the investments would be about 22% of equity (my estimate). According to press reports, Broadcom pays the shortfall if Anthropic defaults on the $35B TPU loan and the chips do not repay it [59]. TSMC is exposed through its customers. Press estimates put Nvidia at about 19% of its 2025 revenue [71].

Neoclouds: equity takes the loss first, then debt. At June 30, CoreWeave had $35.6B of debt against $5.0B of book equity [23]. That debt was about 5.9 times annualized adjusted EBITDA (my estimate). In July, its credit default swaps implied about a 50% chance of default within five years [72]. Contracts give lenders more protection than chips do. Committed contracts were 98% of CoreWeave's revenue in the June quarter, and one customer was 36% of it [23].

Power companies: the risk is to growth and valuation. Constellation sells the output of the Crane restart to Microsoft for 20 years, and all of Clinton's 1.121 GW to Meta for 20 years from June 2027 [73][74]. It also has 20-year agreements with Amazon and Google from September and October 2026 [51][52]. Demand growth is the variable. In the IEA's base case, data center electricity use about doubles, from about 485 TWh in 2025, about 1.5% of world electricity, to 945 TWh in 2030 [75]. A stall would hit Constellation's target of more than 20% yearly earnings growth to 2029 [76], and its valuation. PG&E is regulated and had 12.7 GW of data centers in its pipeline in June [77]. For PG&E, a stall slows growth without losses.

Brookfield: the main risk is to fee growth and valuation, because most of the money at risk belongs to other investors. In November 2025, Brookfield launched a "$100 billion" AI program with Nvidia and the Kuwait Investment Authority. That number is the value of assets that the program could buy with co-investors and borrowed money. The equity comes from the AI Infrastructure Fund. At launch the fund had $5B of commitments from investors that include Brookfield, Nvidia and KIA, against a $10B target [78]. It reported the same $5B at its first close, in BAM's August results [79]. Brookfield Infrastructure (BIP) expects about $2B of equity in about $10B of compute deals, on take-or-pay contracts with investment-grade counterparties [80]. Losses would fall first on the fund investors (pensions, sovereign funds, insurers and wealth clients) and on lenders. BN does not disclose its own share of the AI fund. Its direct AI equity is probably in the low single-digit billions (my estimate). BN has parent debt of $14.7B at about 4.8% [81] and an A3 rating from Moody's [82].

BN owns about 74% of BAM [81], so its risk is slower growth in BAM's fees. BAM earned $808M of fee-related earnings in the June quarter [79]. It targets almost $11B a year by 2031, which needs fee-paying capital to double to $1.3 trillion, with AI, power and nuclear among the sources of growth [83]. A bust would slow fundraising, decrease fee-paying capital and carried interest, and lower the value that the market gives BAM. A less visible channel is BN's insurer, Brookfield Wealth Solutions. It supplied $45B of BAM's $77B of record fundraising in the June quarter [79] and holds $12.2B of investments in Brookfield entities [81]. Money that the insurer puts into AI assets is BN's own risk.

This view is wrong if BN or its insurer paid for a large share of the AI fund. It is also wrong if the tenants are labs or neoclouds. BAM's chief executive, Connor Teskey, says "we don't take technology risks" in infrastructure [83] and cites 20 to 25-year contracts [84]. That is true when Microsoft or Google is the tenant. The OpenAI partnership in BAM's August results covers Brookfield's use of OpenAI's tools in its own businesses [79].

Other alternative managers and private credit: losses fall first on fund investors and insurers, then on the managers' earnings. Blackstone values its data center platform at $185B [85]. A 20% write-down would be about $37B, mostly for fund investors (my estimate). Blue Owl owns 140 data centers with 15.3 GW [86]. Its funds own 80% of Meta's Hyperion campus, which Meta leases four years at a time [57], and part of the Abilene campus that Oracle leases for OpenAI [87]. This year Blue Owl limited withdrawals from two credit funds three times [88]. Apollo's insurer Athene took a "sizable portion" of the Anthropic loan, according to press reports [59]. Regulators are starting inquiries into insurers' data center financing, according to press reports [89]. If 20% of $700B of AI private credit [18] defaulted with 40% recovery, losses would be about $84B. With a 20% recovery, similar to telecom, losses would be about $112B (my illustrative estimates).

Companies: [MSFT](https://picks.kamadorueda.com/company/MSFT.md), [GOOGL](https://picks.kamadorueda.com/company/GOOGL.md), [AMZN](https://picks.kamadorueda.com/company/AMZN.md), [META](https://picks.kamadorueda.com/company/META.md), [NVDA](https://picks.kamadorueda.com/company/NVDA.md), [AVGO](https://picks.kamadorueda.com/company/AVGO.md), [TSM](https://picks.kamadorueda.com/company/TSM.md), [CEG](https://picks.kamadorueda.com/company/CEG.md), [PCG](https://picks.kamadorueda.com/company/PCG.md), [BN](https://picks.kamadorueda.com/company/BN.md)

### What the telecom bust shows

The telecom bust of 2000 to 2002 is the closest comparison, and its differences are larger than its similarities.

The similarity is a build paid for with debt on demand forecasts. In the five years after the 1996 Telecom Act, telecom companies put more than $500B, mostly borrowed, into fiber networks. In July 2002 the FCC chairman told the Senate that the industry owed about $1 trillion [90]. Moody's counted $135B of rated defaults in 2001, with an average bond recovery of 21 cents on the dollar [91]. In 2002, telecom firms were 52% of defaulted high-yield dollars and recovered only 16.6% on average [92].

The differences are in the asset, the counterparties and the use. Fiber lasted for decades, and the industry used it in the end. A GPU loses most of its value in three to five years, so losses on chips would come faster and be more certain than losses on fiber. In 2001 the borrowers were young carriers with little revenue. Today the main customers are hyperscalers with ratings of AA- to AAA [61], and the weak counterparties are the labs. In 2001 only 2.6% of fiber was in use, according to industry analysts [93]. In 2026 capacity is sold out and rents are increasing [19]. This time the main risk is price. The losses would land mostly on private funds and insurers, which would delay them and concentrate them.

### What to watch

I expect trouble to show in this order. First it shows in neocloud shares and debt. Next it shows in private credit funds and Oracle's bonds. Last, if at all, it shows in the bonds of Microsoft, Alphabet, Amazon and Meta.

Late-October results from Microsoft, Alphabet, Amazon and Meta give order books, server lives, 2027 spending plans and the values of lab stakes. Shorter lives or slower conversion of order books make C more likely.

Anthropic's public prospectus follows a draft that leaked in September [11]. A gross margin near 80% supports A. A margin nearer OpenAI's, about 43% on leaked 2025 figures [12], makes C more likely.

Oracle's credit spread, rating and order book are the next signals. S&P now rates Oracle BBB- [20]. A cut to junk, a renegotiated OpenAI contract or more share sales to pay for capital spending move weight from B to C.

GPU rents and open-model prices show the price of compute. The Silicon Data index, reported in the press [19], shows that rents increased in 2026. If rents fall 20% a year or more, the return model goes below a hyperscaler's cost of capital, and C becomes more likely. Open-model prices on Bedrock, Google and Microsoft [32][33][34] show if the managed layer keeps its markup.

Ramp's AI Index comes out each month [8]. A falling price with a falling frontier share points to C. A falling price with fast-increasing volume points to B. A flat number of paying companies points to D.

Brookfield's AI fund has $5B committed against a $10B target [78][79]. A final close near the target shows that investors still want to fund AI on Brookfield's terms. If the fund stays at $5B, the money is getting harder to raise.

Companies: [MSFT](https://picks.kamadorueda.com/company/MSFT.md), [GOOGL](https://picks.kamadorueda.com/company/GOOGL.md), [AMZN](https://picks.kamadorueda.com/company/AMZN.md), [META](https://picks.kamadorueda.com/company/META.md), [BN](https://picks.kamadorueda.com/company/BN.md)

### My current view

The structure supports my view. The debt and the long leases are with Oracle, the neoclouds and private funds. The four biggest spenders can absorb even severe cases, and their balance sheets stay intact. For them a cheaper model is a cheaper input, and they own the customers and the chips. The clouds' own history with S3, EC2 and managed open models shows that a commodity can still pay if the asset stays busy and lasts. But the trigger has not happened. Lab revenue increased, GPU rents increased, and token volume grew faster than prices fell. I put 70% on outcomes where the buildout earns lower but positive returns and losses stay small. I put 30% on outcomes where some companies take real losses. The results in the next two months will change these weights.

This is a mental model to test. It is information only and gives no advice to buy or sell.

### Sources

1. [CNBC: Microsoft fiscal fourth quarter 2026 results (2026-07-29)](https://www.cnbc.com/2026/07/29/microsoft-msft-q4-earnings-report-2026.html)
2. [Alphabet second quarter 2026 results (8-K)](https://www.sec.gov/Archives/edgar/data/0001652044/000165204426000066/googexhibit991q22026.htm)
3. [Investing.com: Amazon Q2 2026, AWS grows and free cash flow turns negative](https://www.investing.com/news/company-news/amazon-q2-2026-slides-aws-surges-37-free-cash-flow-turns-negative-93CH-4826472)
4. [Eastern Herald: Meta Q2 2026 free cash flow and AI capex (2026-07-30)](https://easternherald.com/2026/07/30/meta-q2-earnings-free-cash-flow-ai-capex/)
5. [Oracle quarterly report (10-Q), quarter to 2026-08-31](https://www.sec.gov/Archives/edgar/data/1341439/000119312526389274/orcl-20260831.htm)
6. [Epoch AI: how persistent is the inference cost burden (2026-02-16)](https://epoch.ai/gradient-updates/how-persistent-is-the-inference-cost-burden)
7. [Epoch AI: the open-closed capability gap (2026-05-29)](https://epoch.ai/data-insights/open-closed-eci-gap)
8. [Ramp AI Index, September 2026](https://ramp.com/data/ai-index-sept-2026)
9. [Apidog: GPT-5.5 API pricing (aggregator, 2026)](https://apidog.com/blog/gpt-5-5-pricing/)
10. [Bloomberg: Anthropic sees adjusted operating profit this quarter, FT says (2026-09-13)](https://www.bloomberg.com/news/articles/2026-09-13/anthropic-sees-adjusted-operating-profit-this-quarter-ft-says)
11. [Fortune: Anthropic's draft prospectus, leaked to Reuters (2026-09-29)](https://fortune.com/2026/09/29/anthropic-ipo-s-1-prospectus-income-statement/)
12. [Fortune: OpenAI financials (leak, 2026-06-16)](https://fortune.com/2026/06/16/openai-financials-leaked-losses-revenue-profit/)
13. [TechCrunch: Anthropic's annualized revenue surges to $65B (2026-08-17)](https://techcrunch.com/2026/08/17/anthropics-annualized-revenue-surges-to-65b/)
14. [Constellation Research: Microsoft Q2, OpenAI is 45% of RPO (2026-01-28)](https://www.constellationr.com/insights/news/microsoft-q2-strong-azure-growth-39-openai-45-rpo)
15. [Michael Parekh, summarizing The Information: Anthropic and OpenAI are half of cloud backlog (May 2026)](https://michaelparekh.substack.com/p/ai-anthropic-and-opemai-half-of-2)
16. [Microsoft annual report (10-K), fiscal 2026](https://www.sec.gov/Archives/edgar/data/0000789019/000119312526323660/msft-20260630.htm)
17. [The Register, citing the WSJ: OpenAI's Oracle contract (2025-09-11)](https://www.theregister.com/2025/09/11/openai_reportedly_on_the_hook/)
18. [Bank of England Financial Policy Committee record, September 2026](https://www.bankofengland.co.uk/financial-policy-committee-record/2026/september-2026)
19. [Wccftech, citing Silicon Data: GPU rental rates continue to rise (2026-08-03)](https://wccftech.com/what-gpu-depreciation-nvidia-a100-h100-and-h200-gpu-rental-rates-continue-to-price-upwards-as-b200s-hover-at-just-shy-of-6-hour/)
20. [heise: S&P downgrades Oracle to BBB- (2026-07-09)](https://www.heise.de/en/news/S-P-downgrades-Oracle-to-BBB-only-one-notch-above-junk-level-11363472.html)
21. [Alphabet quarterly report (10-Q), quarter to 2026-06-30](https://www.sec.gov/Archives/edgar/data/1652044/000165204426000071/goog-20260630.htm)
22. [Amazon quarterly report (10-Q), quarter to 2026-06-30](https://www.sec.gov/Archives/edgar/data/1018724/000101872426000026/amzn-20260630.htm)
23. [CoreWeave quarterly report (10-Q), quarter to 2026-06-30](https://www.sec.gov/Archives/edgar/data/1769628/000176962826000366/crwv-20260630.htm)
24. [Amazon annual report (10-K), 2025](https://www.sec.gov/Archives/edgar/data/1018724/000101872426000004/amzn-20251231.htm)
25. [AWS Price List API: Amazon S3 prices, US East (N. Virginia), published 2026-09-28](https://pricing.us-east-1.amazonaws.com/offers/v1.0/aws/AmazonS3/current/us-east-1/index.json)
26. [AWS Price List API: data transfer prices, US East (N. Virginia), published 2026-09-16](https://pricing.us-east-1.amazonaws.com/offers/v1.0/aws/AWSDataTransfer/current/us-east-1/index.json)
27. [HDD Hunt: hard drive price per TB, August 2026 (retail price tracker)](https://hddhunt.com/blog/hdd-price-per-tb-august-2026/)
28. [AWS: Amazon S3 storage classes (durability and Availability Zones)](https://aws.amazon.com/s3/storage-classes/)
29. [AWS News Blog: Amazon S3 turns 20 (2026-03-16)](https://aws.amazon.com/blogs/aws/aws-weekly-roundup-amazon-s3-turns-20-amazon-route-53-global-resolver-general-availability-and-more-march-16-2026)
30. [Bytebase: Amazon RDS db.r6g.xlarge pricing (aggregator of AWS prices)](https://www.bytebase.com/dbcost/rds/instance/db.r6g.xlarge.md)
31. [DevZero: AWS EC2 r6g.xlarge pricing (aggregator of AWS prices)](https://www.devzero.io/instances/aws/r6g.xlarge)
32. [AWS Price List API: Amazon Bedrock model prices, US East (N. Virginia), published 2026-10-06](https://pricing.us-east-1.amazonaws.com/offers/v1.0/aws/AmazonBedrock/current/us-east-1/index.json)
33. [Google Cloud: generative AI pricing, open and partner models (read 2026-10-07)](https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing)
34. [Azure Retail Prices API: DeepSeek models in Microsoft Foundry (read 2026-10-07)](https://prices.azure.com/api/retail/prices?%24filter=contains(meterName%2C%27DeepSeek%27))
35. [DeepSeek API: models and pricing (read 2026-10-07)](https://api-docs.deepseek.com/quick_start/pricing/)
36. [Dataconomy: Chinese AI models hit 61% share on OpenRouter (2026-02-25)](https://dataconomy.com/2026/02/25/chinese-ai-models-hit-61-market-share-on-openrouter/)
37. [CNBC: Michael Burry on hyperscaler depreciation (2025-11-11)](https://www.cnbc.com/2025/11/11/big-short-investor-michael-burry-accuses-ai-hyperscalers-of-artificially-boosting-earnings.html)
38. [amcompute: GPU depreciation and residual value report 2026 (industry blog)](https://www.amcompute.com/blog/gpu-depreciation-residual-value-report-2026)
39. [SemiAnalysis InferenceX: MiniMax M2.7 cost per token, GB200 vs H100 (July 2026 pricing)](https://inferencex.semianalysis.com/compare-per-dollar/minimax-m27-gb200-vs-h100/8k-1k)
40. [Wing VC: how much does an Nvidia NVL72 cost (August 2026)](https://www.wing.vc/content/how-much-does-an-nvidia-nvl72-cost)
41. [BigGo Finance, summarizing Morgan Stanley on generative AI returns (2026)](https://finance.biggo.com/news/212a6131-57b2-49c8-8ccf-20a27bcac88d)
42. [AWS News Blog: up to 45% price reduction for Nvidia GPU instances (June 2025)](https://aws.amazon.com/blogs/aws/announcing-up-to-45-price-reduction-for-amazon-ec2-nvidia-gpu-accelerated-instances/)
43. [Paul Kedrosky, citing The Information: Oracle's GPU rental margins (2025-10-07)](https://paulkedrosky.com/the-godfather-and-oracles-crap-gpu-rental-returns/)
44. [Fortune: Oracle on AI cloud margins at its analyst day (2025-10-17)](https://www.fortune.com/2025/10/17/oracle-declines-on-concerns-about-fulfilling-ai-cloud-demand)
45. [Beancount.io: CoreWeave second quarter 2026 analysis (2026-09-22)](https://beancount.io/blog/2026/09/22/coreweave-fy2026-q2-earnings-analysis)
46. [Yahoo Finance / 24/7 Wall St.: five hyperscaler stocks most at risk (2026)](https://finance.yahoo.com/technology/ai/articles/5-hyperscaler-stocks-most-risk-153110924.html)
47. [SemiAnalysis: H100 vs GB200 NVL72 total cost of ownership (2025-08-20)](https://newsletter.semianalysis.com/p/h100-vs-gb200-nvl72-training-benchmarks)
48. [Broadcom third quarter fiscal 2026 results (8-K, 2026-09-02)](https://www.sec.gov/Archives/edgar/data/1730168/000173016826000076/avgo-08022026x8kxex99.htm)
49. [Hardware Busters: TSMC's second quarter 2026, two thirds of it AI (press)](https://hwbusters.com/news/tsmc-q2-2026-the-best-quarter-in-its-history-and-two-thirds-of-it-was-ai/)
50. [TechTarget: Amazon's $60B Qualcomm deal expands AI inference hardware (September 2026)](https://www.techtarget.com/it-infrastructure/news/366650224/Amazons-60B-Qualcomm-deal-expands-AI-inference-hardware)
51. [Constellation: 20-year power purchase agreement with Amazon at Calvert Cliffs (2026-09-30)](https://www.constellationenergy.com/news/2026/09/constellation-and-amazon-announce-20-year-power-purchase-agreement-at-calvert-cliffs.html)
52. [Constellation: Google agreement for 890 MW of new nuclear capacity in PJM (2026-10-06)](https://www.constellationenergy.com/news/2026/10/google-and-constellation-announce-landmark-agreement-to-bring-890-mw-of-new-nuclear-capacity-to-pjm-grid.html)
53. [S&P Global second quarter 2026 results (8-K)](https://www.sec.gov/Archives/edgar/data/64040/000006404026000040/spgi2q2026-earningsrelease.htm)
54. [Shacknews: Google processes 3.2 quadrillion tokens a month (May 2026)](https://www.shacknews.com/article/149205/google-3-2-quadrillion-monthly-ai-tokens)
55. [Alibaba Group June quarter 2026 results (6-K)](https://www.sec.gov/Archives/edgar/data/1577552/000110465926099220/tm2623667d1_ex99-1.htm)
56. [ServiceNow second quarter 2026 results (8-K, 2026-07-22)](https://www.sec.gov/Archives/edgar/data/1373715/000137371526000072/erq2fy26.htm)
57. [IFRE: the Meta and Blue Owl Hyperion deal broken down](https://www.ifre.com/people-and-markets/2363199/the-metablue-owl-deal-broken-down-off-balance-sheet-gymnastics-24-years-after-enron)
58. [SEC XBRL company facts: Nvidia (quarterly revenue history)](https://data.sec.gov/api/xbrl/companyfacts/CIK0001045810.json)
59. [Yahoo Finance, citing Bloomberg: Apollo and Blackstone lend $35B against chips for Anthropic (2026-06)](https://finance.yahoo.com/sectors/technology/articles/apollo-blackstone-lend-35b-against-220610560.html)
60. [TSMC consolidated financial report, second quarter 2026 (6-K)](https://www.sec.gov/Archives/edgar/data/1046179/000104617926000541/a2026q2consolidatedreport-.htm)
61. [FI-Desk: hyperscaler spending rings credit alarm bells (2026-05-27)](https://www.fi-desk.com/ratings-analysis-hyperscaler-spending-rings-credit-alarm-bells/)
62. [Yahoo Finance, citing Reuters: Anthropic's new AI tools deepen selloff in data and software stocks (2026-02)](https://finance.yahoo.com/news/anthropics-ai-tools-deepen-selloff-054551832.html)
63. [US Census Bureau: AI use by businesses (May 2026)](https://www.census.gov/library/stories/2026/05/ai-use-businesses.html)
64. [Meta quarterly report (10-Q), quarter to 2026-06-30](https://www.sec.gov/Archives/edgar/data/1326801/000162828026050705/meta-20260630.htm)
65. [Constellation Research: Microsoft owns 27% of OpenAI, worth $135B (2025-10)](https://www.constellationr.com/insights/news/microsoft-owns-27-openai-worth-135-billion)
66. [FactSet: hyperscalers tap external financing as AI capex outruns cash flow (2026-07-23)](https://insight.factset.com/hyperscalers-tap-external-financing-as-ai-capex-outruns-cash-flow)
67. [Yahoo Finance, citing the WSJ: hyperscaler debt binge pushes yields (2026)](https://finance.yahoo.com/markets/stocks/articles/hyperscaler-debt-binge-pushes-yields-134825104.html)
68. [SEC XBRL company facts: Oracle (fiscal 2026 10-K data)](https://data.sec.gov/api/xbrl/companyfacts/CIK0001341439.json)
69. [The Next Web: Oracle's S&P downgrade and junk risk (2026-07)](https://thenextweb.com/news/oracle-sp-downgrade-ai-spending-junk-risk)
70. [Nvidia quarterly report (10-Q), quarter to 2026-07-26](https://www.sec.gov/Archives/edgar/data/1045810/000104581026000075/nvda-20260726.htm)
71. [MacRumors: Nvidia replaces Apple as TSMC's biggest customer (2026-01-28)](https://macrumors.com/2026/01/28/nvidia-replaces-apple-as-biggest-tsmc-customer/)
72. [TechTimes: CoreWeave CDS implies 50% default odds (2026-07-30)](https://www.techtimes.com/articles/322222/20260730/coreweave-cds-hits-50-default-odds-bond-market-calls-time-gpu-debt-spiral.htm)
73. [Constellation Energy quarterly report (10-Q), quarter to 2026-06-30](https://www.sec.gov/Archives/edgar/data/1868275/000186827526000104/ceg-20260630.htm)
74. [DCD: Meta signs 20-year PPA with Constellation for Clinton's output](https://www.datacenterdynamics.com/en/news/meta-signs-20-year-ppa-with-constellation-for-entire-output-of-illinois-nuclear-power-plant/)
75. [Our World in Data, citing the IEA: how much energy do data centers and AI use (2026-07-20)](https://ourworldindata.org/how-much-energy-do-data-centers-and-artificial-intelligence-use)
76. [Nasdaq / Zacks: Constellation second quarter 2026 results](https://www.nasdaq.com/articles/ceg-q2-earnings-top-estimates-calpine-contribution-view-raised)
77. [Insurance Journal: PG&E data center pipeline (2026-07-24)](https://amp.insurancejournal.com/news/west/2026/07/24/878798.htm)
78. [BNN Bloomberg, citing Reuters: Brookfield launches $100B AI infrastructure program with Nvidia (2025-11-19)](https://www.bnnbloomberg.ca/business/2025/11/19/brookfield-launches-us100-billion-ai-infrastructure-program-with-nvidia/)
79. [Brookfield Asset Management second quarter 2026 results (8-K)](https://www.sec.gov/Archives/edgar/data/0001937926/000117184326005272/exh_991.htm)
80. [Brookfield Infrastructure Investor Day transcript (September 2026)](https://stockanalysis.com/stocks/bip/transcripts/721844-investor-day-2026/)
81. [Brookfield Corporation second quarter 2026 report (6-K)](https://www.sec.gov/Archives/edgar/data/1001085/000100108526000021/bn-20260630.htm)
82. [Cbonds: Moody's affirms Brookfield Corporation at A3 (2026-04-28)](https://cbonds.com/news/3894095/)
83. [Brookfield Asset Management Investor Day transcript (2026-09-17)](https://stockanalysis.com/quote/tsx/BAM/transcripts/752038-investor-day-2026/)
84. [Semafor: Brookfield is bullish on AI investing (2026-10-04)](https://semafor.com/article/10/04/2026/enough-to-go-around-brookfields-bullish-on-ai-investing-even-if-it-has-to-share-it-with-rivals)
85. [Yahoo Finance: Blackstone leans into AI data centers (second quarter 2026)](https://finance.yahoo.com/real-estate/articles/blackstone-leans-ai-data-centers-062747180.html)
86. [Investing.com: Blue Owl second quarter 2026 earnings call transcript](https://www.investing.com/news/transcripts/earnings-call-transcript-blue-owl-capital-tops-revenue-estimates-in-q2-2026-93CH-4825378)
87. [Crusoe: joint venture with Blue Owl and Primary Digital Infrastructure (Abilene)](https://www.crusoe.ai/resources/newsroom/crusoe-blue-owl-capital-and-primary-digital-infrastructure-enter-joint-venture)
88. [Bloomberg: Blue Owl again caps two BDCs (2026-10-02)](https://www.bloomberg.com/news/articles/2026-10-02/blue-owl-again-caps-two-bdcs-after-39-17-withdrawal-requests)
89. [Startup Fortune: US insurance regulators open inquiries into AI data center financing (2026)](https://startupfortune.com/us-insurance-regulators-are-opening-formal-inquiries-into-the-credit-risks-piling-up-inside-ai-data-centre-financing/)
90. [Wikipedia: Telecoms crash](https://en.wikipedia.org/wiki/Telecoms_crash)
91. [Risk.net: Moody's on 2001 defaults and recoveries](https://risk.net/node/1504555)
92. [Altman and Bana, NYU Stern: defaults and returns on high-yield bonds, 2002 in review (February 2003)](https://archive.nyu.edu/bitstream/2451/27147/2/wpa03010.pdf)
93. [Emerging Moats: AI and the fiber overbuild](https://www.emergingmoats.com/p/ai-and-fiber-overbuilds)