Weeds and Flowers
In late 1971, a saver willing to hold first-rate corporate bonds could collect 7 1/4 percent for the trouble. Everything Benjamin Graham says in Chapter 6 of The Intelligent Investor hangs from that number.
The chapter concerns the line between high-grade bonds and second-grade ones, and the temptation to cross it for extra income. The practice had a flattering name, the businessman’s investment, and Graham retired the name as a fallacy in one sentence: “It is bad business to accept an acknowledged possibility of a loss of principal in exchange for a mere 1 or 2% of additional yearly income.” Applied to his own market: “Since in late-1971 it is possible to find first-rate corporate bonds to yield 7¼%… it would not make much sense to buy second-grade issues merely for the higher return they offer.”
The test does not ask whether the issuer will default; that would require an opinion about the future. It asks whether the extra income is adequate pay for a possibility the buyer has already acknowledged. Rates have moved many times in fifty-five years; the test has not, because the test is a ratio: extra pay, over acknowledged risk, against the first-rate alternative. It can be run on any dated tape without a forecast attached. This Note runs it on July 2026.
The reading
The increment first. On the ICE BofA series, the extra yearly income for descending from Treasuries to the speculative rung stood at 2.69% on July 13, 2026; Bloomberg’s version of the measure closed November 2025 at 2.7%, against a twenty-year average of 4.9%. The pay for the descent is a little more than half its own two-decade norm. One rung up, investment-grade spreads touched 74 basis points in September 2025, the narrowest since roughly 1998, and have held a band of 71 to 80 through the first half of 2026; the reading on July 13 was 78.
Then the base rate, because a conclusion published without one is a pose. Schwab ran the daily history from 2005 through 2025. Measured from days when the spread stood at or below 3%, high yield beat Treasuries over the following twelve months 39% of the time. From days at 5% or above, 83%.
So the test reads out, and this is the conclusion of the Note, the only one it will offer: at these prices, the buyer of the risk-taking rung is not being paid for the risk assumed.
What that sentence does not say matters as much. It does not say spreads will widen; a tight market can stay tight for years, and nothing here predicts when or whether that changes. It does not say the buyer must lose; even from spreads this low, the risk-taker has finished ahead about two years in five. The test grades the price of the bet, not the next draw of the cards. Graham built it that way on purpose, which is why it can be run without pretending to know anything.
This is the first Note on these pages to conclude, so the terms are stated once. A concluding Note carries its date and its base rate on its face; the record of how prices actually move is displayed beside it, without self-grading; errors are corrected at the foot and never silently. For this Note: prices of July 13, 2026, published July 15, 2026, base rate 39%.
The supply behind the price
Investment-grade issuance tied to artificial intelligence reached $218 billion by July 8, 2026, against $80.5 billion in all of 2025, on PitchBook LCD’s US count; Bank of America’s broader measure puts the pace at about 1.6 times last year’s rather than 2.7. With $31.9 billion of high-yield paper, the total nears a quarter trillion dollars.
Meta serves as the specimen because Meta filed the paperwork, and the record holds it. The company entered October 2025 carrying about $29 billion of corporate debt and raised about $57 billion more across October and November. A $27.3 billion joint venture with Blue Owl, named Beignet, rated A+ by S&P and priced at 6.581%, funds the two-gigawatt Hyperion campus; that borrowing sits off Meta’s balance sheet, inside the venture. A $30 billion offering in Meta’s own name came tied for the fifth-largest investment-grade deal ever sold, on a record $125 billion order book. The prospectus states the geometry without ceremony: long-term debt of $28,834 million before the offering, $58,834 million after. The $27.3 billion appears in neither figure.
Behind the deals, the consensus for 2026 hyperscaler capital spending was revised during 2025 from $314 billion to $458 billion to $518 billion on the Janus Henderson and Goldman tallies; CreditSights and MUFG now put it near $600 billion, with about $4 trillion projected through 2030. The portion that operating cash does not carry is the portion the bond market is being asked to carry.
The marginal price
For all that supply, the index has barely moved: investment-grade spreads sit about 10 basis points off their September tights. A market resting near a three-decade low has looked at everything above and declined to be alarmed.
The marginal price is another matter. In mid-September 2025, ten-year BBB technology bonds traded 12 basis points tighter than non-financial corporates; seven weeks of supply erased the premium. Meta’s 6.30% bonds of 2056 reached the widest level of their short trading history in early July, 145 basis points over Treasuries, 13 wide of their April pricing; the 2036s have broken a full point over Treasuries against an investment-grade average near 0.72.
Both readings are true at once. The becalmed index says the market disagrees with any shock thesis; the repricing of the strongest issuers’ newest paper says one dial has moved. A dated ledger can hold both without blinking.
The quieter rung
The louder half of the wave trades under famous names. The quieter half is private.
The Financial Stability Board’s report of May 6, 2026 judged the market reaction to the First Brands and Tricolor failures “contained and short-lived.” The same report catalogs underwriting weakness that “spanned across corporate credit markets,” banks lending at three levels, insurers writing credit insurance on the defaulted firms, and creditors across eleven jurisdictions. The adjective attaches to the reaction; the failures themselves the Board attributed to fraud, opacity, hidden leverage, and off-balance-sheet financing rather than to a turn in the cycle.
Measurement is the livelier question. Outright defaults in private credit run near 1%; counted with selective defaults, the rescue restructurings that spare the word, the figure is closer to 5%, comparable to public high yield. Which column a reader trusts settles much of the argument in advance. About 12% of loans now carry payment-in-kind features, a rising share; a loan that pays its interest in more loan defers the cash test to maturity, so the outright measure reads low.
Software is where AI the borrower meets AI the competitor. Lending to software firms grew from about $8 billion in 2015 to more than $500 billion by end-2025, 19% of all direct lending, on the BIS’s March 2026 count. On S&P’s tally, about $386 billion of software debt comes due in 2028 and 2029, after $80 billion this year. Software equities fell about 30% between October 2025 and February 2026, a decline the BIS attributes to AI-disruption fears, and the shares of software-heavy BDCs, the listed lenders, trailed their peers. Lincoln International’s technology co-head argues the buyouts remain “fundamentally good credit,” and Lincoln’s first-quarter marks agree: enterprise values down 8.8%, loan values down 1.9% of par. Equity is the cushion that exists to be dented. PitchBook adds a dry detail: private-credit funds have been declining new software borrowers. The strain, on today’s evidence, is a matter of appetite. The marks show little else.
The bond letters
A reasonable reader might object that everything above is a market call in a longer coat. The objection gets a straight answer.
Buffett, in the 1988 Berkshire letter: “We do not have, never have had, and never will have an opinion about where the stock market, interest rates, or business activity will be a year from now.” This Note is written under the same restriction, and nothing in it requires an exception. Whether 2.69% pays for an acknowledged possibility of loss is a question about a price on a date, the same shape of question as whether a quoted business is cheap at its multiple.
That is how Buffett read bonds himself. The 1984 letter describes treating bond investment as “an unusual sort of ‘business’ with special advantages and disadvantages,” and wishes the habit on everyone: “many staggering errors by investors could have been avoided if they had viewed bond investment with a businessman’s perspective.” Graham condemned the businessman’s investment because the businessman in it had stopped doing arithmetic. Buffett readmitted him on the condition that he do nothing else.
The test cut both ways in his hands. In the 1990 letter, with the junk market in ruins, Berkshire held RJR Nabisco paper because the yield, as well as the potential gain, “more than compensates for the risk we incur (though that is far from nil),” while the same letter passed sentence on the era: “Mountains of junk bonds were sold by those who didn’t care to those who didn’t think,” and “there was no shortage of either.” That sentence described 1990’s merchandise and its buyers; the record book for Meta’s investment-grade offering is a different market and a different buyer. In the 2002 letter the test said yes at scale, commitments “sextupled, reaching $8.3 billion by yearend.” The 2003 letter recorded the stop: “In 2002, junk bonds became very cheap, and we purchased about $8 billion of these. The pendulum swung quickly though, and this sector now looks decidedly unattractive to us. Yesterday’s weeds are today being priced as flowers.” A market priced as flowers can stay priced that way for years; the sentence carried no forecast in 2003 and carries none here.
The absorption case
A reader who expects this wave to be digested without incident holds serious evidence, and it deserves assembly at full strength.
On this reading, hyperscaler operating cash carries most of the capital spending, and the issuance is strength terming itself out: an intelligent treasurer accepting long money because a deep market offered it. A $125 billion book for a $30 billion deal is a market bidding for more than four times the paper on offer. The erased technology premium is a functioning market charging for supply, promptly and in the right place; a market that reprices the marginal bond while holding the index steady is discriminating, which is health. The FSB traced the season’s two famous failures to fraud, and fraud is idiosyncratic; the broader default column, on this reading, counts mercy as failure, since a lender who amends terms has exercised exactly the flexibility private credit sells. Lincoln’s marks put the software damage in equity, where risk capital is paid to take it. On this ledger the wave is enormous cash generators borrowing at the tightest spreads in three decades because they can, which is what strength looks like when it visits the bond market. A reader holding this view is trusting an index that has watched everything in this Note and moved 10 basis points, and that index has seen more cycles than any of its critics.
None of it quarrels with the conclusion, which grades the buyer’s side of the trade. The same coupon can be intelligent to sell and unrewarding to own; that is roughly what a three-decade tight means. The absorption case can be entirely right and the test still reads the same.
What each side requires
Stated as requirements on the future tape, with no forecast of whether either arrives.
The shock thesis requires the technology-bucket repricing to broaden into index-level widening; the software maturity wall to meet the refinancing reluctance PitchBook observes and produce realized losses; the measurement gap to resolve toward the 5% column; and demand fatigue to show where the primary market prices admission, in concessions and thinning books, beyond any single issuer’s curve.
The absorption thesis requires operating cash to carry most of the capital spending, so issuance remains a term-out of strength; order-book depth to persist if supply approaches the projected $4 trillion; the erased technology premium to remain a market correctly charging for supply; and software stress to stay where Lincoln’s marks place it, in the equity below the loans.
The panel
Each dial carries a dated reading and the zone this publication will treat as changing the reading. The panel predicts nothing, in either direction: dials that hold their readings as supply arrives are the absorption case confirming itself, and where a dial can strengthen, its zone says so. It is written down in advance so that nobody, this publication included, has to rely on memory later.
| Dial | Reading, dated | Zone that changes the reading |
|---|---|---|
| IG index OAS | ~71-80bp band Jan-Jul 2026; 78bp on Jul 13, 2026 | Sustained break through ~130-150bp, or fast widening from the tights |
| HY index OAS | 2.69% on Jul 13, 2026 vs a 4.9% twenty-year average | A sustained move through ~4-5%, the neighborhood of its own average |
| BBB tech vs non-financial corporates | 12bp premium, mid-Sept 2025, erased to flat in seven weeks | Materially positive and widening; indigestion localizing |
| Meta’s curve | 2036s past 1.0pt over Treasuries vs a ~0.72 IG average; 2056s at T+145, 13bp wide of April pricing, early Jul 2026 | New issues pricing wide of secondaries; curves steepening |
| AI-IG issuance | $218B by Jul 8, 2026 vs $80.5B in all of 2025; capex projections ~$600B for 2026, ~$4T through 2030 (CreditSights, MUFG) | Funding-gap issuance replacing term-outs of strength |
| Order books | $125B record on Meta’s $30B deal, Oct-Nov 2025 | Books thinning; deals downsized or pulled |
| PIK share of private credit | ~12% of loans, about half toggles, rising since 2022 | Acceleration, with PIK use and delinquency moving together; a falling share reads the other way |
| Default measures | ~1% outright vs ~5% including selective defaults | The broad measure rising while the narrow one stays flat; the two converging downward reads the other way |
| BDC prices vs marks | Stocks down 10% on average since October 2025; software-heavy names 5pp behind low-exposure peers | NAV discounts deepening while marks hold near par |
| Software maturity wall | $80B due 2026; ~$386B in 2028-29, ~$162B of it in private credit | Refinancings failing, or repricing punitively, as maturities arrive |
The ticket price
Graham’s world paid 7 1/4 percent on first-rate paper, and he still judged an extra point or two inadequate pay for an acknowledged possibility. The market of July 2026 pays 2.69 points over Treasuries for the whole speculative rung, at investment-grade spreads inside a band last seen in the 1990s, with $218 billion of AI-related paper sold by early July against $80.5 billion in all of last year. The test Graham wrote and Buffett kept does not know what happens next, and neither does this publication. It knows what the ticket costs. At these prices, dated and displayed above, the buyer of the risk is not being paid for it. The rest belongs to the record, which will accumulate beside this page whether it flatters the conclusion or embarrasses it.
This note's conclusion is entered, verbatim and permanently, in the Ledger.