The question
Amazon sold a 30-year bond in June 2020. It is worth about 58 cents on the dollar today. The instinct is to reach for a story about artificial intelligence, capital spending, or the market souring on technology debt.
That instinct is wrong, and it is possible to prove it is wrong rather than argue about it.
Where you buy decides almost everything
On 9 March 2020 the 30-year Treasury closed at 0.99%, the lowest reading in a series that starts in 1977. The six 2020 bonds studied here all priced between 26 March and 3 August 2020, when it ran between roughly 1.2% and 1.6%.
As of Thursday 24 September it was 5.47%. The median of every single day since 1977 is 5.42%. Just under half of all recorded days were higher.
Buying long duration at 0.99% was the bond-market equivalent of buying at an all-time high. Price and yield move inversely, so a record-low yield is a record-high price for a fixed stream of dollars. Almost nobody described it that way at the time, because a falling yield feels like safety rather than like a rising price.
One honest qualification on that metaphor, because it matters. Nobody bought these bonds at an all-time high. They were new issues sold at 98.57 to 99.71, which is par. Their own highs came later, and that is the trap: Amazon's 2050s reached 109.05 in July 2020 and Apple's 107.48 in November 2020 before ending forty and thirty-seven points below where they were sold. The duration trap pays you first. That is why it is so hard to see from inside.
What that did to seven bonds
The table above is every megacap long bond we could find from that vintage, plus Meta as a control. Two things are true of all of them.
The Treasury move explains more than 100% of the price fall in every single case. The range is 104% to 160%. Not most of it. All of it, and then some.
The credit spread tightened on all seven, without exception. Minus 21, 25, 28, 39, 45, 73 and 169 basis points, measured at the 30 June 2026 marks. Every one traded at a narrower spread over Treasuries than the day it was sold. The market liked all six of these companies more than it did in 2020.
That is what rules out the credit story. It is not that credit fear was small. It is that credit moved the other way.
Three results worth sitting with
The worst bond belongs to the company nobody worries about
$GOOGL's 2.250% of 2060 is at 50.055, almost exactly half its 99.007 issue price. Eleven coupons later the total return is -36.1%.
It is the worst for an entirely mechanical reason: the smallest coupon over the longest term, priced on 3 August 2020 when the 30-year was 1.23%, the lowest issue-date yield of anything here. Nobody has a thesis about Alphabet's balance sheet. It did not need one.
It is also the only bond in the set that never traded above its issue price. No honeymoon at all, which makes it the one bond for which the all-time-high metaphor is close to literal.
Oracle's 2020 bond beat Amazon's
This is the decisive test, because Oracle is the issuer the market is actually bearish about right now.
$ORCL's 3.850% of 2060 is at 58.715. $AMZN's 2.500% of 2050 is at 57.938. On price they are indistinguishable, -41.1% against -41.5%. On total return Oracle wins clearly, -16.9% against -26.1%, because a 3.850% coupon has paid 24.07 points of income against 2.500%'s 15.21.
And Oracle's spread is 39 basis points tighter than where it priced in the March 2020 panic.
The like-for-like check gives the same answer: Oracle's 3.600% of 2050, same maturity year as Amazon's, is at 60.740 for a -16.5% total return.
The Oracle fear of 2026 does not appear anywhere in Oracle's 2020 paper.
Meta's absence is the finding
Meta carried zero long-term debt at every year-end from 2013 through 2021. It did not sell a bond until August 2022, when it did a $10bn four-tranche debut. We checked this three ways: no 424B of any kind filed between February 2015 and November 2022; the XBRL long-term debt tag reads $0 at 31 December 2021; and the Q3 2022 10-Q describes the debut directly.
Its longest bond, the 4.650% of 2062, entered when the 30-year was already 2.97% rather than 1.2%. It has absorbed only +194bp of Treasury move against Alphabet's +368bp, and it is down 5% against $GOOGL's 36%.
Same kind of company. Same kind of bond. Different entry point.
And the structural point: Meta has since gone from $0 to $58.7bn of long-term debt, making it one of the most levered AI spenders on the list. A phenomenon that excludes the most AI-levered issuer entirely is not an AI phenomenon.
Nvidia proves it from the other direction
$NVDA's 3.700% of 2060 has the best total return of the 2020 vintage at -4.3%, and it is tempting to read that as the market liking Nvidia. It is not.
Nvidia took the same +349bp of Treasury damage as everyone else, a duration cost within 0.15 points of Oracle's and Apple's. It escaped because it priced in April 2020 at a 229bp covid-panic spread that has since collapsed to 61bp, handing back 16.49 points. Its price still fell 27.5%.
So Nvidia is the one row where credit did real work, in the helpful direction. The spread rescued it, and that spread was an artefact of when it sold rather than of what it is.
The qualification, which is worth making rather than hiding
None of this says the market has no credit concern about Oracle in 2026. It does, and it is measurable.
Oracle's own 2026-issued 6.700% of 2056 widened roughly 83 basis points of issuer-specific spread between February and September 2026, while the broad investment-grade index did not widen at all over the same stretch.
So the correct formulation is narrower and more useful than either side of the argument usually allows: the 2020 vintage is a pure rates story, and the 2026 new-issue market is where the credit story lives. Those are different bonds in different windows, and conflating them is the error in both directions.
And the last time this mattered
The assumption that corporate bonds are the dangerous thing in a crisis does not survive the 2008 record either. Total returns with income reinvested: investment-grade corporates were +2.4% in 2008 while the S&P 500 lost 36.8%, and over 2008 and 2009 together corporates made +11.1% against the index's -20.1%.
Treasuries won the panic with +34.0% and then gave almost all of it back, -21.8% in 2009, finishing the two years at +4.7%, less than half the corporates' return.
Not painless: corporates still fell 21.5% peak to trough. But on price alone they were -3.0% in 2008, so the coupon is the only reason the sign flipped, which is the same lesson this whole page keeps arriving at from different directions.
The honest caveats
The marks are 30 June 2026 and the market has moved since. The 30-year went from 4.91% then to 5.47% on 24 September, another 56 basis points. We have not published modelled September prices, because a model is not a mark. Note only that every conclusion here gets stronger with more Treasury move, since none of the damage is credit.
Meta's issue price is derived, not read. The August 2022 debut was a 144A placement with no SEC-filed price to public. The ~98.95 is backed out from the effective interest rate in Meta's own filing, and since that rate includes issuance costs, Meta's true loss is probably slightly worse than stated. Treat it as plus or minus a point, which is also why this page gives both numbers rather than a ratio between them.
A stock analogy imports one intuition that does not hold. A share bought at an all-time high can revisit that high. A 2.250% of 2060 cannot return to its 2020 valuation unless the 30-year returns to 1.2%, and it pulls toward par at maturity regardless. This is a mark-to-market on a locked-in income stream, not a drawdown from a peak.
One statistic we deliberately did not use. Across rolling five-year windows since 1977 the 30-year yield fell 76% of the time with a median fall of 89 basis points. That looks like a compelling base rate and it is not one: it is dominated by the single forty-year disinflation from 1981's 15.21% to 2020's 0.99%, which cannot repeat from 5.47%. Decade medians make the point plainly, running 10.61, 7.88, 5.41, 3.61 and 3.09 from the 1977-86 decade to the present one. This is not a series that rests at a fair value. It appears here only so that nobody mistakes its absence for an oversight.
Educational only. Not advice, and not a forecast.