Interest costs rose 10%. The median firm paid the same

In August we published a panel of 191 companies showing that two years of Fed cuts had not reached corporate interest expense. That article ended with a condition rather than a forecast: the aggregate would turn when the largest borrowers joined the 43% of companies already paying less.

Six months of 2026 are now filed. The aggregate did not turn. It accelerated. And the reason is not the one that article was watching for.

What the comparison measures

The comparison measures reported interest expense across two half-year periods. Fixed-rate debt, new borrowing and refinancing can all change the bill. It does not isolate the effect of a policy-rate change.

The panel

1,519 listed non-financial companies that reported interest expense in all four of the quarters being compared — the first and second quarters of 2025 and of 2026 — using the same XBRL tag in every quarter.

Two constraints are doing work here and both were learned the hard way in the earlier article. Filers migrated from InterestExpense to InterestExpenseNonoperating during this period, so a screen reading one tag loses most of the panel; requiring the same tag in all four quarters means we are subtracting like from like rather than one definition from another. And financial companies are excluded because their interest expense is the cost of funding a lending book — it rises with interest income, and including it measures balance-sheet size rather than borrowing cost.

What happened

First half 2025      $96,351M
First half 2026     $106,064M
──────────
Change                 +10.1%   (+$9,713M)

Ten percent, in a half-year when the average policy rate was 69 basis points lower than the year before. That is the headline, and taken alone it says corporate borrowing costs are still climbing.

Taken alone it is also misleading.

The median company did not experience this

The middle company in this panel saw its interest bill rise 0.4% — flat, within rounding of unchanged. And 724 of the 1,519, or 47.7%, paid less than a year earlier. Nearly half the panel is going the other way from the total.

That combination only happens one way: the increase is concentrated.

Largest single increase (Alphabet)     $1,516M    16% of the rise
Ten largest increases                  $5,437M    56%
Twenty largest increases               $7,276M    75%
────────
All 1,519 companies                    $9,713M   100%

Twenty companies out of 1,519 produced three quarters of the increase.

Who they are

Verizon and AT&T are the panel’s largest interest bills in absolute terms and always have been; their increases are ordinary in percentage terms. The two at the top of the list are not ordinary.

Alphabet’s second quarter alone is 4.3 times its entire interest expense of the prior-year half. The quarterly path — $34M, $261M, $533M, $1,278M across the four quarters — is not a company refinancing at a worse rate. It is a company that was barely a borrower becoming one, fast.

It kept going after this period closed. On 6 August Alphabet priced $25 billion of notes in nine tranches, $10 billion of it maturing after 2045 — which adds roughly $1.33 billion a year of coupon to the line above.

What the money is for is not in dispute and not hidden: both companies are funding capital expenditure at a scale their operating cash flow no longer covers. We have written about the consequence from the other side — Amazon’s free cash flow fell 76.6% in 2025 as capex went from $83.0B to $131.8B, and Oracle went to −$23.7B of free cash flow on the same trade. The interest line is where that borrowing arrives on the income statement, a few quarters later.

Debt is not the only way this build-out is being financed, and the other ways do not show up on this panel at all. NVIDIA has guaranteed up to $105 billion of data-centre leases for a tenant that is not itself — an obligation that creates no interest expense and appears in no aggregate like the one above.

The honest version of the concentration claim

It would be neat to say the entire rise is four data-centre borrowers. It is not.

First half 2025       $86,226M
First half 2026       $90,504M
──────────
Change                   +5.0%

Strip out the ten biggest increases and the remaining 1,509 companies still show a 5.0% rise, while the median company shows 0.4%. Both facts are true, and together they say something more specific than either alone: the increase is concentrated in large borrowers generally, not only in the handful at the very top. Size, not sector, is what sorts this panel.

The panel also keeps changing under you. AMD sits in it at $74 million for the half, up 27.6% — and eleven days after this period closed it sold $4.75 billion of notes at a weighted 5.053%, which on the coupons alone roughly triples that line.

The turn the earlier article was watching for has happened at the typical company. It has not happened in the total, and it will not while the largest borrowers are adding debt faster than the rest are refinancing out of it.

What would make this wrong

• This is not the same panel as our August article. That one was 191 companies across five annual periods; this is 1,519 across four quarters. The two aggregates are not comparable to each other, and neither is a measure of “US corporate debt” as a whole.
• Four quarters is a short window. Half-year comparisons remove seasonality but not one-off items — debt issued or retired mid-quarter lands unevenly, and a single large refinancing can move a company’s line without anything changing about its cost of borrowing.
• 38 companies reported a negative interest expense in at least one quarter, usually where a gain on debt extinguishment was netted into the line. Kraft Heinz reported −$31M for the second quarter of 2026, and that figure is in its 10-Q rather than an artefact of the API. Excluding all 38 moves the aggregate change from 10.1% to 10.4%, so the finding does not depend on them.
• A rising interest bill is not evidence of distress, and this article does not say it is. A company borrowing to build has the same rising line as one refinancing badly. Distinguishing them requires reading the filing.
• Requiring one consistent tag drops companies. Filers that switched tags mid-period are absent from the panel entirely. That is a deliberate trade: a smaller panel measured consistently beats a larger one that subtracts two different definitions.
• Interest expense is not interest paid. Capitalised interest on assets under construction is excluded from this line, which matters most for exactly the companies building the most.

What the median hides

A median is a statement about the middle, and the middle here barely moved. The tails are where the money is, and they are much further apart than a panel of this kind can show.

EchoStar redeemed $3.686 billion of 11¾% secured notes in July 2026, at the closing of a spectrum sale to AT&T — funded by a wire that arrived four weeks after the company reported $440 million of cash. An 11¾% coupon is roughly double what the investment-grade issuers in this panel pay, and the company carrying it was retiring debt out of a shrinking cash balance until an asset sale closed.

That is the same year, the same rate environment, and a different financial universe. Nothing in a median tells you which of those two a company is living in.

Check it yourself

Use the concept endpoints below to inspect the reported interest expense for the periods specified in the panel. Match the form, start date, end date, and units before comparing periods. The panel statistics describe the stated sample; the linked live endpoints do not preserve its historical population and are not a complete reproduction package.

The Alphabet and Amazon quarterly figures can be inspected through the company concept links. Match each quarterly observation to the filing and period before comparing it.

That is what the bill did. For what companies say it would do if rates moved the other way, their own filings answer the question directly — we collected 93 of those disclosures from the June quarter, where the median company puts a 100 basis point rise at $2.9 million a year.

If a figure here does not match a filing, tell us and it will be corrected on the article and on the corrections log, with the date.