Signals Served #029 — The Cost of Capital Moved Somewhere Else

Issue #029 of Signals Served — The Cost of Capital Moved Somewhere Else. Week of September 15–21, 2026. 13 signals with a Founder Signal for each. The repricing of long-duration capital is visible in public markets and unobserved near a seed round; the only early-stage measurement anyone published was about whether rounds finish, not what they cost.

Primary source for the September 16 decision: a unanimous 25bp rise to 3.75–4.00%, the first increase since July 2023 — a gap of thirty-seven months.

The daily series behind the 5.01% print on September 16, its highest daily level since July 2007, and the 4.94% settle on the 17th. The October 2023 peak was 4.98% and never crossed five.

Used to measure the steepening rather than the level: policy down 50bp over twelve months against the long end up 88, a 138 basis point widening between the fed funds upper bound and the 10-year.

The week's only measurement of founders. Of 2025 seed rounds a founder started, 26% ever wired; rounds a third-party deal lead brought, 65%. Both converged on the same $25M median pre-money. AngelList is explicit that the lead is a marker of investor support, not a demonstrated cause of it.

The one company this week with no say over which number it published: a $1.02B net loss against $140.6M of revenue, alongside a company-defined ~$103B order book and $56.4B of booked obligations.

Peter Oppenheimer's mechanism: AI capital expenditure and government borrowing drawing on one pool of savings. He names the mechanism and pointedly declines to size AI's share of the move — a discipline worth copying.

Secondary coverage of the Competition for Capital note. Read alongside the primary: it is a sell-side argument from a bank that underwrites the issuance it is counting.

Note the word the company chose: an initial closing. Two weeks after the press reported a smaller number in progress. That is what an upsizing round looks like, and it means $30.9B is not a settled post-money either.

The fact that argues against this issue, included on purpose. Also: a former Infosys chief executive raising from the chief executives of Intel and Micron — access at the top of the distribution, not a reading of the median.

The cost of capital as a verb rather than a chart. A company that wanted $10B for an equity position went to the bond market in the same week the long end hit a nineteen-year high. The coupon prints September 24.

Torsten Slok argues the spread widening against banks is about the capex, not the plumbing. Check the basket before repeating the number — swap Oracle out and the signal looks materially different — and note Apollo is a private-credit manager whose book benefits from caution about public AI credit.

Secondary coverage of the Apollo note. At least one outlet reported the CDS basket wrong this week — use the argument, and attribute the number to whoever actually said it.

One of three OpenAI numbers this week, and the one that is a negotiation rather than a transaction: a mark investors floated. Nobody has paid it.

The second of the three numbers: a projection the company makes about itself. Three numbers, three different kinds of evidence, and only one of them involved anybody paying.

Reported in the register of a completed round. It is a negotiation: talks, no product, one month old. Note the arithmetic — a $3.0B pre-money plus $700M is a $3.7B post, so the two figures quoted are one number, not two.

Second source on the same negotiation. The transferable point is not the number: whatever your equivalent proof is — a shipped system, a named customer, a checkable result — make sure it is the first thing a stranger finds.

Four funds closed this week, roughly $4.5B, and exactly one writes the cheque a seed founder is asking for. Check which vehicle the money sits in before adding a firm to your list.

Read the general-partner commitment, not the headline size: roughly 24% of the fund came from the manager and its own people. Conviction, or the price of getting a first-time vehicle over the line — both readings are plausible at this size.

The calibration for the rest of the issue: this is what it looks like when capital actually leaves a category — roughly a tenfold decline over five years, concentrated in two funds. Nothing at seed this week looked like it.