25 September: the money got bigger and softer. CoreWeave’s convertible closed upsized at $4.2bn — a 2.875% coupon, paid for in dilution, due 2033. Cheap in cash, expensive in shares. None of the five physical tells moved.
Five checkable tests of whether the AI build-out is real demand or malinvestment, plus the five physical tells that would show a slowdown before any share price does. Every figure is dated and sourced, and we update it every week — including the weeks when nothing moves.
Nothing in the five physical tells moved. There was no new transformer lead-time survey, no interconnection-queue print and no data-centre starts release, so those rows carry the same figures as last week or no figure at all.
What moved was financing, and mood. CoreWeave’s August facility priced at SOFR + 550bp against May’s SOFR + 450bp at the same Ba2 / BB+ rating12 — the same credit grade at a higher price, which is what the marginal dollar getting riskier looks like while the headline number still grows. Separately, on 14 September a public call from Anthropic’s chief executive to slow frontier AI development knocked data-centre power names down sharply (GE Vernova −8.6%, Vertiv −7.7%)34, on the same day the US 10-year Treasury yield touched 5%5.
New this week, 22 September — and it answers last week’s open question. Issue #1 said the thing that would move this test was “the notes pricing and closing”. They closed, and they came in bigger: CoreWeave completed an upsized $4.2bn offering of 2.875% convertible senior notes due 2033 — $3.0bn announced, plus the initial purchasers’ $500m option exercised in full — with net proceeds of $4,137.0m25. The initial conversion price is about $97.85 a share, a 22.5% premium to the 17 September close, and part of the proceeds paid for capped call transactions that push the effective dilution point higher. The notes are unsecured and not collateralised on GPUs.
Set the two prices side by side, because they are the same company in the same month. The August secured facility priced at SOFR + 550bp. This convert pays a 2.875% coupon. The convert is far cheaper in cash because the buyer is also being handed an equity option — the company is paying in dilution instead of interest, and pushing the maturity to 2033, well past the three-to-six-year working life of the hardware the money buys.
Our read: last week we wrote that the thing which would move this test was “the notes pricing and closing”. They closed — and came in $1.2bn larger than announced. Read plainly, that is not a market refusing to fund this. It is a market happy to fund it in a form that costs almost nothing in cash today and settles in shares eight years out. The grade stays SPLIT, moving the wrong way: the hyperscaler core is still cash-funded, while the marginal layer keeps finding new unsecured ways to borrow, each one further from the hardware it pays for. What would move it now: a secured facility priced tighter than August’s SOFR + 550bp, or the first sign that the convert market will not take the next one. Nothing else this week touched a test, and the order books did not move at all.
The 14 September sell-off and the slowdown essays. A chief executive asking for guardrails is not a capital-expenditure cut — one is a statement, the other shows up in an order book — and a share price is not a print, so neither of them changes a test on this page.
Published 10 June 2026, re-scored 17 September 2026. Direction describes the underlying condition since June — the evidence, not a share price and not our opinion of the stock.
Scroll the table sideways on a phone →
| Test | June 2026 | Latest reading | As of | Last week | Direction | Source |
|---|---|---|---|---|---|---|
| 1. Savings or credit? Is it funded out of cash flow, or borrowed? |
SPLIT, moving the wrong way. Hyperscaler core (about $725bn of capex) cash-funded; the marginal layer turning to debt — an $8.5bn GPU-backed facility at A3 / A (low), SOFR + 225bp6 | CoreWeave total debt $35,551m (about $35.6bn) at 30 June 2026, from about $21bn at end-2025 and about $8bn at end-20247. Both publicly syndicated facilities since are below investment grade: May at Ba2 / BB+, SOFR + 450bp1; August at the same rating, SOFR + 550bp2. Same grade, a full point wider. New, 22 September 2026: the convertible closed upsized at $4.2bn, 2.875% due 2033, net proceeds $4,137.0m, unsecured and not GPU-collateralised25. Total debt above is the 30 June figure and does not include it. | 30 Jun 2026 (debt) 10 Aug 2026 (pricing) |
same $35,551m; SOFR + 550bp |
WEAKENING | 10-Q; company release; 8-K712 |
| 2. Capacity ahead of demand? What does it cost to rent an hour of the asset everyone is buying? |
FLASHING. H100 rentals down from about $8/hr to $2.85–3.50, roughly a 64% collapse, as 300+ providers piled in while capex accelerated 77%8 | Median on-demand H100 about $3.39/hr, up about 10% over 90 days and about 11% year over year9. Prices rose through 2026 even as H200 and Blackwell shipped. This is the reading we have seen ourselves; a higher secondhand bank estimate is kept in the sources, not in this cell10. | 19 Aug 2026 (median) | same $3.39/hr, +11% y/y |
STRENGTHENING | Provider index9; BofA via press, secondhand10 |
| 3. Circular financing? Is the seller funding the buyer? |
PRESENT. Nvidia $2bn into CoreWeave; Meta about $21bn of capacity committed11; Anthropic and Google paying SpaceX and xAI about $26bn a year for compute12. We under-counted this in June — see the next cell. | Nvidia had already agreed in February to invest $30bn in OpenAI — public when we wrote in June, and missing from our June list13. New since June: in August Nvidia signed a guarantee of up to $105bn behind OpenAI’s lease and power payments on an Ohio campus — a backstop, not a cash investment14; it was reported at up to $250bn in July and cut before signing15. Oracle plans to raise $45–50bn of debt and equity this year to build for OpenAI16. | 19 Feb 2026 (stake) 17 Aug 2026 (guarantee) |
same $105bn guarantee |
WEAKENING | CNBC; Fortune; DCD13141516 |
| 4. Is the narrative doing the work? | PARTLY. “Scaling laws plus AGI timelines” as the unaudited growth claim of this cycle | Now the loudest argument in the market: a slowdown call from Anthropic’s chief executive, publicly backed by others, a broad AI sell-off on 14 September34, and analysts openly split on whether a credible safety framework makes long-duration capital easier or harder to underwrite17. Nvidia’s chief executive spent August rejecting the “circular financing” label18. We do not grade this one; its measure is vibes. | 15 Sep 2026 | same ungraded |
LIVE | CNBC; Reuters; Benzinga31718 |
| 5. Do obsolete chips still clear near book? Is the second-hand market agreeing with the accounting? |
WEAKENING. One of the five largest buyers has cut the life it expects from its own servers, and said why | Amazon shortened the useful life of a subset of its servers and networking equipment from six years to five, effective 1 January 2025, and separately booked $920m of accelerated depreciation for equipment it decided to retire early. Amazon's own stated reason: “an increased pace of technology development, particularly in the area of artificial intelligence and machine learning”23. The realised effect in 2025 was $1.4bn more depreciation and $1.0bn less net income, mostly at AWS24. Thirteen months earlier the same company had lengthened the same assets to six years. Against this: CoreWeave has an A100 contract running into 2029 and says it re-booked expiring H100s at 95% of their original pricing — but that reaches us through trade press, not a filing, so it does not move the direction. | 1 Jan 2025 (effective) Feb 2025 (filed) FY2025 (realised) |
new test not in issue #1 |
WEAKENING | 10-K2324 |
Last week: from this issue onward that column carries issue #1’s actual figures, published 18 September — not a repeat of this week’s. Where it says no new print, the underlying document has not been refiled and the figure is genuinely unchanged; where it says same, a newer source exists and reads the same. Those are different facts and the column shows both, which is what makes “updated weekly” something you can check rather than a caption you have to take our word for.
We wrote this test's rule on 20 September, before looking at any of the data, because a test invented after you have seen the answer is not a test. The rule says the direction goes to WEAKENING if any of the five largest buyers shortens the useful life it assigns to its own servers. We also wrote down what we expected to find: STRENGTHENING, because the loudest available evidence pointed that way.
We were wrong. Amazon had already shortened, in a filing from February 2025, and said in plain language that AI was the reason. Our own rule marked the cell against the answer we anticipated.
Why this test exists. The big buyers have told their auditors these machines last six years, and that number is fixed in a filing for years at a time. The second-hand market re-prices every week. If demand is real, obsolete chips keep clearing near what the books say they are worth. If it is not, the resale price cracks first — before any income statement can show it, because the accounting cannot move that fast. What would move this cell back: the CoreWeave rebooking figures confirmed in a filing or transcript rather than trade press, and no further shortening in the next round of annual reports.
Its measure is vibes. We track it because it moves money — it moved a great deal of money on 14 September — but we will not convert a mood into a score and then quote the score back at you as evidence. The other three tests have numbers with dates on them. This one does not, and pretending otherwise is how a framework stops being worth reading.
The AI build-out is a physical event. It needs transformers, switchgear, transmission, cooling, memory and optics, and those have lead times measured in years. A physical build-out slows in physical ways, and those show up in order books before they show up in anybody’s stock. These five are the tells. Most weeks, most of them will say no new print. Two of the five carry a figure today; the other three have no baseline of ours yet and sit in their own section below the table until they do.
Scroll the table sideways on a phone →
| Tell | What would signal a slowdown | Last print | As of | Last week | Status |
|---|---|---|---|---|---|
| 1. GPU rents vs capex | Rents falling while capex also falls. The 2024–25 fall happened against rising capex, which is why we now read it as a supply event rather than a demand failure. Both falling together is the real signal. | Rents rising: median on-demand about $3.39/hr, up about 11% y/y9. Capex still accelerating: about $725bn of hyperscaler capex guided for 2026, up about 77% y/y21 | 19 Aug 2026 (rents) Feb 2026 (capex) |
same $3.39/hr vs $725bn |
no slowdown |
| 2. Transformer lead times | Lead times shortening from current levels. If the queue gets shorter, orders are being cancelled. The single cleanest signal available, and it is not a market price. | About 128 weeks from order to delivery on Wood Mackenzie’s mid-2025 survey19; larger substation units now running past 160 weeks20 | mid-2025 survey 2026 trade reporting |
same 128 / 160+ weeks |
no new print |
A tell with no baseline is not yet a measurement, so tells 3 to 5 sit here rather than in the table above, where an empty row still reads like a reading. We will add one source a month, published with its date and its basis, instead of inventing numbers to fill them.
Our conclusion is own the bottleneck, not the boom. We would drop it if transformer lead times shortened materially from about 128 weeks while data-centre starts fell — that combination is a genuine slowdown in the physical chain, and it would mean the scarce inputs are no longer scarce. We would drop the other half of it, the caution on the leveraged middle, if GPU-backed paper started pricing tighter at the same rating and a syndication cleared without a concession. A sell-off in the shares of companies that sell into the build-out is neither of those things. Next full re-score: December 2026, on all four tests plus these five tells.
Each entry gives the date the figure refers to — not the date of the article that reported it — and the date we last checked it. Every source below is one used in our September re-score or the original June piece.
No figure on this page is new reporting. All of it is carried from our 17 September 2026 re-score and our 10 June 2026 piece, with the source and the date it refers to attached. Where a number is reported rather than confirmed by a filing, the row says so. Three of the five physical tells carry no baseline of ours yet; they are listed under “Baselines we are still building” instead of being shown as readings, and we will not borrow a number from somewhere else to fill them.
Educational research, not personalized investment advice. Dragonfly Lens is not a registered investment advisor. Companies are named to explain the framework, not as buy or sell recommendations. Past performance does not guarantee future results.