On 10 June we published four specific, checkable tests for whether the AI build-out is real demand or malinvestment — and said plainly that these numbers, not the arguing, would settle it. This week the market had exactly the scare those tests were written for: Anthropic chief executive Dario Amodei published an essay calling for a slowdown in frontier AI development, OpenAI’s Sam Altman and Elon Musk publicly backed it, and on 14 September data-centre power names fell sharply (GE Vernova -8.6%, Vertiv -7.7%) alongside chipmakers. The same day, the US 10-year Treasury yield touched 5% for the first time since 2023. So it is a good moment to do the unglamorous thing and mark our own work. The result is not what we expected. The number we were most worried about has got better. The two we treated as secondary have got worse — and on both, we missed things in June that were already public.
| Test | June 2026 verdict | September 2026 | Direction |
|---|---|---|---|
| 1. Savings or credit? | SPLIT, moving the wrong way | Debt ~$35.6bn (30 June), newest paper rated junk and priced wider | WORSE |
| 2. Capacity ahead of demand? | FLASHING | H100 spot +33% y/y (BofA); median on-demand +11% | BETTER |
| 3. Circular financing? | PRESENT | Under-counted in June (a $30bn Nvidia stake already existed); since then, a signed guarantee of up to $105bn | WORSE |
| 4. Is the narrative doing the work? | PARTLY | Now the loudest argument in the market | LIVE |
The short read: the overcapacity signal that worried us most in June has reversed, while the financing underneath the boom has kept deteriorating — and was already worse in June than we said. That is close to the opposite of the story the market told itself this week.
It is easy to publish a framework. It costs nothing, it always sounds wise, and nobody comes back to check. The only thing that makes a framework worth reading is whether the person who wrote it returns to it when the numbers are inconvenient.
So: in June we did not say “it is a bubble” or “it is not.” We said an investment boom turns into malinvestment when specific conditions hold, listed four of them, and wrote down what each looked like at the time. Three months later, three of the four have moved enough to score; the fourth we mark but decline to grade. Here they are, one at a time — including two places where our June piece left out things that were already public.
Why it matters, in plain terms. A company spending money it already earned can be wrong without taking anyone else down — it simply stops. A company spending borrowed money has lenders, covenants and maturity dates, and when it stops, the stopping is disorderly. The same building gets built either way; what differs is what happens if demand disappoints.
What we said in June. Split, and moving the wrong way. The hyperscaler core — roughly $725bn of capex from Microsoft, Alphabet, Amazon and Meta — is funded mostly by genuine operating cash flow, which makes it unlike 1999 in an important way. But the marginal layer had turned to debt: CoreWeave’s $8.5bn facility secured against GPUs, the first investment-grade rating on financing secured by GPUs plus a customer contract (in this case Meta’s), plus $3.1bn publicly syndicated in May.
What happened since. It has deepened, and the quality has fallen.
| Facility | Size | Closed | Rating |
|---|---|---|---|
| DDTL 4.0 | $8.5bn | Mar 2026 | A3 / A (low) — investment grade; floating tranche SOFR + 225bp; backed by Meta contracts |
| DDTL 5.0 (publicly syndicated) | $3.1bn | May 2026 | Ba2 / BB+ — below investment grade; priced at SOFR + 450bp |
| DDTL 5.5 | $2.6bn | Aug 2026 | Ba2 / BB+, priced at SOFR + 550bp |
CoreWeave raised more than $30bn in debt and equity in 2026 through mid-August, and its debt reached about $35.6bn at 30 June 2026 — up from about $21bn at the end of 2025 and about $8bn at the end of 2024 (company 10-Q, filed 11 August).
Score: WORSE. We flagged this as split and moving the wrong way. It kept moving that way, and faster than we expected.
June also named trip-wires, so we score those too. GPU rentals sliding through $2/hr while capex stays high: not triggered (about $2.80 spot). Widening spreads on GPU-backed paper: partly triggered (SOFR + 450bp in May, + 550bp in August, at the same rating). Failed syndications: not triggered — both deals were reported oversubscribed.
Why it matters, in plain terms. If you want to know whether a hotel-building boom is justified, do not count cranes — check the nightly room rate. Building while the rate collapses means capacity has outrun anyone’s willingness to pay for it. For AI, the room rate is what it costs to rent an hour of GPU time.
What we said in June. Flashing. H100 rental prices had fallen from roughly $8/hr to $2.85–3.50 — about a 64% collapse — as more than 300 providers piled in, while capex accelerated 77%. Falling utilisation prices alongside an accelerating build-out is the textbook gap. We did note a counter-signal and said the market was “genuinely contested, not one-way.”
What happened since. The counter-signal won, and decisively.
Score: BETTER. Materially. And note where that leaves the argument: the test most often cited as proof of an AI bubble is the one that has improved.
Why it matters, in plain terms. If a shopkeeper lends you the money to buy his goods, his sales figures partly measure his own lending, not your appetite. The revenue is real on the way in and vulnerable if the lending ever stops.
What we said in June. Present. Nvidia had invested $2bn in CoreWeave, whose business is buying Nvidia chips; Meta committed roughly $21bn to CoreWeave capacity; Anthropic and Google together pay SpaceX and xAI around $26bn a year for compute built partly to serve them. Real contracts and real money — but the same names kept appearing on both sides of the table.
What happened since. It is bigger than we wrote — partly because we under-counted it in June.
The honest counter-argument, which deserves stating properly. Nvidia’s Jensen Huang publicly rejects the “circular financing” label, writing in August that “OpenAI will pay the lease”, and that OpenAI’s plans could represent roughly $600bn of Nvidia compute through 2030 — his estimate, not a signed contract. Asset manager Janus Henderson, among others, describes the same web as a “virtuous circle” — long-term purchase commitments paired with financing are a normal way to lock in supply in a capacity-constrained industry. Both points are fair. Vendor financing is not fraud and is not new.
Score: WORSE. Not “much worse”: much of the gap between June and now is our own under-count, and we would rather say that than inflate the change.
This was the softest of the four in June, and it is now the loudest thing in the market. Anthropic’s chief executive called for slowing frontier AI development and OpenAI’s backed him; AI stocks fell broadly, and power-equipment suppliers such as GE Vernova and Vertiv fell around 8%. Nvidia’s chief executive spent August publicly rejecting a label. Analysts are split on whether a credible safety framework makes long-duration capital easier to underwrite rather than harder.
We are not going to score a test whose measure is essentially vibes. But one distinction is worth holding on to, because a great deal of money moved on it this week: a chief executive asking for guardrails is not a capital-expenditure cut. Those are different events. One is a statement; the other shows up in orders. Which brings us to the only thing we would actually watch.
The whole point of this site is that the AI build-out is a physical event. It needs transformers, switchgear, transmission, cooling, memory and optics, and those things have lead times measured in years. A physical build-out slows down in physical ways, and those show up in order books before they show up in anybody’s stock.
None of those five is what moved this week. What moved this week was a statement (and, on the same day, a 5% ten-year yield), and the price of things that sell into the build-out. That may turn out to have been early. It is not, on its own, evidence.
The June conclusion was own the bottleneck, not the boom: the leveraged, fast-depreciating middle — rented single-purpose compute financed at fixed rates against a three-to-six-year asset — only wins if the boom continues, while the scarce, slow-depreciating inputs are twenty-to-forty-year assets with demand from grid replacement, electrification and reshoring whether or not AI keeps compounding.
Three months of evidence has not changed that conclusion. It has sharpened it, and in a slightly uncomfortable direction: the demand looks more real than we thought, and the financing looks more fragile. Those two findings do not cancel out. They point at the same place: the framework points at the part of the chain that gets paid in both scenarios, and flags the part that only survives one.
Next check: December 2026, on the same four tests, with the five physical tells above added as a fifth. If transformer lead times have not shortened from those levels by then, whatever happened this week was a mood, not a slowdown.
Figures dated where cited. June baselines are from our 10 June 2026 piece, Real Demand or Malinvestment?. September figures: CoreWeave facility sizes, dates and ratings from company announcements and rating-agency reporting; H100 pricing from published provider comparisons and a Bank of America estimate for August 2026; Nvidia-OpenAI figures from company statements and contemporaneous reporting, where the $600bn figure was reported and the $105bn guarantee signed. Where a number is reported rather than confirmed by a filing, we have said so. 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.