Effector Loop

Pacing the Advantage, Revised: What Ends the Cycle and How We'd Know

· updated ai credit compute cycle-analysis

Revised October 2026. This is a new draft of the whole argument, not an update appended to the old one. The September 13 edition is archived here. What changed, and why:

  • Withdrawn: the bearish overlay on CoreWeave and IREN. GPU rental prices rose all year, and at today’s prices a GPU repays its financing with room to spare. The stress is not in the unit economics.
  • Moved: where the fragility sits. “Fixed payments versus reliable cash” was true and pointed at the wrong layer. The loans secured on GPUs are well covered. The risk is in the unsecured debt above them, the assumed life of the chips, two customers, and the need to keep borrowing.
  • Split: the release bottleneck. We implied a control problem would lower demand for infrastructure. It can do the opposite, so the section now has two branches.
  • Replaced: the indicator list. “Renewal pricing” and “usage growth beyond coding” could not be shown to be wrong. Each signal now has a reading, a threshold, a source and a caveat.
  • Added: a base rate and a timeline. How earlier buildouts ended, and when the supply side of this one could catch up.

A technology can keep improving while the companies financing it get into trouble. The economy needs useful work at an acceptable cost. A borrower needs cash on the dates its contracts specify. Those are different requirements.

That opening has held up. What did not hold up was where the September post looked for the trouble. It treated a compute provider’s debt load as the weak point and suggested a small bearish position against two of them. A year of rental prices and two episodes in the bond market point somewhere else.

The revised claim is narrower. The debt in this buildout is sized to the price of compute, not the volume of it. Falling demand for AI is therefore the wrong thing to wait for. The things to watch are the price, and whether the borrowers can keep borrowing.

This is a scenario analysis, not a forecast. Rental prices run through the week ending October 4, 2026; bond and index quotes through October 9. Model inputs and thresholds are the author’s and are labelled as such.

Buildouts do not end because demand falls

Eight buildouts ended badly for the people who financed them. In each, more of the thing being built was being sold after the bust than before it, and in most the volume never dipped.

01 / Cited · eight chosen cases

In eight busts, demand rose or held and the price fell

Change in demand and in price between two cited years around each bust. Step through the three parts, or open a case for its figures and sources.

The failures followed the price. Rising volume did not prevent any of them.

British railways1845–50 −40% +56% revenue per mile of track total revenue Shareholders. Dividends averaged 1.9% in 1849, against the 7% promised.
Demand through the bust
Total revenue rose by more than half.
Supply and price
Miles in service rose 2.6×; revenue per mile fell 40%.
What failed
Shareholders. Common dividends averaged 1.9% in 1849 against the 7% in the prospectuses.

Years the bars compare: total revenue, 1845 to 1850; revenue per mile of track, 1845 to 1850.

US railroads1873–79 −45% +75% freight rate freight carried Bondholders. Defaults in 1873–75 equalled 36% of the corporate bond market.
Demand through the bust
Freight carried by a sample of thirteen lines rose every year, 75% in total.
Supply and price
Track kept opening, though new construction fell by three-quarters. What the same thirteen lines earned for moving a ton one mile fell 45%.
What failed
Defaults in 1873–75 equalled 36% of the corporate bond market.

Years the bars compare: freight carried, 1873 to 1879; freight rate, 1873 to 1879.

Fiber1997–2002 −86% about 4× cross-country line price long-distance demand Carriers, and the suppliers that had lent to them.
Demand through the bust
Demand on long-distance networks grew about fourfold between the end of 1997 and the end of 2001. US internet traffic, the fastest-growing part, roughly doubled every year, including 2001 and 2002.
Supply and price
Capacity rose about 500× as new equipment let one strand carry 100× more. A New York–Los Angeles circuit fell 86% in two years.
What failed
Carriers, and the suppliers that had lent to them. Lucent had committed up to $6.7B of customer financing.

Years the bars compare: long-distance demand, end of 1997 to end of 2001; cross-country line price, 2000 to 2002.

US cinemas1995–2001 −11% +19% tickets sold per screen tickets sold Six chains, which filed for bankruptcy in 2000–01.
Demand through the bust
Admissions rose 19% in four years, dipped 4% in 2000, then set a record in 2002.
Supply and price
Screens rose 33% as megaplexes opened and drew audiences away from older theaters, so tickets sold per screen fell 11%. Ticket prices did not fall.
What failed
Six chains filed for bankruptcy in 2000–01, among them United Artists, Carmike, General Cinema and Regal.

Years the bars compare: tickets sold, 1995 to 1999; tickets sold per screen, 1995 to 1999.

Memory chips2011–12 −45% +59% price per unit of memory memory sold Elpida, which filed in February 2012 owing ¥448B.
Demand through the bust
Micron sold 59% more DRAM, the main kind of computer memory, measured in bits.
Supply and price
The price per bit fell 45%, so its DRAM revenue fell 12%.
What failed
Elpida filed in February 2012 owing ¥448B, the largest failure in Japanese manufacturing.

Years the bars compare: memory sold, Micron’s fiscal 2012 against 2011; price per unit of memory, Micron’s fiscal 2012 against 2011.

Solar panels2011–13 −50% flat panel price panels installed in a year Suntech, which defaulted on $541M of bonds in March 2013.
Demand through the bust
About 30 GW was connected in 2011 and again in 2012; installed capacity then grew 35% in 2013.
Supply and price
Panel prices halved in 2012, and were down 80% from their 2008 peak.
What failed
Suntech defaulted on $541M of bonds in March 2013.

Years the bars compare: panels installed in a year, 2012 against 2011; panel price, during 2012.

US shale2014–16 −47% +1.9% oil price world oil use 105 producers, with about $68B of debt, by October 2016.
Demand through the bust
World oil consumption rose 1.9 million barrels a day in 2015, or 1.9%, nearly twice its usual pace.
Supply and price
US output still rose 1.0 million barrels a day in 2015 while the price roughly halved: Brent crude averaged $52 a barrel, 47% less than in 2014.
What failed
105 North American producers filed for bankruptcy by October 2016, with about $68B of debt.

Years the bars compare: world oil use, 2015 against 2014; oil price, 2015 average against 2014.

Container shipping2015–16 −19% +2–3% freight rate containers shipped Hanjin, the seventh-largest carrier, with about $5B of debt.
Demand through the bust
Global container demand grew 2–3% in 2016.
Supply and price
The fleet grew about 4% after years of ever-larger ships; Maersk’s average freight rate fell 19%.
What failed
Hanjin, the seventh-largest carrier, sought receivership in August 2016 with about $5B of debt.

Years the bars compare: containers shipped, 2016 against 2015; freight rate, 2016 against 2015.

Each bar is the change between two cited years, listed when a case is opened. The windows differ, so compare direction across cases, not size. Where no price is cited, the left bar is revenue or sales per unit of capacity (British railways, cinemas); demand for British railways is total revenue. Memory figures are one producer’s and the shipping rate is one carrier’s. The fiber demand bar is cut off at the edge of the scale: demand grew about fourfold.

These are chosen cases, not a census. Some busts do start with demand falling. The claim here is narrower: a buildout does not need one.

The pattern is clearest in fiber. Between 1997 and 2001 demand grew about fourfold. Capacity grew about five hundredfold, because new fiber arrived together with equipment that let each strand carry a hundred times more. The borrowing had been sized to the price of scarce capacity. When that price fell, the volume did not help.

Two of the cases sit close to this one. Memory chips are the same industry: Micron shipped 59% more memory in a year and collected 12% less for it, and the weakest producer went under. Cinemas show what a new format does to the one before it. The megaplex did not shrink the audience. It moved the audience away from theaters that were still being paid for, which is what a new chip generation can do to the last.

AI has its own version of that equipment. Quantized models, distillation, architectures that activate a fraction of their parameters for each token, and chips built for inference all reduce the compute behind a unit of work. Epoch AI estimates that the cost of reaching a fixed level of capability has fallen about 47% a quarter for three years, roughly 13× a year.

The financing has taken a familiar shape too. Meta’s Louisiana campus is 80% owned by Blue Owl funds, with Meta guaranteeing its residual value. Nvidia is obligated to buy CoreWeave’s unsold capacity through April 2032, and has guaranteed up to $105 billion of lease shortfalls on a campus whose tenant is an OpenAI affiliate. A supplier standing behind its customers’ obligations is what Lucent was doing in 2000.

Layered financing marks the stage of a cycle. It does not time the end of one. The position the new long-haul carriers held then belongs to OpenAI now: about $750 billion of planned compute spending through 2030, paid for by funding rounds that each need a higher price than the last.

What the data showed this year

Two things happened together. Capability got cheaper, and the hardware that produces it got more expensive to rent.

02 / Reported · rate cards and bond quotes

GPU rental prices rose while CoreWeave’s borrowing cost jumped twice

Top: median on-demand price per GPU-hour across providers. Bottom: market yield on CoreWeave’s 9.25% notes due 2030. Shaded bands mark the two yield spikes.

On-demand rental price, US$ per GPU-hour

CoreWeave 9.25% 2030 notes, yield to worst

Table view: prices and provider counts, about every eight weeks
Week ofH100H200B200B300
Oct 6, 2025$2.98(36)$3.50(21)$5.74(10)—
Dec 22, 2025$2.99(38)$3.50(24)$5.45(12)$5.45(4)
Feb 16, 2026$2.99(35)$3.54(20)$5.32(10)$6.13(4)
Apr 13, 2026$2.99(37)$3.50(25)$5.50(12)$7.25(6)
Jun 8, 2026$3.00(40)$4.00(25)$6.11(17)$8.16(8)
Aug 3, 2026$3.38(40)$4.29(27)$6.64(15)$7.99(9)
Sep 28, 2026$3.38(41)$4.50(35)$6.79(20)$8.63(16)

Provider counts in brackets. Listing counts fell by roughly half in the week of February 2, 2026 while provider medians barely moved, which reads as a change in what the scrape counts; it is one reason to use provider medians.

Sources: GetDeploying, GPU rental price history (CC BY 4.0) · Bondsupermart — CoreWeave 9.25% notes due 2030, factsheet. Retrieved October 9, 2026.

Rental prices are published rate cards, weekly, through the week ending October 4, 2026; each provider counts once. B300 is shown only from the first week with at least 5 providers. The yield is an indicative bid quote, weekly closes plus the labelled daily extremes, through October 9, 2026. A higher yield means a lower bond price and a higher cost of new borrowing. H100 and H200 are the older Hopper generation; B200 and B300 are the current Blackwell generation, and the second chip in each pair has more memory.

Across published rate cards, the median on-demand price of an H100 rose 13% between October 2025 and the week ending October 4, 2026, to $3.38 an hour across 41 providers. The H200 rose 29% to $4.50 and the B200 18% to $6.79. All four chips in the chart ended at their highest reading in the dataset, including the H100, which first shipped in 2022. Prices were flat until mid-May. Nearly all of the increase came after.

Capability, meanwhile, kept getting cheaper; the Epoch estimate above runs through this September, though it is a three-year average and not a reading for 2026 alone. Buyers did not pocket the difference. YipitData estimates that the average price actually paid per token fell only 6% in the first five months of 2026, because customers moved up to more capable models. They also used far more of them: Google reported processing about seven times as many tokens a month in May as a year earlier.

The September post modelled this with a sensitivity explorer: cheaper intelligence could require more hardware. It is now an observation. Spending on compute rose while the price of a unit of capability fell, which is what demand elasticity above one looks like. The cycle holds together for as long as that remains true.

The bond market disagreed, twice

CoreWeave’s 9.25% notes due 2030 are the most visible public price for this kind of credit. A higher yield means investors want more to hold the company’s debt.

The yield rose from 8.8% at the end of October 2025 to 12.1% on December 19. Nothing in the rental market had changed: prices were flat. By early June the yield was back at 8.3%. A short position taken on the credit signal alone would have lost money.

The second spike came in July. CoreWeave launched a $2.6 billion loan at 4.25 to 4.50 points over the benchmark rate. It cleared at 5.50 points and about 97 cents on the dollar, in a week when an AI-focused hedge fund was reported to be selling its public holdings under pressure. The notes went from 9.1% to 12.5% in a month, and the stock closed at $60.82 on July 29, 56% below its May high. This time rental prices were rising.

Change in yield, in percentage pointsCoreWeave 2030 notesSingle-B indexCCC index
First spike
Oct 31, 2025 to Dec 19, 2025		+3.35
8.77% to 12.12%		−0.10
6.77% to 6.67%		+0.14
12.33% to 12.47%
Second spike
Jun 30, 2026 to Jul 30, 2026		+3.46
9.07% to 12.53%		+0.06
7.21% to 7.27%		+0.39
13.90% to 14.29%
Since the end of August
Aug 31, 2026 to Oct 9, 2026		+0.41
11.32% to 11.73%		+0.90
7.21% to 8.11%		+2.51
14.81% to 17.32%

In both spikes almost none of the move came from the wider high-yield market. Since the end of August the opposite holds. The CCC index yield is up 2.5 points and the 30-year Treasury is at 5.6%, while CoreWeave’s notes are up 0.4, to 11.7%.

So credit and rental prices decoupled, with credit moving first and, the first time, moving alone. That is the reason for the rule this post ends on: act only when both agree.

The GPU pays for itself

03 / Modeled · author assumptions

What a B200 has to earn to pay for itself

Hourly rental price that repays debt and equity in full over the chip’s useful life, against published B200 prices.

12%
5 yrs
100%

Breakeven rental price

$2.60per GPU-hour

Published B200 medians: 24-month contract $4.50 (3 providers) · 12-month contract $4.86 (5 providers) · on-demand $6.79 (20 providers).

Debt service $1.43 Equity return $0.57 Operating cost $0.60

At these settings the chip clears its full cost with $1.90 per hour to spare against the lowest published contract price.

One input at a time, others at the starting values
Cost of debt 8% $2.4612% $2.6014% $2.67 $0.21 across the range
Useful life 4 yrs $2.955 yrs $2.606 yrs $2.36 $0.59 across the range
Hours rented 100% $2.6085% $2.9570% $3.45 $0.86 across the range

Sources: GetDeploying, GPU rental price history (CC BY 4.0). Retrieved October 9, 2026.

Every input is the author’s assumption, not a reported figure: $60,000 per installed B200, 75% debt and 25% equity, equity wanting 20%, 60 cents an hour of operating cost, and full repayment of both over the chip’s life, which is what an amortizing loan requires. Capital is recovered only on rented hours; operating cost is charged per rented hour, which flatters low utilization (treating it as fixed puts the 70% case at $3.71). Market prices are provider-median published rates for the last complete week; the 24-month and 12-month medians rest on three and five providers.

The starting case is a B200 that costs $60,000 installed, financed 75% with debt at 12% and 25% with equity that wants 20%, repaid in full over five years, with 60 cents an hour of operating cost. It needs $2.60 an hour. The lowest published contract median is $4.50 and the on-demand median is $6.79.

The sliders show which inputs matter. Two points on the cost of debt are worth 7 cents an hour. One year of useful life is worth 23 to 36 cents. Thirty points of utilization are worth 86 cents. The input the bond market reprices, the cost of capital, is the least important of the three.

The debt CoreWeave took on bought assets that earn well above their carrying cost, in a market that is still tightening. That is not a short. The same arithmetic covers IREN, the other company named in September.

Where the fragility sits

Most of CoreWeave’s GPUs are financed with delayed-draw term loans. A subsidiary borrows against a specific pool of GPUs and the customer contracts they serve, draws the money as hardware arrives, and repays principal from those contracts’ payments, reaching zero in about five years. The rate follows the customer. At June 30 the effective rate on these loans ranged from 7% on a facility backed by a hyperscaler contract to 15% on the oldest. On the arithmetic above, they are well covered.

What the bond market is pricing sits around those loans.

  • The unsecured layer on top. Net interest was $640 million in the second quarter, 25% of revenue. In the breakeven case, first-year interest on the secured loan is about 13% of contract revenue. The rest belongs to parent-company notes and convertibles with no GPUs behind them.
  • Useful life. CoreWeave depreciates its hardware over six years. The July loan runs five, and the contracts behind it average about three. After that the chips have to be re-rented at a price that still services the loan.
  • Two customers. Two customers supplied 62% of second-quarter revenue.
  • The need to keep borrowing. Capital spending exceeded operating cash flow by $5.7 billion in the quarter. Each new contract needs a new loan. The day the market will not fund one at a workable spread is the day CoreWeave cannot sign new business.

The shape of a bad outcome follows from who stands behind what. Nvidia’s obligation to buy unsold capacity, worth $6.3 billion at signing, supports the revenue that services the secured loans. That makes Nvidia something close to a lender of last resort to its own customer. A company in that position can keep operating and paying its secured lenders for years while its equity absorbs the loss. We read the likely bad outcome as a long restructuring, not a bankruptcy.

The release bottleneck has two branches

The September post, following Dario Amodei’s essay, considered a world where labs can build models faster than they can safely release them. It implied this would reduce demand for infrastructure. It need not. There are two branches, and they leave opposite marks.

Branch 1: hard capBranch 2: release bottleneck
What happensLabs stop training at the frontier because the result is unusableLabs train it but cannot safely ship it
Compute demandFalls for training-class systemsRises. Post-training, evaluation, red-teaming, monitoring and interpretability all use compute
What you would seeThe newest systems become easy to rent, the memory premium falls, hyperscalers cut capex guidanceRental prices firm and the newest systems scarce while released capability stalls
Who it reachesHardware suppliers and neocloudsThe labs’ next funding rounds

We withdraw the chain that ran from a control problem to lower GPU prices. Branch 2 is the worst configuration for the financing layer, spending without matching revenue, and it leaves rental prices firm. It points at the labs’ funding rounds, not at CoreWeave’s rental book.

Neither branch is active. Epoch’s capabilities index rose about 17 points in the twelve months to late September, above its trend of 14, though it was nearly flat from April to July.

One caution about treating model size as the signal. By outside estimates, deployed frontier models became smaller after GPT-4, and rental prices rose anyway. The spending moved to compute used when a model runs: reasoning, long-running agents, many agents at once. The question is whether labs run out of new ways to spend inference, not whether models stop getting bigger.

The signals

The September post ended with a list: renewal pricing, unrestricted liquidity, usage growth beyond coding. None of those could be shown to be wrong. The replacement sorts signals by the kind of stress they detect and the order in which they should move, with a reading and a threshold for each.

Start with ratios between chips, because they should move before prices do.

04 / Derived · ratios of published rental prices

Renters paid more for memory and less extra for a newer chip

Price ratios between chips, on-demand, weekly. The H200 is an H100 with 141 GB of memory instead of 80 GB.

The ratios this post tracks, about every eight weeks
MemoryGenerationOld vs newOn-demand ÷ 12-month
Week ofH200 ÷ H100B300 ÷ B200B200 ÷ H200H100 parityH100H200B200B300
Oct 6, 2025 1.17—1.641.560.98†0.95†1.37†—
Dec 22, 2025 1.171.00†1.561.651.130.95†1.30†—
Feb 16, 2026 1.191.15†1.501.690.861.03†1.52†—
Apr 13, 2026 1.171.321.571.631.151.101.181.28†
Jun 8, 2026 1.341.341.531.471.091.091.261.45†
Aug 3, 2026 1.271.201.551.531.111.211.37†1.21†
Sep 28, 2026 1.331.271.511.491.041.221.401.25

† One side of the ratio rests on fewer than 5 providers. The on-demand and 12-month medians come from different sets of providers, so that ratio can print below 1.0; read its direction, not its level.

Sources: GetDeploying, GPU rental price history (CC BY 4.0). Retrieved October 9, 2026.

Ratios of provider-median on-demand prices. H100 parity divides the H100 price by one-third of the B200 price, on the assumption that one B200 does the inference work of about 3 H100s; at 1.0× the older chip would cost the same per unit of work.

The H200 is an H100 with more memory and nothing else changed, so the ratio of their prices isolates what renters pay for memory. It rose from 1.17× to 1.33× over the year as long contexts and inference-heavy work bid for it. Over the same year the premium for a newer chip fell from 1.64× to 1.51×. Renters have been paying up for memory more than for raw compute. If labs stopped growing models and context, the memory premium should fall back before any absolute price does.

One tempting measure does not work. The dollar gap between a B200 and an H100 narrows early in an oversupply, as the B200 falls and the H100 holds, then widens sharply once the H100 is stranded. A measure that moves both ways cannot be a trigger.

One early mark of new supply is already there on thin data. The cheapest B300 listing fell from $7.39 to $4.99 between late July and the end of September while the median rose.

The board below replays the four signals that have weekly history against today’s lines, then traces where each kind of stress would travel.

05 / Signals · readings as of October 9, 2026

How close each signal has come to its line, and who a break would reach

Replay the year against today’s thresholds, then pick a kind of stress to trace its route through the financing stack.

Replay the year

Letters match the kinds of stress below.

DCoreWeave 2030 notes, yield

11.7%0.8 pts short of its line

AB200 on-demand premium over a 12-month contract

1.40×0.30 above its line

AH100 parity with Blackwell

1.49×0.49 above its line

BSame-chip memory premium, H200 ÷ H100

1.33×0.18 above its line

Rental test○Prices flat or lower●More providers listing○Premium down to 1.1×1 of 3

Credit test○Yield above its line0.8 pts short

Both requiredNot met

October 9, 2026, latest

October 9: the yield is 0.8 points short of its line and one of three rental conditions holds. Neither test is met.

Trace a route

3 of the 5 routes end at Nvidia.

Who it reaches, in order

  1. 1Neocloud equity
  2. 2Loans secured on Hopper-generation chips
  3. 3Nvidia’s CoreWeave backstop
  4. 4Nvidia’s gross margin

Slow. Appears in rental prices over weeks to quarters.

What we watch

  • Blackwell on-demand premium over a 12-month contract

    NowB200 1.40× (5 providers on the 12-month side)

    LineToward 1.1× while provider and listing counts rise. Spare capacity removes the premium for not committing, and it should show on the newest chip first.

  • B200 on-demand price

    Now$6.79 on 20 providers

    LineBelow $6.00, its floor since June, with more providers listing than today.

  • H100 parity with Blackwell

    Now1.49×

    LineFalling toward 1.0×. It moves last and matters most, because it strands loans secured on the older chips.

Who it reaches, in order

  1. 1Demand for training-class systems
  2. 2Hyperscaler capex guidance
  3. 3Nvidia

That is Branch 1. Branch 2 reaches the labs’ funding rounds instead: follow C.

Slow as a plateau. Sudden if an incident forces a pause.

What we watch

  • Same-chip memory premium

    NowH200 ÷ H100 1.33×; B300 ÷ B200 1.27×

    LineH200 ÷ H100 drifting back toward 1.15× before absolute prices move. That would say labs have stopped growing models and context (Branch 1).

  • Released capability against its price

    NowEpoch’s index rose about 17 points in 12 months, above its 14-point trend, with a flat stretch from April to July. Not stalled.

    LineFrontier indices flat for six to twelve months while the price of a fixed capability keeps falling.

  • A visible control incident

    NowNone that has led to mandated evaluations or a pause.

    LineAny. Unlikely in a given quarter, and it would reach financing first.

Who it reaches, in order

  1. 1OpenAI
  2. 2Oracle, CoreWeave and Nvidia as its counterparties
  3. 3A sudden stop in which Nvidia’s $105B guarantee is drawn

Lumpy. Arrives at funding rounds.

What we watch

  • Price of the labs’ own equity

    NowOpenAI’s last primary round valued it at $852B in March. An August tender was reported flat at that price. A round near $1.4T is reported in talks. Anthropic raised at $965B in May.

    LineA primary round that is flat or down, or a tender below the last primary.

  • Public proxies for lab credit

    NowSoftBank five-year CDS near 385 bp in mid-September, a three-year high. Oracle five-year CDS reported at a record 261 bp on October 8.

    LineNew highs that persist after a lab round closes.

  • Gap between open and closed models

    NowOpen models trail by about four months, wider than a year ago.

    LineNarrowing toward parity while lab revenue growth slows.

Who it reaches, in order

  1. 1CoreWeave’s ability to sign new contracts
  2. 2Its equity
  3. 3Its unsecured notes
  4. 4Secured lenders last

Live. Quoted daily and free to read.

What we watch

  • CoreWeave 2030 notes, yield to worst

    Now11.73% bid

    LineAbove 12.5%, the July high: refinancing stress. Below 10%: funding fully open.

  • Market or name

    NowSince the end of August the notes are up 0.4 points and the CCC index 2.5. Led by the market.

    LineThe notes widening while the index does not.

  • Terms on the next GPU loan

    NowJuly’s $2.6B loan cleared at SOFR + 5.50% and about 97 cents on the dollar, with a 1.35× debt-service covenant.

    LineWider than SOFR + 6.00%, or a tighter covenant. Either slows new capacity.

  • Landlords whose tenant is CoreWeave

    NowGalaxy sold $3.5B of secured notes at 9.875% in July on a site leased to CoreWeave. A second leased site raised $1.1B at a 9.25% yield in September.

    LineA deal of this kind failing to price, or these notes falling without CoreWeave’s own.

Who it reaches, in order

  1. 1Financing for projects not yet built
  2. 2Developers and landlords
  3. 3Rental prices last

Slow for power politics. Sudden for regulation.

What we watch

  • Electricity-price backlash and data-center moratoria

    NowPJM’s capacity auction cleared at its price cap for a third year in July. New York paused permits for data centers of 50 MW or more for up to a year.

    LineA moratorium in a major data-center state, or regulators shifting grid costs onto data centers.

All 14 signals, with sources and caveats

AToo much supply

Slow. Appears in rental prices over weeks to quarters.

1Blackwell on-demand premium over a 12-month contract

Now
B200 1.40× (5 providers on the 12-month side)
Changes at
Toward 1.1× while provider and listing counts rise. Spare capacity removes the premium for not committing, and it should show on the newest chip first.
Source
GetDeploying. Thin tier. B300’s premium printed 1.03× in July on two providers, so B200 is the one to watch.

2B200 on-demand price

Now
$6.79 on 20 providers
Changes at
Below $6.00, its floor since June, with more providers listing than today.
Source
GetDeploying. Published rate cards, not the negotiated multi-year contracts the loans are written against.

3H100 parity with Blackwell

Now
1.49×
Changes at
Falling toward 1.0×. It moves last and matters most, because it strands loans secured on the older chips.
Source
GetDeploying, SemiAnalysis InferenceX. Assumes one B200 does the work of three H100s. Measured ratios range from about 1× to more than 4× by model.

Who it reaches, in order: Neocloud equity → Loans secured on Hopper-generation chips → Nvidia’s CoreWeave backstop → Nvidia’s gross margin

BCapability plateau or control problem

Slow as a plateau. Sudden if an incident forces a pause.

1Same-chip memory premium

Now
H200 ÷ H100 1.33×; B300 ÷ B200 1.27×
Changes at
H200 ÷ H100 drifting back toward 1.15× before absolute prices move. That would say labs have stopped growing models and context (Branch 1).
Source
GetDeploying, Nvidia. B300 also has about 1.5× the low-precision compute of B200, so its ratio is less clean.

2Released capability against its price

Now
Epoch’s index rose about 17 points in 12 months, above its 14-point trend, with a flat stretch from April to July. Not stalled.
Changes at
Frontier indices flat for six to twelve months while the price of a fixed capability keeps falling.
Source
Epoch AI, capabilities index, Epoch AI, price of capability. Index values are re-estimated between releases.

3A visible control incident

Now
None that has led to mandated evaluations or a pause.
Changes at
Any. Unlikely in a given quarter, and it would reach financing first.
Source
Amodei. A step change, so it gives no early reading.

Who it reaches, in order: Demand for training-class systems → Hyperscaler capex guidance → Nvidia That is Branch 1. Branch 2 reaches the labs’ funding rounds instead: follow C.

CLab financing

Lumpy. Arrives at funding rounds.

1Price of the labs’ own equity

Now
OpenAI’s last primary round valued it at $852B in March. An August tender was reported flat at that price. A round near $1.4T is reported in talks. Anthropic raised at $965B in May.
Changes at
A primary round that is flat or down, or a tender below the last primary.
Source
SiliconANGLE, Bloomberg, TechCrunch, Clarity, Caplight. Secondary-market prices are thin, delayed and indicative.

2Public proxies for lab credit

Now
SoftBank five-year CDS near 385 bp in mid-September, a three-year high. Oracle five-year CDS reported at a record 261 bp on October 8.
Changes at
New highs that persist after a lab round closes.
Source
Bloomberg, SoftBank CDS, TradingKey, Oracle CDS. These CDS levels are press reports, not a data feed.

3Gap between open and closed models

Now
Open models trail by about four months, wider than a year ago.
Changes at
Narrowing toward parity while lab revenue growth slows.
Source
Epoch AI, open-model gap. A benchmark gap is not a revenue gap.

Who it reaches, in order: OpenAI → Oracle, CoreWeave and Nvidia as its counterparties → A sudden stop in which Nvidia’s $105B guarantee is drawn

DFinancing layer

Live. Quoted daily and free to read.

1CoreWeave 2030 notes, yield to worst

Now
11.73% bid
Changes at
Above 12.5%, the July high: refinancing stress. Below 10%: funding fully open.
Source
Bondsupermart. Indicative quotes. Check trade prints on FINRA TRACE.

2Market or name

Now
Since the end of August the notes are up 0.4 points and the CCC index 2.5. Led by the market.
Changes at
The notes widening while the index does not.
Source
FRED, CCC index yield, FRED, single-B index yield. Rated BB− by Fitch, but trading between the single-B and CCC indices.

3Terms on the next GPU loan

Now
July’s $2.6B loan cleared at SOFR + 5.50% and about 97 cents on the dollar, with a 1.35× debt-service covenant.
Changes at
Wider than SOFR + 6.00%, or a tighter covenant. Either slows new capacity.
Source
CoreWeave 8-K, Bloomberg, loan repricing. Each loan is tied to different customers, so terms are not like for like.

4Landlords whose tenant is CoreWeave

Now
Galaxy sold $3.5B of secured notes at 9.875% in July on a site leased to CoreWeave. A second leased site raised $1.1B at a 9.25% yield in September.
Changes at
A deal of this kind failing to price, or these notes falling without CoreWeave’s own.
Source
Galaxy, Octus, Bloomberg, data-center notes. Secured on buildings and leases, so not the same risk as the parent’s notes.

Who it reaches, in order: CoreWeave’s ability to sign new contracts → Its equity → Its unsecured notes → Secured lenders last

EOutside shocks

Slow for power politics. Sudden for regulation.

1Electricity-price backlash and data-center moratoria

Now
PJM’s capacity auction cleared at its price cap for a third year in July. New York paused permits for data centers of 50 MW or more for up to a year.
Changes at
A moratorium in a major data-center state, or regulators shifting grid costs onto data centers.
Source
PJM, Davis Polk. Restricts new supply, so it reaches financing before rental prices.

Who it reaches, in order: Financing for projects not yet built → Developers and landlords → Rental prices last

Lines are today’s thresholds, applied to the whole year; they were not in force a year ago. Each track runs from a far point on the left to its line, so a marker moving right is moving toward the level that would change our view. Rental ratios are weekly provider medians. The yield is the weekly close, with the December and July daily peaks shown in their weeks. The last frame pairs the October 9 yield with rental prices for the week ending October 4. The memory premium starts beside its line because that line marks where it stood before renters began paying up for memory. Thresholds are the author’s judgment, recorded in advance so they can be checked later. They are not forecasts and carry no probability.

The replay shows something the tables did not. Measured against today’s lines, December 2025 came closer on both tests than the market is now: prices were flat, providers were multiplying, and the yield was within half a point. The premium for not committing never narrowed, and that one condition is what kept the rental test from tripping on a false alarm. It is in the rule for that reason.

When

06 / Cited constraints and author’s judgment

When supply could catch up

How long each constraint is reported to last, the contract expiries already on the calendar, and where we place the most likely ending.

Supply constraint, as reportedScheduled eventAuthor’s judgment

Chip packagingeasing through 2027

The shortfall in advanced packaging is reported narrowing from about 20% to about 10% by the end of 2026. TrendForce, June 2026.

High-speed memorytight through 2027

The memory used in AI chips is forecast to stay supply-constrained through 2027, with prices still rising. TrendForce, September 2026.

Powerinto the 2030s

The largest turbine maker is taking reservations for 2031 delivery. Utility Dive.

Early rental contracts expire2027 to 2029

CoreWeave’s contracts averaged about four years at the end of 2024, on terms of two to five. Those signed in 2023 and 2024 come up for renewal across these years. CoreWeave S-1.

Most likely ending2027 to 2028

Rental prices drift toward cost, Nvidia’s margin falls, and leveraged neoclouds are restructured. Tight memory pushes this toward the later year.

Tail risk: financing fails first2027

A lab round that does not clear at a higher price, or a leveraged backer under strain, a year before supply catches up.

Constraint bars show how long each shortage is reported to last, not a date on which supply and demand balance; no supplier has claimed balance. The last two rows are the author’s judgment about a scenario, not a forecast, and carry no probability.

The supply side is not close to catching up. Packaging is easing, memory is forecast to stay tight through 2027, and power is the binding constraint into the next decade: the median US generation project finished in 2025 took more than five years to get from its grid-connection request to operation.

The nearest thing to a scheduled supply event sits in the rental book. CoreWeave’s contracts averaged about four years at the end of 2024, so those signed in 2023 and 2024 come up for renewal from about 2027. That is when Blackwell supply should be at its peak and its successor, which Nvidia said would begin shipping in the second half of 2026, is ramping.

Our base case is a normalization in 2027–28. Blackwell rentals drift toward the $2 to $3 an hour where the breakeven sits. Nvidia’s gross margin, 75.0% last quarter and guided to 74.0%, moves toward the 60s as its customers’ alternatives improve; TrendForce expected shipments of cloud providers’ own chips to grow 44.6% in 2026 against 16.1% for GPUs. Leveraged neoclouds are restructured. It does not look like 2000, because the hyperscalers’ balance sheets hold.

The tail risk is that the financing layer fails a year before supply arrives: an OpenAI round that does not clear at a higher price, or a leveraged backer under strain. SoftBank repaid its OpenAI bridge loan in September and funded its next OpenAI payment with new senior notes, which moves the leverage without removing it.

Position

Re-check monthly: five ratios (B200 on-demand ÷ 12-month, H100 parity, H200 ÷ H100, B300 ÷ B200, B200 ÷ H200), one bond yield (CoreWeave 2030), and one index (the CCC yield on FRED).

Capability can keep advancing through every scenario above. The debt was sized to this year’s prices, and it will continue on its original schedule.

Sources and data notes

The September 13, 2026 edition is kept unchanged with its original charts and source notes.

Rental prices. GetDeploying, GPU rental price history, https://getdeploying.com/gpus, CC BY 4.0. The file is a scrape of published rate cards: the uncontracted retail end of the market, not the negotiated two-to-five-year contracts the loans are written against. That market is visible only when contract values appear in filings. All prices here are the provider median (each provider counts once) for the last complete week, September 28, 2026 to October 4, 2026; the week still being captured when the file was retrieved on October 10, 2026 is dropped. A provider count travels with every price. Medians over fewer than 5 providers are not plotted and are marked in tables. Reservation tiers beyond 12 months rest on one to five providers and jump when one joins or leaves.

Bond and index quotes. The CoreWeave yield is an indicative bid quote from a retail bond platform, not a trade print; the chart uses the last quote of each week plus the labelled daily extremes. Index comparisons use ICE BofA effective yields from FRED at the dates shown, with the last reading on October 8. Share prices are daily closes.

Model and thresholds. Every breakeven input is the author’s assumption. The three-to-one throughput figure behind “H100 parity” is also an assumption: the specification ratio is 2.3× and measured results range from about 1× to more than 4× depending on the model. Thresholds in the signals table and the position box are judgments recorded in advance, not estimates of probability.

Reported second-hand. OpenAI’s August tender and pending round, the hedge-fund selling in July, the final discount on the July loan, the two credit-default-swap levels, and Moody’s contract-life figure come from press reports, not primary documents, and are worded as reported.