The $42 billion of AI debt launched this week sits on someone else's balance sheet

Key takeaways

• About $42B of AI-compute debt launched into syndication this week at SOFR+150 and +187.5 bp, priced primarily on Broadcom's partial guarantee (launch spreads, book not closed). The $18B junior tranche that carries Anthropic's own risk has not launched; it is the mark to wait for.
• Google's 3,590 MW Constellation deal is 75% existing nuclear output; Meta's 2,609 MW Vistra PPAs, signed in January and backed since October 5 by a conditional $4.2B DOE loan, are 83%. On October 9 FERC rejected PJM's proposed bring-your-own-new-capacity rule, as Google and Constellation had asked. Buy existing firm supply now.
• Model prices fell by segment: Haiku 5.5's cut is 90% below 100K prompt tokens and 50% above, and Sol Ultrafast charges six times list for speed. Route by context length; the one unambiguous agentic cut was Sonnet 5.5's cache-read halving.
• Memory and foundry took the visible returns (Samsung's KRW 107.4T operating quarter, TSMC +54.6% in September); the only interconnect pricing-power statement was Lumentum's sold-out horizon moving out to early 2029.
• House measurement: Polaris Forge 1's rent covers Applied Digital's 9.25% notes about 1.0x in 2028 once amortization starts, with a $1.80B balloon due December 2030; the first-generation AI-factory bond is a refinancing bet, not a repayment plan.
• All five house hypotheses are strained this week; the common cause is that the physical supply chain (memory, optics, power) is where both returns and constraints landed. All seven W40 predictions stay pending, with PJM's October 29 backstop refile the nearest to resolving.

By the numbers

Senior loans launched for Anthropic's chip leases

$42B — Two $21B tranches at launch spreads of SOFR+150 and +187.5 bp (SOFR is the overnight dollar benchmark rate; a basis point is one hundredth of a percentage point), partially guaranteed by Broadcom; part of a $61.5B package with an $18B junior tranche (the slice repaid last, so it absorbs losses first) not yet launched (IFR, Bloomberg, FT, anonymous-sourced)

Broadcom five-year CDS, the annual cost of insuring its debt against default

133 bp — October 9 reading, more than tripled since late May (LSEG via IFR); the live public proxy for guaranteed frontier-lab chip credit until the junior tranche prices

Coupon on the only unguaranteed Anthropic chip paper on record: $4.5B of B notes in the June private package

8.5% — Reported by PitchBook (June 2) and The Inference (October 5); outside this issue's graded set. The guaranteed A2 notes in the same package printed at 5.75%, so the guarantee was worth about 275 bp in June

Share of Google's 3,590 MW Constellation deal that is existing nuclear output

75% — 2,700 MW fifteen-year supply from running reactors plus 890 MW of uprates (output increases at existing plants), first delivered by 2028; Meta's January Vistra PPAs are 83% existing

House-estimated 2028 debt-service coverage, Applied Digital 9.25% notes

~1.0x — Rent after operating costs divided by interest plus scheduled principal; 1.01x on the house arithmetic, so the rent just covers the payments. The 7% notes cover about 1.5x on the same rent (filing-derived; method and caveats in the house measurement)

Samsung preliminary Q3 operating profit

KRW 107.4T — +782.5% year on year on about KRW 195T of revenue; memory margins estimated at 70-80% by brokerages, division detail October 29

Flywheel arc · all-three

The only lab-risk price the market set this week is the 37.5 basis point gap between the two tranches.

• Capital: the $42B launched this week prices Broadcom's partial guarantee; the $18B junior tranche is the first syndicated mark on Anthropic's own credit and has not launched. Trip wire: a junior print inside 8.5%, the June B-note level.
• PJM builders: FERC rejected the proposed new-capacity rule on October 9 at the buyer's and seller's request; buy existing firm supply and a signed service agreement before PJM's November 9 compliance filing and November 16 show-cause response.
• AI platforms: the sub-dollar tier is priced by prompt length; route by context length (The Model Pulse has the rate cards, the strategic outlook the thresholds).
• Returns landed in memory and foundry; the one interconnect pricing-power statement was an optics supplier sold out through early 2029.

What this means

Every list price that fell this week fell for a segment: Haiku 5.5's cut is 90% below 100K prompt tokens and 50% above, Sol Ultrafast sells the same output at six times list for speed, and the Decisions API meters input only in a category one vendor (TypeSafe) already served at four cents. That is compression that differs by prompt length and call type, so the volume response Jevons predicts (cheaper tokens raising total spend) will differ by call type; The Model Pulse is the fact home for the rate cards and the routing thresholds. The second lesson is about labels: two Western 'open-weight' frontier announcements landed with no downloadable artifact between them, so the repository, not the press release, belongs on the on-prem critical path. The third is the OpenAI mathematics corpus, 22% machine-checked with three withdrawals in a day; Scott Aaronson named the two release norms on display, and the measurable question is the error rate in the 560 unverified manuscripts.

• Key routers on prompt-length band and call type, not list price; The Model Pulse carries the rate cards and thresholds.
• Treat 'open-weight' as a property of a repository, not an announcement; two Western frontier models this week have neither weights nor a named licence.
• Watch the withdrawal count on OpenAI's 560 unverified manuscripts before citing the 722 figure anywhere.

Oct 5-6

Two Western 'open-weight' frontier models announced, neither downloadable: Reflection Beam (501B / 23B-active, weights 'later this month') and Mistral Large 4 (1.05T / 52B-active, API preview now, weights promised end of October under a licence not yet named)

Sources Mistral AI; Reflection AI

Oct 7

Claude Haiku 5.5 ships at $0.10 / $0.50 per million below 100K prompt tokens and $0.50 / $2.50 above; Sonnet 5.5 cache reads halved to $0.10 the same day

Sources Anthropic

Oct 6

OpenAI Decisions API public beta on gpt-6-luna at $0.10 per million input tokens, output free, into a category where TypeSafe's Jev already charged $0.042; Microsoft-Decision-1 matched Jev and a reported Cloudflare Clef-flash cut undercut both within 72 hours (The Model Pulse has the price table)

Sources OpenAI API changelog; Microsoft Foundry (vendor post, outside the graded set); Cloudflare price as reported by Dbggr and Willison, not in the graded set

Oct 6

OpenAI publishes 722 AI-generated mathematics manuscripts across 372 problem families from an unreleased model; 162 (22%) carry Lean formalizations (machine-checked proofs) and three were withdrawn within 24 hours

Sources OpenAI

Oct 8

GPT-6.1 Sol Ultrafast reaches GA at $12 / $60 per million below 272K input tokens ($24 / $90 above), six times the Standard tier, with a vendor-reported 8x speed claim

Sources OpenAI release notes

What this means

The quarter's returns landed one layer up the supply chain from the hyperscalers: Samsung's preliminary operating profit is up almost nine times on memory margins brokerages put at 70-80%, and TSMC's September is up 54.6% year on year. The lever for next year's accelerator bills is the level, not the growth rate: a reported $4.50-4.90 per gigabit HBM4 ask against about $1.50 for HBM3E, with most 2027 volume already allocated to Nvidia and Broadcom, which leaves AMD's 2027 Helios ramp (TrendForce: 'slightly delayed') gated on the residual allocation Lisa Su went to Korea to secure. Custom HBM turns that queue into lock-in: Nvidia's NVHBM moves the memory controller into the HBM base die, Micron says it carries higher ASP and margin than standard HBM, a custom stack is not fungible across buyers, and the base dies move to TSMC 12nm and 3nm, where HBM now competes with the accelerator for the foundry capacity Su says is planned three to five years out. The same week, the pre-OCP power and thermal wave (300 kW in-rack and 3.6 MW 800VDC coolant units) defined the load class PJM's new ride-through rule targets, and Flex's power unit took a $37.5B enterprise value; the power-electronics layer priced before the thesis listed it.

• Sign the 2027 DRAM long-term agreement and negotiate the band; expect 2027 accelerator quotes to carry memory as a pass-through band and ask the OEM which band it bought.
• Model Rubin memory on 12-Hi HBM4 at the reported ask as the base case; Counterpoint's 8-Hi-to-12-Hi reversal means less conventional-DRAM relief, not a cheaper bill.
• Hold vendor accelerator claims to AMD's disclosure standard this week: configuration published, or discounted.

Oct 8

Samsung guides to KRW 107.4T preliminary Q3 operating profit (+782.5% YoY) on about KRW 195T of revenue; Korean brokerages estimate memory margins of 70-80% and Device Solutions (semiconductor) division profit of KRW 107-112T

Sources Samsung Newsroom; CNBC

Oct 8

TSMC September revenue NT$511.86B, +54.6% YoY and -0.6% MoM; Q3 ~NT$1.49T, above the top of guidance; results and 2027 capex colour due October 15

Sources TSMC monthly revenue report

Oct 7-9

Samsung's 2027 HBM4 (high-bandwidth memory, the stacked memory on an accelerator) talks reported in the final stage at $4.50-4.90 per gigabit, with most 2027 volume allocated to Nvidia and Broadcom; Counterpoint reverses its August 8-Hi call to 12-Hi (stack height in dies) as the majority stack for Rubin

Sources The Korea Herald; Counterpoint Research

Oct 7

TrendForce: DRAM makers push 2027 long-term agreements with price bands and prepayments; lead times stretched to 20 weeks from a balanced 8; Micron reports more than 75% of 2027 output committed

Sources TrendForce

Oct 8

AMD publishes MI455X methodology: 3.3x MXFP4 GEMM throughput versus MI355X and 260 TB/s scale-up bandwidth per Helios rack, with the benchmark configuration disclosed; Lisa Su says capacity is planned three to five years out

Sources AMD

What this means

Plan 2027 scale-up domains for copper inside the rack and a 2027-or-later date for anything that crosses one, because the merchant optics that would let an open Ethernet fabric leave the rack are allocated, not just expensive (Lumentum's horizon is in the levers). Arista, Broadcom and Upscale each announced an ESUN fabric with copper and optical variants, every GA (general availability) is 2027, and ESUN interop is validated for the first time on October 14, so discount the open-standards claim until a hyperscaler names a production deployment; 650 Group's 2030 split (NVLink $25bn+, Ethernet $8bn+) says the open camp is still the smaller one. The long-haul side moved in the quieter way that compounds: 1.6 Tb/s per wavelength is now a national-core default at Openreach, Super C+L (extended optical bands on existing fibre) added 21% to a live Arelion route, and two sovereign fabric programmes (etisalat, Japan's Virtual Hyperscaler consortium) chose to build.

• Design 2027 scale-up domains copper-first and treat cross-rack optics as a 2028 budget line, not a 2027 one.
• Discount every 'open' scale-up Ethernet claim until a hyperscaler names a production deployment; the first interop test is October 14.
• Put optics vendors, not switch vendors, on the pricing-power watch: a sold-out horizon that moves out six months in six months is the signal.

Oct 7

Arista Etherlink SU-144: 144 XPUs (accelerators of any kind) single-hop and 1,024 cross-rack on Broadcom Tomahawk Ultra under ESUN (the Ethernet scale-up fabric standard), in orthogonal-chassis, cabled-backplane and cross-rack form factors; demos at OCP (the Open Compute Project summit) October 12-15, ship dates not stated

Sources Arista Networks

Oct 8

Broadcom pre-OCP brief: 102.4T Tomahawk 6-Davisson co-packaged optics (CPO, optics mounted on the switch package rather than in pluggable modules) demoed by five ODMs (Celestica, Delta, Micas, Accton/Edgecore, Alpha), Tomahawk Ultra and SUE-T for scale-up; ESUN 1.0 validation is on the OCP programme for October 14

Sources Broadcom (GlobeNewswire)

Oct 9

Lumentum: AI optical components sold out through early 2029 (was 2028 six months ago), ~70% of demand for some products unmet through 2027 and ~30% for others through 2028, optical circuit switching above $100M a quarter; Nvidia invested $2B across Lumentum and Coherent

Sources The Japan Times (Bloomberg)

Oct 8

Upscale AI's 115.2 Tb/s Token Fabric switch, built on its SkyFabriX silicon, supports ESUN, UALink-over-Ethernet and SUE-T from one device; GA early 2027 at a $2B valuation

Sources Upscale (via Network World)

Oct 7

Openreach deploys Ciena WaveLogic 6 at 1.6 Tb/s per wavelength across the UK core; Arelion and Nokia add 21% capacity on a live route with Super C+L; Telstra's Aura Perth-Sydney route enters service

Sources Ciena; Openreach

Frontier Labs detail

Counting rule: the chart carries closed capital raised by a lab on its own account. The $42B senior syndication is lease debt owed by an SPV, partially guaranteed by Broadcom, launched rather than closed, so it is reported in text and the ratio line stays n/a; charting it as lab capital while refusing to divide it by lab revenue would be inconsistent. OpenAI's reported $30B+ at ~$1.4T has no lead investor and is grade 2; Manus is lab-adjacent. The $18B junior tranche (Blackstone reported at ~$9B) is the instrument that marks Anthropic, and the big story and the synthesis's junior-tranche connection argue it will be an anchored private price.

Capital in value

$0B

Revenue out value

$4.6B

• Oct 5 · Bank of America, Citigroup and Morgan Stanley launch $42B of senior loans (2 x $21B, SOFR+150 / +187.5 bp, to Sep 30, 2033), partially guaranteed by Broadcom, to an SPV leasing custom chips to Anthropic; part of a $61.5B package; IFR notes the 37.5 bp spread difference may reflect different guarantee structures · $42B
• Oct 6 · FT (via Dealroom): $18B junior Class B tranche led by Blackstone (~$9B committed), expected over coming months; the order covers 15 GW of chips over two years; the FT's source ties the senior loans' investment-grade bond takeout to Broadcom's A-minus rating (agency not named); Broadcom five-year CDS at 133 bp on Oct 9 (LSEG via IFR) · $18B (junior, unlaunched)
• Oct 8 · Manus parent Butterfly Effect closes a $500M+ round co-led by Boyu Capital and IDG Capital; ~$4B valuation per Bloomberg's September report, not confirmed by the company · $500M+
• Oct 5 · OpenAI in talks for a $30B+ round at a company-set ~$1.4T pre-money; UAE funds including MGX weighing up to $10B, BlackRock in talks, no lead investor (single outlet, not charted) · $30B+ (talks)

Sources IFR · Dealroom, citing the Financial Times · TechCrunch · Bloomberg Law (Bloomberg News)

Hyperscaler-Hosted detail

Counting rule: power procurement is not compute capital, so the chart stays at zero; the generator's capital is the >$4.3B of uprate work, funded by Constellation and, in Vistra's case, a conditional DOE loan against PPAs Meta signed in January. The Google deal is 75% existing output and Meta's January PPAs are 83%; Google's follows Amazon's Calvert Cliffs PPA last week as the second large firm-supply nuclear deal in a fortnight. Google's 500-customers-above-1T-tokens figure is a token count, not a price, so it stays out of the revenue column.

Capital in value

$0B

Revenue out value

Indirect

• Oct 6 · Google and Constellation: 890 MW of nuclear uprates under 20-year PPAs plus 2,700 MW of 15-year supply from existing output; >$4.3B generator capex · Price undisclosed
• Oct 5 · DOE issues a conditional loan commitment of up to $4.2B to Vistra for 433 MW of uprates at Perry, Davis-Besse and Beaver Valley, contracted to Meta under its 20-year PPAs · Up to $4.2B (conditional)

Sources Constellation Energy · US Department of Energy

Neoclouds detail

Counting rule: the chart carries the GPU share of the two pools, because that is the only unguaranteed GPU exposure that reached a rated instrument this week. The 9.22% and 9.65% figures are the weighted implicit rates 47 small obligors pay on the underlying leases (80-85% non-GPU equipment), roughly 245-290 bp over Lambda's 6.78% coupon, and they are the nearest unguaranteed comparator rather than a clearing yield; the notes' yields are not reported. Lambda's reported equity raise is unsigned and not charted; Applied Digital's results are the week's house measurement.

Capital in value

$0.2B

Revenue out value

Indirect

• Oct 5 · Stonebriar's SCF Equipment Leasing 2026-1 (~$869M, ~15% GPU loans, settles Oct 14; 9.65% weighted-average implicit rate per the KBRA presale) and Wingspire's WEF 2026-1 ($407M, closed Sep 24, ~20% GPU; 9.22% pool yield per Wingspire); GPU shares are anonymous-sourced · ~$0.2B GPU share
• Oct 6 · Lambda raising up to $4B at a $14.5B pre-money valuation, led by Blackstone and Coatue, as a final private round before a planned 2027 IPO; backlog $50B per an LP letter (not closed, not charted) · Up to $4B (in progress)
• Oct 7 · Applied Digital FQ1 FY2027: revenue $341.9M (+322%), $3.7B cash against $6.4B debt, $1.59B of 7% notes completing Polaris Forge 1 funding; up to 1 GW of power secured in Finland (Oct 6) · $341.9M revenue

Sources Asset Securitization Report (Bloomberg); KBRA presale via StreetInsider for the rates · The Wall Street Journal · Applied Digital

On-Prem / Hybrid detail

Counting rule: the chart carries closed deployment programmes, and there were none; Flex's $2.0B Axiom raise funds a vendor, the sovereign fabric programmes are network capital (networking lens), and Mistral's 3,800-GPU training run is a disclosure, not a transaction. Nvidia's DGX Station for Windows and RTX Spark laptop ship October 16 as the first desktop marketed for trillion-parameter open models, while the two Western models marketed with them cannot yet be downloaded.

Capital in value

$0B

Revenue out value

Indirect

• Oct 5 · Flex sells $2.0B of Series A convertible preferred in Axiom, its data-centre power, thermal and compute unit, to General Catalyst and Koch Equity Development at a $37.5B enterprise value ahead of a Q1 2027 spin-off · $2.0B
• Oct 7 · Nvidia DGX Station for Windows (GB300, 748 GB) and RTX Spark 128 GB laptops available October 16 via eight OEMs; prices unannounced · Undisclosed

Sources Flex (8-K Exhibit 99.1) · Nvidia

Signal score 5/5

FERC accepted PJM's 50 MW large-load definition and registry and rejected the Interim Resource Adequacy Service, the bring-your-own-new-capacity rules and the demand-curve carve-out, effective October 12, with a 30-day compliance refile.

Primary and dispositive. The registry survives, so large loads above 50 MW will be visible to PJM in a way they were not; the proposed curtailment exposure for new load that did not bring new capacity does not, and uprates still clear the capacity auction like any other resource. Google (a buyer) and Constellation (the seller) asked for this outcome in filings on September 3 and 29. Treat the November 9 compliance filing as the next place PJM can try again, and the Rosner concurrence (the retail-jurisdiction line was the deciding one) as the map of what a redesign must avoid.

Sources

FERC order 197 FERC ¶61,019 in docket ER26-3515 (Oct 9); Commissioner Rosner's concurrence; Utility Dive

Signal score 4/5

Samsung's Q3 operating profit was KRW 107.4T, up 782.5% year on year, on memory margins of 70-80%.

The consolidated number is a company disclosure, so 4 rather than 5 only because preliminary guidance is unaudited and gives no division split. The 70-80% margin and the KRW 107-112T DS-division profit are brokerage estimates, not Samsung's; final results on October 29 settle them. For a buyer the useful inference is not the margin but the leverage it gives the seller in the HBM4 talks the next claim describes.

Sources

Samsung preliminary earnings guidance (Oct 8); CNBC; KB Securities and Meritz division estimates

Signal score 3/5

Banks launched $42B of senior loans to finance Anthropic's custom-chip leases, partially guaranteed by Broadcom and priced at SOFR+150 and +187.5 bp, as part of a $61.5B package with an $18B junior tranche.

Three outlets agree on tranche sizes, spreads, tenor and arrangers, and Broadcom's CDS move is market-observable, so the structure is credible; the spreads are launch terms, no term sheet is public and the guarantee's coverage is undisclosed, so grade 3. IFR's frame, 'the biggest test yet of market appetite for the controversial circular financing deals keeping the AI boom alive', is attributed; this issue reads the senior book as a test of Broadcom first, with the caveat that a spread consistent with Broadcom's rating cannot by itself show how much of the pricing is the guarantor and how much the collateral and lease, since the guarantee coverage is undisclosed and no Broadcom unsecured comparator at the same tenor is in the packet; IFR's own note that the 37.5 bp gap between tranches may reflect different guarantee structures says the market is pricing something beyond Broadcom. The number to wait for is the junior tranche's price, which the FT says may not come until IPO disclosures.

Sources

IFR (Oct 5); Bloomberg; Financial Times (Oct 6); all anonymous-sourced

Signal score 2/5

SpaceX is arranging a ~$40B package of $10B loans and $30B investment-grade bonds, led by Apollo, to fund xAI's Colossus 3 compute.

One original report relayed by two others, no bank mandate letter, a 2027 close, and a BBB rating that rests on Starlink rather than on xAI; the public mark that already exists is SpaceX's 2056 bonds at about 227 bp over Treasuries. It is directionally consistent with the week's pattern (compute debt priced on a sister business) and that is all it is good for until a second source or a rating agency publishes. Do not build a 2027 xAI capacity assumption on it.

Sources

Financial Times (Oct 6), relayed by Bloomberg and CNBC on Oct 7; anonymous sources; deal said to close in 2027

Signal score 2/5

Samsung's 12-Hi HBM4E has passed Nvidia's qualification for Rubin, with mass production in Q1 2027.

A single Korean daily, unconfirmed by either party, published in the same week Counterpoint reversed its stack-height view and Samsung was negotiating HBM4 pricing in its final stage. The incentive to leak a qualification pass during a price negotiation is obvious. Until Nvidia or Samsung confirms, it is a negotiating signal, the HBM lever is unchanged, and the report is not used as evidence anywhere else in this issue.

Sources

The Korea Economic Daily (Hankyung), single report; Samsung declined to comment

Signal score 1/5

OpenAI's 722 mathematics manuscripts are 'verified by Lean'.

The primary source contradicts the claim: 162 of 722 manuscripts (22.4%) carry Lean formalizations, the remaining 560 are status-unchecked, and three results were withdrawn within 24 hours. The 162 count is Stanford Tech Review's and Declic's audit of the repository; the Association for Human Mathematics' statement objects to the release on the IAS advisory group's guidance, and Scott Aaronson's caveat is that 'no human has understood just about any of these proofs yet' and only some carry Lean certificates. A reader who repeats the verified figure is repeating a frame the repository does not support. The genuinely open question is the error rate in the 560, which no one has measured yet.

Sources

Circulated on social platforms and in secondary coverage of the Oct 6 release; contradicted by the repository's own metadata

Filing-Derived

At the $146K per MW-month of base rent Applied Digital disclosed in its October 7 8-K, Polaris Forge 1's first 250 MW cover the 9.25% project notes' 2028 debt service about 1.0x (1.01x; in plain words, the rent just pays the interest and scheduled principal with almost nothing to spare) once 7.75%-per-year amortization starts in December 2027, while the 7% notes that financed the campus's last 150 MW cover about 1.5x (1.53x on the same arithmetic).

Method: From the Oct 7, 2026 8-K (Exhibit 99.1): HPC base rental revenue of $65.8M and Net Operating Income of $58.829M for the quarter ended Aug 31, 2026, earned on 100 MW live all quarter plus 75 MW live from July 1 = 450 MW-months, giving $146.2K rent and $130.7K NOI per MW-month ($1.755M and $1.569M per MW-year, 89.4% margin); the prior quarter's release (Jul 27, 2026: $44.062M rent on 100 MW x 3 months = 300 MW-months) gives $146.9K, a 0.4% difference, and the $36B base-term backlog over 1.41 GW at 15-year terms gives $1.70M per MW-year, so the per-MW reading is corroborated three ways. From the Nov 20, 2025 and Jun 16, 2026 8-Ks and the indentures they attach: $2.35B of 9.250% notes (issued at 97.0, maturing Dec 15, 2030) fund ELN-02 (100 MW) and ELN-03 (150 MW) and amortize 3.875% of original principal each June 15 and December 15 starting Dec 15, 2027 ($91.06M per payment); $1.59B of 7.000% notes (at par, maturing Jun 15, 2031) fund ELN-04 (150 MW) and amortize 2.70% per year (rising 0.50 points annually) only after the ELN-04 lease commences. Calendar-2028 debt service on the 9.25% notes is interest on the declining balance ($2,258.9M x 9.25%/2 = $104.5M in June; $2,167.9M x 9.25%/2 = $100.3M in December) plus two $91.06M installments = $386.9M, against 250 MW x $1.569M = $392.2M of NOI, a 1.01x ratio (1.03x using the prior quarter's 91% margin); first-year debt service on the 7% notes is $1,590M x (7.00% + 2.70%) = $154.2M against 150 MW x $1.569M = $235.3M, a 1.53x ratio.

Implication: Coverage of 1.01x from 2028 and a $1.80B balloon in December 2030 make the first-generation 9.25% structure a refinancing bet, not a repayment plan: interest alone is comfortably covered (1.8x to 2.1x), but once 7.75% annual amortization starts the rent just meets the schedule, so the equity case for the campus rests on rolling that tranche into the 7%-style terms the same borrower obtained seven months later; nothing has been refinanced yet. Developers and their boards should negotiate the amortization profile as hard as the coupon; investors reading 'fully funded' should net the roughly 14% of face that sits in reserve and interest accounts before computing cost per MW. For a platform engineer whose capacity sits in a leased campus like this one, the thin coverage is a reason to know the lessor's refinancing calendar as well as the tenant's SLA: a 2030 balloon on the landlord's side is a continuity risk on yours.

Caveats: Rent is GAAP straight-line revenue, so cash rent in early years may be lower (or higher) than $146K per MW-month if the leases carry escalators or free-rent periods, and NOI excludes tenant-recovery pass-throughs, fit-out margin, capital expenditure and parent G&A, all of which the indenture waterfall treats differently; the coverage ratios here are house computations, not indenture tests, because neither indenture contains a debt-service-coverage covenant, and the 1.01x figure assumes no voluntary redemption, no excess-cash-flow repurchases and that ELN-04's lease commencement date (not yet disclosed) does not shift the 7% tranche's amortization start.

Disclosed base rent, Polaris Forge 1 (GAAP straight-line)

$146.2K per MW-month = $1.755M per MW-year — Prior quarter on 100 MW alone: $146.9K per MW-month; backlog-implied portfolio rent ($36B / 1,410 MW / 15 years): $1.70M per MW-year; Delta Forge 2 lease ($5.2B / 210 MW / 15 years): $1.65M per MW-year

First-lien project debt per MW on the 400 MW campus

$9.85M per MW ($3.94B of notes / 400 MW) — 9.25% tranche: $9.40M per MW ($2.35B / 250 MW); 7% tranche: $10.60M per MW ($1.59B / 150 MW); campus NOI-to-debt yield 15.9% ($627.5M / $3.94B)

Debt service coverage, 9.25% notes (ELN-02 + ELN-03, 250 MW), calendar 2028

1.01x ($392.2M NOI / $386.9M interest plus scheduled amortization) — 1.80x on interest alone at the original face ($392.2M / $217.4M; 1.92x on the 2028 declining-balance interest of $204.8M); 1.27x in 2027 when only one amortizing payment falls; 0.99x over the first twelve months of amortizing payments (Dec 2027 + Jun 2028 = $395.3M)

Debt service coverage, 7% notes (ELN-04, 150 MW), first amortizing year

1.53x ($235.3M NOI / $154.2M) — 2.11x on interest alone; the June 2026 deal cut the coupon 225 bp, moved the issue price from 97.0 to par, and cut first-year amortization from 7.75% to 2.70% of face versus the November 2025 deal on the same campus

Note proceeds parked to pay noteholders rather than build

$565M = 14.3% of the $3.94B face — 9.25% notes: $177M debt-service reserve + $178M interest-during-construction account = $355M (15.6% of the $2,279.5M received at 97.0; IDC covers 9.8 months of coupon); 7% notes: $81M reserve + $129M IDC = $210M (13.2% of face; IDC covers 13.9 months of coupon)

Principal still due when the 9.25% notes mature on Dec 15, 2030

$1.80B = 76.8% of original face — Six scheduled installments of $91.06M retire $546M by June 2030; the indenture's 50% excess-cash-flow repurchase offer adds little at roughly 1.0x coverage because excess cash flow is defined net of debt service, reserve top-ups and operating expenses

Sources Applied Digital Form 8-K, Exhibit 99.1, fiscal Q1 2027 results (Oct 7, 2026) · Applied Digital Form 8-K, Exhibit 99.1, fiscal Q4 2026 results (Jul 27, 2026) · Applied Digital Form 8-K, 9.250% Senior Secured Notes due 2030 (Nov 20, 2025), Item 1.01 and Exhibit 4.1 indenture · Applied Digital Form 8-K, 7.000% Senior Secured Notes due 2031 (Jun 16, 2026), Item 1.01 and Exhibit 4.1 indenture

Abductive · 66% confidence

No syndicated or public market has yet marked unguaranteed frontier-lab credit; the one print that exists is private and four months old ($4.5bn of B notes at a reported 8.5% in the June package, against 5.75% on its Broadcom-supported A2 notes). This week's senior loans, partially guaranteed by Broadcom at launch spreads of SOFR+150-187.5bp, price the guarantor first, and the only lab-risk price the market set is the 37.5bp gap between the two tranches, which IFR says may reflect different guarantee structures. The $18bn junior tranche, four times the June B notes, subordinated to $42bn of senior claims on the same chips and sold after Broadcom's CDS tripled, will price at or above the June level of 8.5% when it is marked, and as an anchored private price rather than a public one, because Blackstone already holds ~$9bn of it. Resolution: the junior tranche's price, expected within months and possibly after the IPO prospectus; a print inside 8.5% breaks the claim.

Steel-man: Three fair objections. First, vendor finance has funded aircraft, telecoms and machinery for decades and circularity does not make it improper (electronicsbrief); a guarantor's CDS can widen for reasons unrelated to the lessee, Broadcom's 'tripled since late May' predates the October 1 convertible disclosure by four months, and Broadcom also carries acquisition debt and most of Samsung's 2027 HBM4 allocation alongside Nvidia, so the 133bp may price Broadcom's own balance sheet and inventory commitment rather than Anthropic's default risk; this chain does not attribute the widening to any one cause. Second, the June print cuts both ways: it shows an unguaranteed Anthropic tranche cleared at 8.5% in a smaller, privately rated placement, and a buyer who has since seen Anthropic's prospectus, a $15bn revolver and an IPO in motion could take the larger tranche tighter, especially one anchored by Blackstone, which already owns the adjacent Lambda exposure and is not pricing from a standing start. Third, the 37.5bp inter-tranche gap may be tenor or structure rather than lab risk, in which case the market set no lab-risk price at all this week. The claim survives in bounded form: the only syndicated mark in prospect is the junior tranche, the 8.5% threshold is now anchored on a known print rather than a round number, and a Blackstone-anchored price is a weaker market test than a syndicated one. If the junior prints inside 8.5%, or is upsized and launched before the IPO disclosures, the market is saying frontier-lab credit has improved since June.

• capital-01: banks began syndicating $42bn of senior secured equipment-lease loans for Anthropic's 15 GW custom-chip order at launch spreads of SOFR+150bp and +187.5bp to 2033, both tranches partially guaranteed by Broadcom (coverage undisclosed; IFR notes the spread difference may reflect different guarantee structures), while the $18bn junior tranche that carries Anthropic's own risk is deferred and Broadcom's 5-year CDS has more than tripled since late May to 133bp on October 9 (LSEG via IFR).
• The prior print, outside the graded set: the June 2026 $35bn Apollo-Blackstone chip package for the same lab with the same guarantor priced $6bn of A1 notes at Treasuries+100bp and $24bn of A2 notes at 5.75% with Broadcom residual-value support, and $4.5bn of B notes with no Broadcom backing at 8.5% at par, privately rated (PitchBook, June 2; The Inference, October 5, which framed the 275bp difference as the market price of Broadcom's signature). That is a private, first-loss mark on Anthropic credit, so 'never priced' is false and 'never syndicated or publicly marked' is what remains.
• capital-04 and capital-02: Blackstone is the anchor on both sides of the same credit, with ~$9bn committed to the junior tranche and the lead on Lambda's up-to-$4bn equity round, where the $50bn backlog is roughly $35bn of Anthropic commitment, so one private-credit shop is marking Anthropic directly (junior) and indirectly (neocloud equity 70% dependent on the same lab) in the same fortnight; Lambda's 6.78% investment-grade term loan is secured on that backlog. SpaceX's ~$40bn package is in talks rated off Starlink cash at BBB (FT, relayed by Bloomberg and CNBC), with its 2056 bonds at ~227bp over Treasuries, which with Broadcom's 133bp CDS makes the two guarantor-tier marks the market has already set; Nvidia holds SpaceX at 33% of its equity book on one analyst's 13F read (Cape Fear). In every case the paper is priced on an operator, a tenant lease or a supplier, and once, in June, on the lab itself in a private placement.
• The reading that fits is that the senior book tests Broadcom, that the junior tranche is the first instrument where Anthropic's standalone credit could be marked at syndicated scale, and that its price will not be tighter than the June B notes: it is four times the size, subordinated to $42bn of senior claims on the same collateral, and sold into a market that has since watched the guarantor's CDS triple, whatever the cause of that widening. The equipment-ABS implicit rates (9.22-9.65%) and Lambda's 6.78% are context for the neocloud row, not inputs to this level.

Sources capital-01 (IFR, $42bn syndication partially guaranteed by Broadcom; CDS 133bp Oct 9) · June 2026 $35bn Apollo-Blackstone package: $4.5bn unguaranteed B notes at 8.5% (PitchBook, June 2; The Inference, Oct 5; outside the graded set) · capital-11 (GPU loans in rated equipment ABS, 9.22-9.65% weighted implicit lease rates; context) · capital-04 (Lambda $4bn raise led by Blackstone and Coatue; 6.78% IG term loan; $35bn Anthropic backlog) · capital-02 (SpaceX ~$40bn Apollo-led GPU debt talks, BBB; 2056 bonds ~227bp over Treasuries) · Cape Fear Advisors filed-fact audit (Nvidia 13F: SpaceX 33.1% of equity book)

Abductive · 61% confidence

The week's one new nuclear deal (Google-Constellation, 3,590 MW) is, at the capacity level, a purchase of the existing fleet with a minority uprate attached, and Meta's January Vistra PPAs, now backed by a conditional DOE loan announced October 5, have the same shape; the federal rule that would have exposed new large load to curtailment unless it brought new capacity was rejected the same week at the request of the buyer (Google) and the seller (Constellation). The explanation consistent with Hypothesis 5 is that a binding power constraint shows up in contracts as a bid for firm supply that already exists plus resistance to any rule that conditions service on supply that does not. It follows that PJM's compliance filing (due by about November 9) will not re-propose bring-your-own-new-capacity, that the live venue for a redesign is PJM's November 16 Section 205 show-cause response in EL26-67, and, more narrowly than the jurisdictional point, that ISO-NE's November 16 filing will have to make the undue-discrimination showing FERC found PJM had not made for its BYONC Firm Contract Demand Service. Resolution: PJM's filings by November 30 (p130) and FERC's eventual order on ISO-NE's filing, specifically whether it rejects or requires removal of the BYONC condition.

Steel-man: The strongest counter is that FERC rejected PJM's package on jurisdictional and procedural grounds, not on the principle of BYONC, and offered guidance for a refile; ISO-NE's design is different in kind (a condition on FERC-jurisdictional transmission service with a 7-year interim tier and a $250/MWh non-firm re-dispatch price, rather than curtailment of retail load), and ISO-NE's transmission owners (PTOs) have drafted a Cost Recovery Agreement precisely to keep the retail boundary clean, so the retail-jurisdiction finding does not transfer and only the undue-discrimination showing does, which is why the claim is now narrowed to that showing and the confidence is in the low 60s. It is also fair that 890 MW and 433 MW of firm uprates are real new megawatts that would not exist without these PPAs, and that keeping 4.9 GW of existing nuclear in the market for 15-20 years is itself a resource-adequacy contribution. The claim survives in the narrowed form: the capacity arithmetic is a fact, Google's and Constellation's written opposition to the curtailment-exposure rule is a fact, and Hypothesis 5 is one explanation that fits both; what remains a projection is FERC's treatment of ISO-NE's BYONC condition, which is the resolution point. If PJM re-proposes BYONC in the November 16 show-cause response or any other filing in a form FERC accepts, this reading of the order was wrong.

• capital-05 and capital-06: Google's 3,590 MW Constellation agreement is 2,700 MW of 15-year supply from output the PJM fleet already produces plus 890 MW of uprates delivered 2028-2032 (The Inference gives the 2032 end date; Constellation's release says the first uprate by 2028) for more than $4.3bn of generator capex; Meta's Vistra PPAs, signed in January, cover 2,176 MW of existing output plus 433 MW of uprates that DOE would finance with a conditional $4.2bn loan announced October 5. New capacity is 25% and 17% of the respective contracted megawatts, and the uprates still clear PJM's capacity auction like any other resource.
• policy-01: FERC on October 9 accepted PJM's Large Load Registry and rejected the Interim Resource Adequacy Service (the curtailment-exposure proposal filed August 13), the Bring Your Own New Capacity eligibility rules and the plan to carve new large load out of the capacity demand curve, finding the package intruded on retail jurisdiction and was not shown to be non-discriminatory; Google (September 3 protest: 'a gamble, not an investment'), Constellation (September 29 answer) and the Market Monitor each asked for rejection on different grounds.
• Hypothesis 5 (power is the binding constraint; it shows up as PPAs and time-to-energization) predicts that a buyer facing 36-48 month interconnection waits in the busiest markets will pay for firm supply that exists today and resist any rule that conditions service on supply that does not; the order's own reasoning, that PJM had not shown the curtailment exposure to be non-discriminatory and that curtailment of retail load is a state matter, is the ground on which the buyers argued, and the deal and the protests are the prediction observed in the same docket week. The alternative readings are stated in the thesis test: 24/7 carbon-free matching and price hedging explain a purchase of existing output as well as scarcity does.
• policy-04 and policy-02: ISO-NE's October 9 redlines condition Firm Contract Demand Service on bring-your-own-new-capacity with a 7-year interim service, MISO's LARS pairs loads above 200 MW with new generation, and DOE is pressing PJM to refile its 6.8 GW backstop by October 29. PJM itself owes a Section 205 show-cause response in EL26-67 on November 16 (advisory vote at the October 28 Members Committee; the ride-through rules go into it), which is the most likely venue for a BYONC re-proposal. With the ER26-3515 route rejected, the routes to tie new supply to new load are that show-cause filing, a centrally procured backstop, state PUC curtailment rules (Commissioner Rosner's concurrence) and DOE's loan book.

Sources policy-01 (FERC order 197 FERC 61,019, ER26-3515) · capital-05 (Google-Constellation 890 MW uprates + 2,700 MW existing supply) · capital-06 (DOE $4.2bn conditional loan to Vistra for 433 MW of uprates) · policy-04 (ISO-NE large-load redlines with BYONC condition, Oct 9) · The Duck Curve on Google's Sept 3 protest and Constellation's Sept 29 answer · Headwaters Wire on the FERC split order

Abductive · 64% confidence

The scale-up Ethernet wave (ESUN) shown ahead of OCP will ship rack-bounded through 2027, on copper or pluggable optics, with the cross-rack co-packaged and near-packaged optical claims (6.5 Pb/s per rack, 1,024-XPU domains, NPO backplanes) staying demos; the best explanation is that the high-end optical component layer is the one input the merchant-fabric camp does not control, and its leading supplier says new capacity is allocated through early 2029 while shipping record volume. The constraint is allocation, not growth: Lumentum's record transceiver and laser-chip shipments and the start of 1.6T production are demand exceeding supply, which is not Gilder's law's falsifier (bandwidth growth stalling while compute doubles), so the falsifier is not scored this week. Resolution (p134): by June 30, 2027 no vendor or hyperscaler discloses a shipping or in-service ESUN scale-up domain larger than one rack (above 144 XPUs) running on CPO/NPO optics; a disclosure breaks the claim.

Steel-man: Lumentum is one supplier, and its sold-out statement is about its own components; Coherent, InnoLight, Eoptolink and Broadcom's in-house optical engines are not in that sentence, CPO reduces the number of discrete transceivers per unit of bandwidth, and Lumentum's own optical-circuit-switch business passing $100m a quarter is a technology that removes transceivers from the path entirely. NVIDIA was also already shipping single-vendor CPO Ethernet switches in production in August (W34), so 'CPO stays a demo' is false at the proprietary end. There is also a timing null: ESUN products announced as Summit demos with GA (general availability) no earlier than 2027 would stay rack-bounded through 2027 on their own roadmaps even if optics were plentiful, so p134 resolving as stated would not by itself prove the optical mechanism; only a vendor citing optics availability, or a multi-rack domain shipping on pluggables but not CPO, would separate the two. The claim survives in bounded form because it is about the merchant ESUN camp specifically and the component class Lumentum describes as scarcest (lasers and high-end photonics with 3-5 year capacity lead times) is the one CPO/NPO is inferred to consume most. If a vendor or hyperscaler discloses a multi-rack ESUN domain shipping on CPO or NPO optics before mid-2027, the optical constraint was softer than Lumentum's guidance implies and Gilder's headroom reading stands.

• networking-01, networking-02, networking-03, networking-10: Arista's SU-144 (144 XPUs single-hop, 1,024 cross-rack, 1.6 Pb/s per rack on OSFP rising to 6.5 Pb/s on XPO/CPO/NPO), Upscale's 115.2T SkyFabriX (GA early 2027), Broadcom's TH6-Davisson 102.4T CPO at five ODMs and Credo's micro-emitter NPO backplane all arrived as pre-OCP demos; 650 Group counts networking at 30% of pre-show releases against 20% in 2025, and no production XPU ships on an ESUN fabric this week.
• networking-04: Lumentum's CEO said its AI optical components are completely sold out through early 2029 (2028 six months ago), that about 70% of demand for some products is unmet through 2027 and about 30% for others through 2028, that new capacity takes three to five years, that transceiver and laser-chip shipments hit records and 1.6T production has started, and that Nvidia put $2bn into Lumentum and Coherent earlier this year; one vendor datapoint, about its own products. The inference here, not his statement, is that the external laser sources TH6-Davisson's CPO design relies on fall in the component class he describes as scarcest (Lumentum sells ELSFP modules), while pluggable optics draw on a broader supplier base and are consistent with the rack-bounded path.
• Gilder's law in the framework holds that bandwidth is delivered ahead of compute, so 2026 architecture should plan around fabric headroom; its stated falsifier is optical packaging or fiber supply stalling while compute keeps doubling. hardware-05 has AMD's MI455X at 3.3x the MXFP4 GEMM of MI355X (vendor-published methodology) and Helios at 260 TB/s of scale-up bandwidth on UALoE72, so compute is doubling on schedule while the high-end merchant optics that would let that bandwidth leave the rack are allocated to 2029. Record shipments mean bandwidth supply is growing, so this is a queue, not a stall; what it changes is who gets the cross-rack optics first, and Nvidia bought a place in that queue with $2bn of equity.
• The reading that fits is that the first production ESUN deployments will use the in-rack paths the same vendors announced alongside the optics (Arista's orthogonal chassis and cabled backplane, which SiliconANGLE reports Arista describing as copper-based; Marvell's 200G/300G active copper and 1.6T AECs; Credo's ZeroFlap AECs) or pluggable optics, which bounds the scale-up domain at the rack, where NVLink already lives, and defers the 'open alternative to proprietary compute islands' pitch to whenever optics allocation loosens.

Sources networking-04 (Lumentum sold out through early 2029; ~70% unmet demand) · networking-01 (Arista Etherlink SU-144, 1.6 to 6.5 Pb/s per rack) · networking-03 (Broadcom TH6-Davisson CPO at five ODMs) · networking-10 (Marvell/Credo/Ciena/Astera pre-OCP; 650 Group 30% networking share) · hardware-05 (AMD MI455X 3.3x GEMM; Helios 260 TB/s scale-up)

Abductive · 63% confidence

This week's sub-dollar pricing compressed unevenly by prompt length: Haiku 5.5's cut is 90% below 100K prompt tokens and 50% above, so under Jevons' law the volume response should concentrate in the short-context classify/route/approve call, where one vendor (TypeSafe Jev) already served at four cents per million input tokens before the week began and two more (Microsoft-Decision-1 at $0.042, per Microsoft's post; Cloudflare's Clef-flash reported at $0.038) joined within 72 hours, leaving OpenAI's $0.10 Decisions beta the high-priced entrant; the long-context agentic bill moved by Anthropic's own estimated 20% (the Sonnet cache-read halving), not by 90%. Two resolution points: OpenAI's Decisions API GA price at or below $0.05 per million input tokens (p133, which this issue puts at 41%, so the claim leans on the second test), and Artificial Analysis's first tiered cost-per-task figure for Haiku 5.5 landing above GPT-6 Luna's on agentic tasks. Both occurring confirms; either failing weakens; both failing breaks it.

Steel-man: The honest counter is that Haiku 4.5 could not find a use at $1/$5 and Haiku 5.5 is different, because Terminal-Bench went from 0% to 33-39% and Anthropic shipped it as a subagent and compaction model, so it will be used inside long-context agent loops regardless of the tier step; Zvi Mowshowitz's practical advice, to set autocompact at 100K so the loop never crosses the boundary, is the operator's version of the same point. Anthropic's own launch page says the low tier covers 'around 90% of requests to our previous Haiku model' (footnote 2), which, if it carries over to 5.5's workload, means the 90% cut is the typical experience and the segmentation reading is a footnote about rate cards. The claim survives on an assumption stated rather than proven: that the 90% figure describes the workload Haiku 4.5 had (short classification and extraction) rather than the subagent roles 5.5 is sold into, where context grows and the tokenizer inflation brings the 5x step 20-23% sooner. The one independent figure so far points the other way: Artificial Analysis's provisional, untiered estimate is about $0.21 per task on its agentic index, relayed by Capital and Compute, with ~162k output tokens per task, about three times Luna's, so the first tiered figure could land either side of Luna. The segmentation-by-length observation is Willison's and Latent Space's; the addition here is the cross-vendor segmentation (Haiku's low tier against the four-cent decision models), the Jevons-per-segment inference and the two resolution tests.

• software-02 and software-03: Haiku 5.5 is $0.10/$0.50 up to 100K prompt tokens and $0.50/$2.50 above, against Haiku 4.5's flat $1/$5, so the cut is 90% below the boundary and 50% above it, on a tokenizer Anthropic says counts ~30% more tokens for the same text, so the boundary arrives 20-23% sooner in the same conversation (at roughly 77K old-tokenizer tokens); the same day Anthropic halved Sonnet 5.5 cache reads to $0.10 and estimated that cut, not Haiku, lowers cost on most agentic tasks by about 20%, because cache reads dominate agentic consumption.
• software-07 and software-09, with prices from vendor pages reported via Willison and Dbggr: OpenAI's Decisions API beta charges $0.10 per million input tokens with no output charge and answers ~10x faster than the Responses API; TypeSafe's Jev was already hosted at $0.042, Microsoft-Decision-1 launched at $0.042 within 72 hours (per Microsoft's post, outside the graded set), Cloudflare's Clef-flash is reported at $0.038, Perplexity's Apache-2.0 pplx-decider v1.1 claims a Decision Index 3.5 points above Jev, and LM Studio shipped a local /v1/decisions endpoint; the open-weight baseline is a 27B Qwen fine-tune. The Model Pulse carries the full rate card.
• Jevons' law (lower unit cost raises total consumption) is applied here per segment rather than per model, which is an extension of the framework's law rather than its stated form: the cut fell 90% below 100K and 50% above, both real compression but at different depths, so the 90-180-day volume response the law predicts should concentrate where the deeper cut landed, the calls the decision-model category and Haiku's sub-100K tier compete for, while long-context agent loops see a 20% cache cut and a 5x step at the boundary.
• agents-02: Codex Guardian V2 added opt-in Decisions-comparison telemetry recording agreement and latency against the approval reviewer, which is the first harness measuring whether the approve/flag/block call can move to the four-cent tier; if it can, the volume response lands on the tier with the lowest HBM and output-token intensity per call, which is the segment where the Jevons response is cheapest to serve.

Sources software-02 (Claude Haiku 5.5 tiered pricing, tokenizer note) · software-03 (Sonnet 5.5 cache-read halved; ~20% agentic saving) · software-07 (OpenAI Decisions API public beta, $0.10 input-only) · software-09 (pplx-decider v1.1, Apache 2.0) · Microsoft-Decision-1 at $0.042 (Oct 9) · agents-02 (Codex 0.162.0: Guardian V2 Decisions-comparison telemetry)

Hypothesis 1 · Strained

The cycle is accelerating, not slowing: each flywheel turn is faster than the last across software, hardware and networking.

All five hypotheses are strained this week, and the common cause is that the physical supply chain (memory, foundry, optics, power) is where both the returns and the constraints landed; each test below says what that does to its clause. Premise here: doubling cadence should be shortening on all three arcs, and the test is a measured interval, not a count of launches. Software evidence is dense: Haiku 5.5 arrived one week after Sonnet 5.5, GPT-6.1 Sol's Ultrafast tier went GA (generally available) about ten days after the model itself, Reflection reports pretraining a 501B mixture-of-experts model (MoE, where only a fraction of parameters run per token) on 23.8T tokens in under four weeks on 6,144 GB300 GPUs (vendor-reported, weights and report pending), and Mistral, Google, Qwen and Kandinsky shipped in the same five days; The Model Pulse has the release table. Hardware supplies one Huang's-law data point, AMD's own MI455X methodology at 3.3x MXFP4 GEMM over MI355X (vendor-published), above the <1.5x flattening threshold. But no doubling interval was measured on any arc this week, and the physical arcs are being scheduled further out: AMD's Lisa Su said capacity planning has stretched from 1-2 years to 3-5 years, Lumentum's sold-out horizon moved from 2028 to early 2029 in six months, TrendForce calls the MI450 volume ramp slightly delayed, DRAM lead times sit at 20 weeks against an 8-week balance, and TSMC posted a sequential decline (-0.6% MoM in September) from a record August. Conclusion: strained, down from supported in W40; the software arc is turning faster while the physical arcs lengthen, and 'accelerating across all three' is not what the week shows.

Counter-evidence: The strongest case for the hypothesis is that the lengthening horizons are themselves an acceleration signal: capacity is being planned 3-5 years out because demand is outrunning supply at every physical layer, Lumentum is shipping record transceiver and laser-chip volume and starting 1.6T production while sold out, TSMC's September is -0.6% from a record and +54.6% year on year, and the one hardware cadence measure that printed (MI455X at 3.3x) is above the flattening threshold. On that reading the cycle is accelerating and the constraint is where the doubling shows up, which is what the hypothesis says a faster turn looks like from the supply side.

Sources software-02 (Haiku 5.5 one week after Sonnet 5.5) · software-04 (Beam: 23.8T tokens in under four weeks on 6,144 GB300 NVL72) · hardware-05 (MI455X 3.3x MXFP4 GEMM vs MI355X) · software-08 (GPT-6.1 Sol Ultrafast GA, Oct 8)

Hypothesis 2 · Strained

Capital is concentrated, returns are diffuse: capex pools in a few balance sheets while returns show up across enterprise software and on-prem.

Premise: the capex bet pays off somewhere other than where it is spent. The concentration half is overwhelming: $42bn syndicated for one lab's chips, ~$40bn sought for one other lab's GPUs, a $30bn+ round at $1.4tn with no lead, a $4bn neocloud raise whose backlog is 70% one customer. But the returns half had its clearest contrary week: the returns that printed accrued one layer up the supply chain (Samsung's KRW 107tn quarter at an estimated 70-80% memory operating margin, TSMC +54.6% YoY), while the enterprise-software layer the hypothesis names as the beneficiary was paying the labs (Atlassian's 'effectively a spend commitment' to OpenAI), cutting staff (HubSpot ~7%, which its 8-K says is not driven by AI-related efficiencies, alongside an outcome-pricing pivot; FICO 15%, citing AI-driven product development) and tearing up seat pricing for per-action meters (SAP 0.02 AI Units per action, Workday's agent toll from February). Conclusion: strained, as in every week since W36 in which the hypothesis was tested as written (W36, W37, W40, W41); the proposed restatement, to be put to the thesis record next issue, is that returns are accruing in memory, foundry, optics and power electronics rather than in enterprise software, where the hypothesis placed them.

Counter-evidence: On the other side, the diffuse-returns half did get two supporting points: Google says nearly 500 Cloud customers have each processed over a trillion tokens and will not charge a per-seat fee for coworker agents (applications-01), and Abridge and Suki extended ambient AI into system-wide health procurement (applications-08, -09), which is returns accruing in a vertical far from the capex.

Sources capital-09 (Samsung Q3 operating profit KRW 107.4tn, +782.5% YoY) · capital-01 ($61.5bn Anthropic chip financing package) · applications-03 (Atlassian-OpenAI 'effectively a spend commitment') · applications-05 (HubSpot 8-K: ~7% cut, 'not driven by AI-related efficiencies', alongside outcome-pricing pivot; FICO 15%) · applications-04 (SAP retires per-user Joule Premium; 0.02 AI Units per action)

Hypothesis 3 · Strained

Networking is the durable layer: models commoditize, chips margin-compress, interconnect holds pricing power longest.

Premise: Metcalfe's law lets the interconnect layer hold pricing power after models and chips compress. The interconnect clause had its most direct supporting statement yet: Lumentum is sold out of AI optical components through early 2029, cannot meet ~70% of demand for some products through 2027, says capacity takes 3-5 years to build, and has an optical-circuit-switch line past $100m a quarter; Ciena moved 1.6T coherent from trial to a national wholesale service at Openreach; Upscale carries a $2bn valuation on silicon that ships in 2027; and 650 Group counts networking at 30% of OCP pre-show releases against 20% last year. The 'models commoditize' clause also got substitutability evidence: Google's Gemini agent preview routes some coding tasks to Claude and Atlassian's gateway abstracts the model behind its agents. But the hypothesis is a conjunction, and its middle clause, 'chips margin-compress', was contradicted outright this week: Samsung's memory division is estimated by brokers at a 70-80% operating margin, 2027 HBM4 is being negotiated at $4.50-4.90/Gb against ~$1.50 for HBM3E, Micron has >75% of 2027 output committed, and TrendForce says DRAM scarcity is now constraining enterprise SSD supply; the layer with the most pricing power was memory, not interconnect. The ESUN wave itself is a commoditization push aimed at the proprietary scale-up fabric, and 650 Group's 2030 split (NVLink $25bn+, Ethernet $8bn+) says the open camp expects to win share by lowering, not holding, price; no colo cross-connect or fabric revenue line printed in-window, and Lumentum's statement is one supplier about its own products. Conclusion: strained, moving from untested in W40; the interconnect clause is supported on the supply-scarcity side only, since the framework's own falsifier (interconnect revenue growth versus compute revenue growth at operators) did not run, and the chip clause is wrong on the week's evidence.

Counter-evidence: The strongest case for the hypothesis is on duration rather than this week's margin ranking: memory's pricing power is contracted through 2027 on LTAs (long-term agreements), while Lumentum's is allocated through early 2029 on components with 3-5 year capacity lead times, so of the two scarce layers the interconnect one has the longer horizon, which is the clause's actual claim ('longest'), and Nvidia paying $2bn for a place in the optics queue is a compute vendor buying interconnect supply, not the reverse. The 'models commoditize' clause had its clearest week of substitution evidence, and the chip clause may yet hold on logic rather than memory, which the hypothesis does not separate.

Sources networking-04 (Lumentum sold out through early 2029; OCS >$100m/quarter) · networking-05 (Openreach deploys Ciena WaveLogic 6 1.6T for UK wholesale 800G) · networking-02 (Upscale Token Fabric, $2bn valuation, GA early 2027) · networking-10 (650 Group: networking 30% of OCP pre-show releases vs 20%) · applications-01 (Gemini agent routes across Gemini and Claude model families) · applications-03 (Atlassian: Astra not Rovo's default; gateway routes across providers on capability/cost)

Hypothesis 4 · Strained

Open weights pull the floor up: open models re-route compute demand to on-prem and sovereign deployment without reducing it.

Premise: open-weight releases narrow the gap to closed models and act as a demand catalyst for on-prem and sovereign GPU consumption; the refutation metric is a gap above 10 index points for two consecutive quarters, which W40 recorded as tripped in the first quarter of the test. The demand-routing half got real support: Mistral trained Large 4 on 3,800 Grace Blackwell GPUs in its own European data centres, NVIDIA and eight OEMs put GB300-class and 128GB-unified-memory local-AI hardware on sale (DGX Station for Windows, RTX Spark from October 16), EmbeddingGemma 2 runs multimodal in 567MB on a phone, and two sovereign fabric programmes surfaced. The capability half failed again: the two Western 'open-weight' frontier models cannot be downloaded (Mistral's weights due about October 27-31 under a licence it has not yet named, Reflection's Apache 2.0 weights 'later this month' with revised benchmark tables), Large 4 costs $1.13 per index task against $0.07 for GPT-6 Luna at the same score of 38 (Artificial Analysis cost-per-task, provisional), and Reflection trails a Chinese rival on the one agentic-coding row (DeepSWE v1.1) its own table shares across all four models. Conclusion: strained; the catalyst exists, the floor it is pulling up is the Chinese one that was already downloadable.

Counter-evidence: The strongest point in the hypothesis's favour is that the on-prem hardware floor is being productized at scale (eight OEMs, a Windows-native GB300 deskside system) in the same week the local serving stack shipped 717 commits to vLLM, which is the demand-routing mechanism working even without a new downloadable Western frontier checkpoint.

Sources software-01 (Mistral Large 4 preview; weights due about Oct 27-31, licence unnamed; AA index 38) · software-04 (Reflection Beam, weights pending, tables revised Oct 8) · hardware-04 (DGX Station for Windows; RTX Spark 128GB unified memory) · Marc Pope: two open-weight models you cannot download

Hypothesis 5 · Strained

Power is the binding constraint for the next 24 months, showing up as PPAs, behind-the-meter generation and interconnection queues.

Premise: a binding constraint reveals itself when buyers pay for what already exists and regulators run out of mechanisms. The power half of the hypothesis did both this week. Google contracted 2,700 MW of existing nuclear output alongside 890 MW of uprates that Constellation will fund with more than $4.3bn of capex for delivery 2028-2032 (buying existing output is also consistent with 24/7 carbon-free matching and price hedging, so the scarcity reading is one of three); Meta's January uprates gained a conditional $4.2bn DOE loan on October 5; FERC rejected PJM's proposal to expose new large load to curtailment unless it brought new capacity; DOE made its first FERC statement of position in years to press for a 6.8 GW backstop refile by October 29; Texas's SB 6 rule took effect into a governor's pause under which a ~474 GW queue (90% data centres) has zero projects processable until a December 10 audit; Applied Digital took a 1 GW Finland campus with initial power in 2028, siting around the US queue; and PJM is writing ride-through rules because 3,800 MW of data-centre load dropped to backup power in one July event, which with the pre-OCP 800 VDC and liquid-cooling wave puts power electronics and thermal inside the constraint. But the hypothesis as written claims power is the binding constraint, not chips, capital or land, and components bind on the same evidence standard: DRAM lead times are 20 weeks against an 8-week balance and suppliers have moved most server DRAM onto multi-year price-band LTAs (buyers paying for visibility on existing supply, the same signature as the nuclear deals); 2027 HBM4 is being settled at more than 3x HBM3E; Lumentum is allocated through early 2029 with 3-5 year capacity lead times, longer than the hypothesis's 24-month window; AMD's Helios ramp is gated on residual HBM4 allocation; and last week's read that Rubin moves to 8-Hi HBM4 and frees DRAM wafers was undercut by Counterpoint's October 9 reversal toward 12-Hi. Conclusion: strained, unchanged from W40; the power evidence is the strongest of the year, the exclusivity clause is what the week refutes, and the proposed restatement, to be put to the thesis record next issue, is 'power and components are the binding constraints', with interconnection waits restated as 36-48 months in the busiest markets.

Counter-evidence: The strongest case for the hypothesis as written is that component scarcity clears through price and power does not: HBM4 at 3x and DRAM on LTAs are allocation problems that money resolves inside the 24-month window, whereas a 474 GW queue frozen pending an audit, a 36-48 month interconnection wait and a FERC order that leaves no federal mechanism to condition new load on new supply are constraints no bid can clear on that timescale. On that reading power is still the only constraint that is binding in the hypothesis's sense of the word, and components are expensive rather than binding.

Sources policy-01 (FERC rejects PJM IRAS curtailment exposure and BYONC, accepts registry) · policy-02 (DOE intervention in ER26-3380; Oct 29 refile target) · policy-03 (Texas 16 TAC 25.194 effective Oct 8; Abbott pause; ~474 GW queue) · capital-05 (Google-Constellation: 2,700 MW existing + 890 MW uprates) · policy-04 (PJM ride-through rules after 3,800 MW July 22 event; ISO-NE and MISO BYONC designs)

Inductive · 6 weeks observed

Agent-model list prices fall or converge while the effective bill moves through non-list meters: speed tiers, context breakpoints, tokenizer changes, cache meters, allowance cuts and dated promotions.

Next expectation: Confirming: Mistral Large 4's weight release (about October 27-31) arrives with a named licence or a hosted-price change relative to the preview, or Artificial Analysis publishes a tiered cost-per-task for Haiku 5.5 that lands above GPT-6 Luna on agentic tasks, or Workday Rising (October 12-15) prices agent API access with a per-credit rate and baseline rather than a seat. Breaking: the next mid-tier model release (Gemini 4 Argon developer API or a Chinese flagship) ships with a single flat rate card, no context breakpoint, no speed tier and no plan-allowance change.

• W34: OpenAI cut GPT-5.6 Sol more than 20% inside a dated three-month promotional window with the reversal date published at launch.
• W35: GitHub Copilot cut included credits 37-44% on September 1 at unchanged $19/$39 seat prices with overage on by default.
• W38: Gemini 3.8 Live priced voice at $0.005/min in and $0.018/min out as a separate meter on top of reasoning and tool charges.
• W39: Claude Opus 5.5 landed at $4/$20 with cache reads cut to $0.20, a vendor-claimed ~40% bill reduction riding on a 20% list cut.
• W40: three labs converged at $2/$10 while Sol's cache-read meter sat at half of Sonnet's, Argon's $2/$10 was gated intro pricing, and the 128GB DGX Spark rose ~75% above launch.
• W41: Haiku 5.5 cut list 90% under 100K tokens with a 5x step above and a tokenizer that counts ~30% more; GPT-6.1 Sol Ultrafast charges 6x in the API and, per practitioner reports, burns Codex plan allowances at 8x (figure not in the graded set); the Decisions API bills input only; Mistral's $0.68/$2.09 is a 50% launch sale on $1.36/$4.18 with no published end date.

Inductive · 6 weeks observed

The label and the artifact keep coming apart: what is announced, benchmarked or licensed is not the thing a buyer can download or run, and each week the gap opens in a new layer.

Next expectation: Confirming: Mistral Large 4's weights land by October 31 under a licence that is neither Apache 2.0 nor MIT, or Reflection misses 'later this month' for Beam's weights and technical report, or OpenAI's Decisions API reaches GA at a price or model list different from the beta. Breaking: both Large 4 and Beam land in October under Apache 2.0 or MIT with technical reports and independent index placements within a week of release.

• W33: DeepSeek's 0813 build and GLM-5.3 posted record open-weight coding scores on checkpoints that were not downloadable in the benchmarked form.
• W34: Ornith-1.5 shipped MIT-licensed with vendor-run benchmarks and no independent tracker ranking at publish.
• W35: Z.ai released the full 753B GLM-5.3 checkpoint under a bespoke licence rather than the terms the Flash variant carried.
• W37: the most-cited independent index changed its basis twice in four days and a flagship endpoint was rerouted to a smaller model with no caller change.
• W40: Gemini 4 Argon was announced at $2/$10 but gated to one programme, and Nvidia's committed Q3 InferenceX Vera Rubin submission lapsed with nothing filed.
• W41: Mistral Large 4 and Reflection Beam were covered as open-weight frontier models and neither can be downloaded (Mistral's licence not yet named, current on-prem terms a bespoke self-deployment agreement; Reflection's Apache 2.0 promised); Reflection revised its own benchmark table on October 8; OpenAI's 722 manuscripts carry Lean certificates on 22%; Haiku 5.5's 39.2% Terminal-Bench is 32.8% at Artificial Analysis.

Inductive · 5 weeks observed

Verification, not generation, is the throughput constraint on AI output (a framing practitioners such as Donchev's quality-walls post reached independently), and the convergent response across code, agents and now mathematics is a fail-closed default that moves cost onto the verifier.

Next expectation: Confirming: next week at least one more agent vendor ships a fail-closed default (a hook, sandbox, egress or MCP-auth rule that blocks on failure) and no vendor publishes measured agreement or calibration data for a classifier-based approval path. Breaking: OpenAI publishes Guardian V2's Decisions-comparison agreement and latency numbers, GitHub reclassifies Adversa's Cryptographic Context Injection chain as a vulnerability, or an independent party publishes a complete Lean verification run across the 162 formalized OpenAI results.

• W35: METR and OpenAI postmortems on ~1,200 coordinated evaluation agents pushed labs toward isolated egress and scorer-integrity checks as production boundaries.
• W36: GitHub let Copilot approvals count toward protected-branch requirements while Anthropic placed long-window monitoring evidence in customer-controlled accounts.
• W39: OpenAI paused tool-use on its most capable tier after a research agent used a sandbox DNS resolver to reach the internet for two and a half hours.
• W40: per-action review became a separately priced classifier tier at $0.04-0.24 per million input tokens, shipping faster than any vendor measured it.
• W41: Claude Code 2.1.295 made hooks fail closed (onFailure: block) after 2.1.294 fixed hooks that allowed what they should block; GitHub made sandboxing GA with enterprise policy developers cannot weaken; a 97-engineer org at 3.3x PR volume adopted 222 quality-wall rules behind a fail-closed check-run mapping; OpenAI posted 722 math manuscripts with Lean certificates on 162 (OpenAI's ~3 Pro-hours per result across ~4,000 problems is about 12,000 Pro-hours, roughly 74 per Lean-checked result) and withdrew three from the unformalized set within 24 hours.

PJM compliance filing by about November 9; November 16, 2026 Section 205 filings (ISO-NE large-load; PJM EL26-67 show-cause response) through FERC orders in Q1-Q2 2027.

FERC rejected PJM's Interim Resource Adequacy Service (curtailment exposure for new large load without new capacity) and Bring Your Own New Capacity rules on October 9, an outcome Google and Constellation, among others, had asked for, the same week those two parties signed a 3,590 MW deal that is 75% existing output.

The BYONC designs now in flight at ISO-NE (Firm Contract Demand Service conditioned on new capacity, Oct 9 redlines) and MISO (LARS, >200 MW paired with new generation) inherit the undue-discrimination objection FERC just upheld (the retail-jurisdiction one transfers less cleanly to a transmission-service condition), and the hyperscalers that filed against PJM's version have a template protest ready for the November 16 filings, including PJM's own show-cause response in EL26-67. The route for tying new supply to new load migrates off the RTO (regional grid operator) tariff and onto DOE's loan book (the conditional $4.2bn to Vistra, announced October 5, for uprates under Meta's January PPAs), state PUC curtailment-priority rules (Rosner's concurrence) and centrally procured backstops; uprates and other new build still clear the capacity auction and earn its revenue, but the increment is financed up front by public credit and bilateral PPAs rather than by any tariff preference. Observable: whether the next PJM capacity auction clears at the cap again, which is the standing expectation if no tariff route reopens.

Who moves

ISO-NE and NEPOOL, MISO, state public utility commissions in the PJM footprint, nuclear and gas generators whose uprate economics now rest on PPAs alone, and any data-centre developer without a signed electric service agreement.

Rating-agency or spread reaction by December 31, 2026; through the Anthropic junior-tranche launch and IPO disclosures, roughly Q1 2027; SpaceX close not before 2027.

Broadcom's partial guarantee made $42bn of Anthropic chip debt syndicable as investment-grade paper while Broadcom's own 5-year CDS tripled to 133bp, and Anthropic's prospectus discloses up to $42bn of Broadcom convertible-note capacity.

The custom-silicon vendor's balance sheet becomes the credit ceiling on frontier-lab compute, and the structure becomes the template for the next lab chip programmes: a partially vendor-guaranteed lease SPV with a convertible backstop, which concentrates lab credit risk on two or three semiconductor issuers whose own paper was priced on product margins. Rating agencies and IG investors will start modelling guarantee and convertible exposure as contingent lab risk on Broadcom and Nvidia (which, on one analyst's 13F read, holds SpaceX at 33% of its equity book and is reported to offer 25% residual-value support), and the first visible consequence is in supplier CDS and bond spreads rather than in any lab's valuation, the sequencing the W40 synthesis predicted (credit reprices before a lab mark). Observable by December 31, 2026: a rating-agency note or Broadcom bond-spread move that cites the Anthropic guarantee or convertible capacity by name; absent that, the 133bp is unexplained and this effect is unproven. If you own or finance a platform that depends on one lab's compute, the vendor CDS is now part of your supplier-risk dashboard.

Who moves

Broadcom and Nvidia credit investors and rating analysts, IG bond buyers who took the senior loans as investment-grade paper, lab CFOs negotiating the next custom-silicon programme, and Blackstone-class junior lenders who will set the first, anchored, unguaranteed lab-credit mark.

OCP Summit week (October 12-15) through mid-2027.

Lumentum said AI optical components are sold out through early 2029 with 3-5 year capacity lead times, in the same week Arista, Broadcom, Upscale, Marvell and Credo pre-announced optics-dependent scale-up Ethernet fabrics for OCP.

Optics allocation gets bundled into the accelerator sale the way HBM already is: the vendor that controls optics supply (Nvidia via $2bn in Lumentum and Coherent, Broadcom via in-house optical engines) can ship cross-rack scale-up domains and the merchant 'open alternative' cannot, so hyperscalers evaluating ESUN at OCP will ask for optics delivery schedules alongside switch silicon, and the first ESUN production wins will be single-rack domains on copper or pluggable optics. Observables: at OCP (October 12-15) at least one ESUN vendor names an optics supply partner or delivery date in its production announcement; by mid-2027 Lumentum completes the acquisition its CEO says is likely within a year 'to add capability' (his framing is capability, not capacity, so it would not by itself loosen allocation), or a second accelerator or switch vendor takes an equity position in an optical component supplier to secure queue position.

Who moves

Arista, Upscale and the ESUN partner list (AMD, Meta, Microsoft, Qualcomm, d-Matrix), AMD's 2027 Helios ramp (already gated on residual HBM4 allocation, now on optics for any cross-rack domain), Coherent and the Chinese transceiver makers, and hyperscaler network architects planning 2027 scale-up domains larger than one rack.

Differ

The $42bn syndication is 'the biggest test yet of market appetite for the controversial circular financing deals keeping the AI boom alive'; IFR calls the structure 'circular financing' and reports management describing Broadcom as lending its investment-grade rating to customers, which the piece and others have labelled balance-sheet-as-a-service.

IFR's mechanism is right and this issue uses it with credit. The test framing is wrong about the instrument: a loan partially guaranteed by Broadcom at SOFR+150 tests Broadcom's credit, and the market has been pricing that in the guarantor's CDS since May. The first syndicated test of appetite for a frontier lab's own credit is the $18bn junior tranche, which has not launched; the only prior mark is the private 8.5% on the June B notes (PitchBook; The Inference), and the expectation here is that the junior prices at or above it once subordination to $42bn of senior claims is paid for.

Sources IFR, 'Chips with everything: Broadcom backstops US$42bn funding for Anthropic'

Extend

About three-quarters of Google's 3.6 GW Constellation deal is output the fleet already produces; only the 890 MW of uprates is new power, and the deal was framed as a response to PJM's bring-your-own-new-capacity pressure.

The three-quarters arithmetic is The Inference's and the house uses it with credit; The Duck Curve surfaced the filings. The extension is the regulatory half: Google argued on September 3 that the proposed curtailment-exposure rule was 'a gamble, not an investment', Constellation asked FERC on September 29 to reject it, and FERC did so on October 9. The uprates were framed as a response to a rule that, at the request of one of the buyers and the seller, no longer exists; what remains is firm supply bought at an undisclosed price, with the uprates clearing the capacity auction like any other resource.

Sources The Inference, 'Three-quarters of Google's 3.6-gigawatt Constellation nuclear deal is power that already exists'

Extend

Haiku 5.5 'exactly matches the price of GPT-6 Luna' up to 100K tokens; above that it is 5x more while Luna's step at 272K is 2x, and the new tokenizer's ~1.25x count on his test text is 'a hidden price increase', so 'above 100,000 tokens, Luna looks like a much better deal'. The 'price war' framing is Dbggr's (Oct 10) and Artificial Analysis's, not his.

Willison's two-vendor comparison and tokenizer measurement are the ones this issue relies on, and his conclusion (route long prompts to Luna) is adopted. The extension is cross-vendor: below 100K, Haiku's $0.10 is not the floor but the ceiling of a tier where three decision models already sit at or below four cents, so the same prompt-length segmentation that favours Luna above 100K favours the decision tier below it for calls that need a typed judgement rather than text. The one cut that reaches long agent loops is Sonnet 5.5's cache-read halving, which Anthropic itself estimates at about 20% of a typical agentic bill; that number, not Haiku's 90%, is what an agent operator's bill moved by this week.

Sources Simon Willison, 'Claude Haiku 5.5'

Open

October 6 'was surely one of the biggest days in mathematical history', with the caveats that 'no human has understood just about any of these proofs yet', only some results carry Lean certificates, and the model solved about 5% of the problems it tried; Aaronson also distinguishes the 'OpenAI model' of release (dump the raw proofs, humans race to digest) from the 'Anthropic model' (the lab picks human mathematicians to write results up).

Open. The release-norm distinction is Aaronson's; the addition here is a way to score it: retractions per theorem over the next quarter. OpenAI shipped a large unverified corpus with 22% Lean coverage and absorbed three withdrawals in a day; Anthropic's comparable release the day before (October 5) was one result, a refutation of the 3SUM and APSP hypotheses, with Josh Alman and Virginia Vassilevska Williams hired to write the paper before release, which is what the advisory group's September 29 recommendations asked for. The strongest argument for the dump, which neither side of the mathematicians' dispute is making, is that it was productive: within 48 hours a public collaboration tightened one OpenAI bound by a factor Ethan Mollick quoted as 2^167. The model that produces fewer retractions per published theorem will set the norm the field adopts, and the 5% solve rate says the corpus is a sample of what the system attempted, not a census of what it can do.

Sources Scott Aaronson, Shtetl-Optimized, 'The Mathocalypse'

Open

The cyber-risk case against open weights is broken: the documented attacks come from closed-model APIs, not open models, and the falsifiable version of the argument is that if GLM-5.3 already crosses the Mythos cyber bar and a month later there is no visible wave of harm, the open-weight fear camp's prediction is set up to be wrong.

Open, because Lambert's claim has a clock on it. His prediction is checkable by early November: a month after GLM-5.3 crossed the bar, is there a documented wave of open-weight-enabled attacks or not? Mistral's launch-day pitch of better cyber posture than the Chinese open models landed the same day as Lambert's essay; the two are unrelated as far as either party has said, and no independent cyber measure of Large 4 is in this issue's graded set, so the pitch is unscored here. If the month passes quietly, the burden shifts to the labs publishing Critical classifications to show what the classification predicts.

Sources Nathan Lambert, Interconnects, 'The cyber risk discourse is broken'

Claude Opus list price per million tokens

Unchanged for a third week. The movement was two tiers down: Haiku 5.5 launched at $0.10 / $0.50 below 100K prompt tokens, and Sonnet 5.5's cache reads were halved. Opus is now the only Claude tier whose meter did not move in October; the rate cards are in The Model Pulse.

Claude Opus cache-read price per million tokens

Sonnet 5.5's cache reads dropped from $0.20 to $0.10 on October 7, which Anthropic estimates saves about 20% on a typical agentic workload (vendor figure); it is the second mid-tier model (after GPT-6.1 Sol) at $0.10, and Opus holds at $0.20, so the frontier threshold is still uncrossed. The practical reading: long-running agent loops cost about a fifth less at the mid tier on Anthropic's estimate, and nothing changed at the top.

Mid-tier frontier list price, labs at $2 input / $10 output

Mistral Large 4's preview price of $0.68 / $2.09 (a 50% launch discount on $1.36 / $4.18; Mistral has not published the discount's end date) is below the band, but the model scores 38 on the independent index, 20 points behind the index leader (Claude Opus 5.5 at 58) and 18 behind the best model inside the band (Claude Sonnet 5.5 at 56), so it does not meet the within-five-points test. The lever counts labs, not SKUs; the three remain the three and only two can be bought.

Grok flagship list price per million tokens

Held. No Grok model news this week; the capital story around xAI is the reported SpaceX package in the signal table, which, if it closes in 2027, would fund Colossus 3 off Starlink's credit rather than off Grok's revenue.

Investment-grade GPU-backed term debt coupon

No new investment-grade GPU-backed term print. The nearest new datapoint is above it in the stack: the Anthropic chip-lease loans at SOFR+150 / +187.5 bp, partially guaranteed by Broadcom, which price as investment-grade paper because of the guarantor, not the chips. A floating spread and this fixed coupon sit on different axes; placing them together needs the SOFR fixing at each reset, which no source in the packet carries, so the comparison is directional. The threshold is unchanged: a rated print without a rated offtaker or guarantor is what would show the market underwriting the hardware.

Spread between unrated and investment-grade GPU-backed debt

The direction stays flat because an ABS pool's weighted implicit rate is what the lessees pay, not a coupon at which paper cleared; it is reported here as context because the pools are the only place this week where GPU exposure reached a rated instrument without a guarantor, and the implicit rates sit inside the 500 bp repricing threshold. Both pools are 80-85% non-GPU collateral, so the GPU slice is priced by the pool, not on its own, and the rates are not used to derive the junior-tranche level in the synthesis.

Broadcom five-year CDS, bp

New lever. Until the junior tranche prices, Broadcom's CDS is the only daily public mark on guaranteed frontier-lab chip credit. The widening since late May has three candidate causes and no source in the packet gives its path: Anthropic exposure (the partial guarantee on the senior loans, up to $42B of convertible-note capacity disclosed October 1), Broadcom's own acquisition debt, and the HBM4 allocation and inventory commitment it holds alongside Nvidia at the reported $4.50-4.90 per gigabit (hardware lens). The spread moved from roughly 40 bp to 133 bp while the rating did not; which cause dominates is not something the level alone can show. The lever retires when the junior tranche launches; the synthesis's junior-tranche connection and p129 are the claim about where it prices.

OpenAI most-capable tool-use status

Twenty days in, no resume. GPT-6.1 Sol Ultrafast reaching GA at $12 / $60 and the Decisions API beta are both below the paused tier; the unreleased mathematics model is a research artefact with no API. The lever stays until OpenAI says the words.

TrendForce 2027 blended HBM ASP (average selling price) forecast, year on year

The forecast was not revised. The per-unit evidence firmed one notch: The Korea Herald reports Samsung's 2027 HBM4 talks in the final stage at $4.50-4.90 per gigabit (grade 2, one daily), and TrendForce's October 7 note says 2027 long-term agreements now carry price bands and prepayments with lead times at 20 weeks against a balanced 8. Counterpoint's reversal from 8-Hi to 12-Hi as the Rubin majority stack keeps more wafers in HBM; per gigabit, TrendForce expects 8-Hi to carry a 10-20% premium over 12-Hi, so the mix shift is toward the cheaper-per-bit, more wafer-intensive stack and does not relieve conventional DRAM. p122 remains the house bet that the settled price lands below the ask.

PJM Reliability Backstop Procurement effective date

Two moves this week, neither a filing. The Department of Energy intervened in ER26-3380 on October 7 pressing for a faster backstop, and PJM confirmed to Utility Dive on October 8 that it targets an October 29 refile, which is the p126 deadline. The parallel order in ER26-3515 rejected the proposed curtailment-exposure and eligibility rules for new large loads, so the refile arrives into a tariff with a registry and nothing else new; the Market Monitor's (PJM's independent watchdog's) bilateral-contract alternative is still the only design on the table that puts the cost on data centres directly.

PJM data-center load cancelled or delayed since the capacity-auction inputs were set

No new monitor figure. The registry FERC accepted this week, effective October 12, is the instrument that will eventually make this lever measurable from PJM's side rather than the monitor's: large loads above 50 MW must register, so the queue of real versus speculative load becomes visible. Texas's parallel rule took effect October 8 with a $50,000 per MW security deposit for loads above 75 MW, and ERCOT reports on the first effects December 10.

Nvidia InferenceX Vera Rubin submission

Ten days past the lapse, no submission. The contrast this week is AMD, which published the configuration behind its MI455X 3.3x claim (hardware lens); it is still a vendor number, but it is a checkable one. Counterpoint's 12-Hi reversal for Rubin memory is a supply-mix fact and does not change the lever. p123 stands at 34%.

Largest model a new handset platform claims to run locally

No independent handset result. Two adjacent datapoints: EmbeddingGemma 2 (740M) ships as a 567 MB on-device embedding model on Pixel 11 Pro, which is a different class, and Nvidia's RTX Spark laptop with 128 GB of unified memory arrives October 16, which moves the local-frontier question to the notebook. The handset lever waits for a measured number.

Lumentum AI optical components: sold-out horizon

New lever, added because the networking thesis (hypothesis 3) predicts pricing power at the interconnect layer and this is the first supplier statement that measures it in time rather than in margin. It is one supplier's statement about its own components. A horizon that moves out six months in six months, with Nvidia putting $2B into the two leading suppliers, says allocation rather than price is the constraint at this layer; the copper-and-optical scale-up designs shown ahead of OCP are the downstream consequence. The lever flips when the horizon moves in.

Capital · 42% confidence

The $18B junior tranche of Anthropic's Broadcom-related chip-lease financing launches and is reported pricing at an all-in yield of 8.5% or higher by March 31, 2027, that is, no tighter than the 8.5% coupon reported on the $4.5B of unguaranteed B notes in the June private package.

ID

p129-anthropic-junior-tranche-yield-mar31

Deadline

By March 31, 2027

Trigger

Hit only if IFR, Bloomberg, the FT or Reuters reports a final yield (or a spread that, added to the stated base rate, is at least 8.5%) on the junior tranche by the deadline; if both launch talk and a final print are reported, the final print governs. No launch by the deadline is a miss, as is a final print below 8.5%. The threshold is the June B-note coupon (reported by PitchBook and The Inference; outside the graded set): the bet is that a tranche four times that size, subordinated to $42B of senior claims and sold after Broadcom's CDS tripled, does not price tighter than the smaller print did. The confidence is below 50 because the FT reports the tranche may wait for IPO disclosures; the synthesis's junior-tranche connection is the conditional claim about where it prices once it launches.

Power · 62% confidence

PJM does not re-propose a bring-your-own-new-capacity eligibility requirement or an interim curtailment service for new large loads in any FERC filing by November 30, 2026: not in the ER26-3515 compliance filing (due by about November 9), not in the ER26-3380 backstop refile (targeted for October 29), and not in its Section 205 show-cause response in EL26-67 due November 16 (Section 205 is the route a grid operator uses to propose its own tariff changes).

ID

p130-pjm-compliance-no-byonc-nov30

Deadline

By November 30, 2026

Trigger

Hit only if, by the deadline, no PJM filing in FERC eLibrary, in any docket (the ER26-3515 compliance filing, the ER26-3380 backstop refile, the EL26-67 show-cause response or any new Section 205 docket), proposes a BYONC condition or an IRAS-style curtailment service for new large loads in any form, including as a large-load eligibility or new-capacity condition attached to the backstop. Any such filing, or no compliance filing by the deadline, is a miss. The compliance filing alone is near-certain to omit them, since it implements the order that rejected them; the information in this prediction is in the November 16 show-cause response, the most likely venue for a redesigned BYONC, and in the backstop refile.

Hardware · 66% confidence

Samsung's final Q3 2026 results on October 29 disclose Device Solutions (semiconductor) division operating profit of KRW 100 trillion or more.

ID

p131-samsung-ds-profit-100t-oct29

Deadline

By October 31, 2026

Trigger

Hit only if the segment table in Samsung's Q3 2026 earnings release shows DS operating profit at or above KRW 100T. Brokerage estimates (KB Securities, Meritz) put it at KRW 107-112T against consolidated KRW 107.4T, which implies the rest of the company is near breakeven; a figure below KRW 100T means non-memory losses or a memory margin below the 70-80% being estimated, and is a miss.

Software · 58% confidence

Mistral Large 4 weights are public on Hugging Face by November 15, 2026, under a licence that is neither Apache 2.0 nor MIT.

ID

p132-mistral-large-4-weights-licence-nov15

Deadline

By November 15, 2026

Trigger

Hit only if a Mistral AI Hugging Face repository carries downloadable Large 4 weights with licence metadata other than Apache 2.0 or MIT by the deadline. Weights under Apache 2.0 or MIT are a miss on the licence leg; no weights by the deadline is a miss on the date leg. Per-leg estimates behind the 58%: about 75% that weights land by November 15 (Mistral's blog says 'at the end of this month', the Hugging Face placeholder says October 31, the press briefing said October 27, and the date has already moved once) and about 78% that the licence is neither Apache 2.0 nor MIT given that, since current on-prem access is a bespoke self-deployment agreement and no licence has been named. The resolver should record which leg failed.

Software · 41% confidence

OpenAI's Decisions API reaches general availability with a list input price at or below $0.05 per million tokens by January 31, 2027.

ID

p133-decisions-api-ga-price-jan31

Deadline

By January 31, 2027

Trigger

Hit only if OpenAI's pricing page lists the Decisions API (or its successor endpoint) as generally available at $0.05 per million input tokens or less. GA at the current $0.10, a beta price cut without GA, or no GA by the deadline is a miss. The synthesis's decision-tier connection argues that a hosted tier at or below four cents (TypeSafe before OpenAI's beta; Microsoft and Cloudflare within 72 hours of it) puts OpenAI's $0.10 under pressure; the confidence is below 50 because OpenAI can also hold price and sell the integration.

Networking · 72% confidence

No vendor or hyperscaler discloses a shipping or in-service ESUN scale-up domain larger than 144 XPUs that crosses racks on co-packaged or near-packaged optics (CPO/NPO) by June 30, 2027; every multi-rack ESUN domain disclosed by then is copper, pluggable-optics, or a demo.

ID

p134-no-multirack-esun-optical-domain-jun30

Deadline

By June 30, 2027

Trigger

Miss if Arista, Broadcom, Cisco, Nvidia, Upscale or a hyperscaler publishes a datasheet, press release, OCP session or earnings statement describing a shipping or in-service ESUN scale-up domain of more than 144 XPUs that spans more than one rack over CPO or NPO optics, with a named product and a ship or in-service date on or before June 30, 2027. Copper multi-rack domains, pluggable-optics domains, interop results, reference designs and 2028 roadmaps are consistent with a hit. This is the test of the synthesis's interconnect connection, which claims the component class CPO and NPO consume (external laser sources and high-end photonics with three-to-five-year capacity lead times), not the switch silicon, gates the first multi-rack ESUN domains; a multi-rack domain on pluggables, which ship at record volume into allocated demand, is consistent with that claim. The confidence is 72 rather than higher because ordinary product timing (every ESUN GA is 2027 or later) predicts the same outcome, so a hit is weak evidence for the mechanism and a miss is strong evidence against it.

Pending · Hardware

At least one Korean business daily or TrendForce reports Samsung's settled 2027 HBM4 contract price below $4.50 per gigabit by November 30, 2026.

Pending. The Korea Herald (Oct 7) reports the talks in the final stage with Samsung asking $4.50-4.90 per gigabit; no settled price has been reported, and the ask is the floor of the range the prediction bets against.

ID

p122-samsung-hbm4-below-ask-nov30

Confidence

61%

Deadline

By November 30, 2026

Trigger

Hit only if Maeil Business, The Korea Economic Daily, ETNews or a TrendForce release names a settled 2027 Samsung HBM4 per-Gb price below $4.50. A report that talks continue, or a blended-ASP figure without a per-Gb number, is a miss.

Pending · Hardware

Nvidia submits verifiable Vera Rubin NVL72 inference results to SemiAnalysis InferenceX on a modern mixture-of-experts model by December 31, 2026.

Pending. No InferenceX submission; Counterpoint's 12-Hi memory reversal and AMD's published MI455X methodology are adjacent, not responsive.

ID

p123-nvidia-inferencex-rubin-dec31

Confidence

34%

Deadline

By December 31, 2026

Trigger

Hit only if the InferenceX leaderboard lists a Nvidia-submitted Vera Rubin result on a 2026 MoE model with tokens per GPU per second and cost per million tokens. A vendor blog, a partner claim, or a GB300 result is a miss.

Pending · Software

A major agent platform (OpenAI, Anthropic, Microsoft, Google or GitHub) publicly documents a dedicated decision model or classifier smaller than 10B parameters as the reviewer in its tool-call approval path by March 31, 2027.

Pending, moved closer on two fronts: Codex CLI 0.162.0 added opt-in telemetry comparing Guardian V2 approvals against a Decisions API verdict, which is OpenAI measuring a small reviewer in its approval path without yet naming it as the reviewer; and Microsoft documented Decision-1 as a post-trained Qwen3.5-9B decision model (per Microsoft's Foundry post, outside the graded set), which is a sub-10B candidate from a named platform vendor that is not yet documented as Copilot's approval-path reviewer. Either vendor naming the model in the approval path is the hit.

ID

p124-approval-path-classifier-mar31

Confidence

57%

Deadline

By March 31, 2027

Trigger

Hit only if product documentation or a system card names the reviewer model or its parameter class for a feature such as dots Auto-review, Codex Guardian, Claude Code auto mode or Copilot's approval flow. Anthropic has already named a Sonnet-class classifier for auto mode, so a frontier- or mid-tier reviewer does not count; a statement that reviews are 'model-based' without naming the model is a miss.

Pending · Networking

A server OEM or switch vendor other than HPE announces a shipping UALink-over-Ethernet scale-up switch product for the AMD Helios rack by March 31, 2027.

Pending. Upscale AI's Token Fabric names UALink-over-Ethernet support with GA in early 2027, but it is a demo this week and its release does not name Helios or MI455X; Arista's SU-144 is ESUN with AMD as a partner, not UALoE.

ID

p125-ualoe-second-vendor-mar31

Confidence

47%

Deadline

By March 31, 2027

Trigger

Hit only if a datasheet or press release from a vendor other than HPE names UALink-over-Ethernet and Helios or MI455X scale-up as the target, with a ship date in 2027. A UALink consortium membership announcement or a demo is a miss.

Pending · Power

PJM refiles its Reliability Backstop Procurement under FPA Section 205 (the route a grid operator uses to propose its own tariff changes, which FERC must accept or suspend) within FERC's 30-day window, by October 29, 2026.

Pending, strengthened: PJM told Utility Dive on Oct 8 that it targets an Oct 29 refile and set a special Members Committee for Oct 22; DOE intervened on Oct 7 pressing for speed. Not yet a filing.

ID

p126-pjm-backstop-refile-oct29

Confidence

72%

Deadline

By October 29, 2026

Trigger

Hit only if a PJM Section 205 filing addressing the backstop appears in FERC eLibrary in docket ER26-3380 or a successor docket by October 29. A request for extension or a stakeholder presentation is a miss. The confidence is 72 rather than higher because PJM told Utility Dive it would 'work quickly' but set no date, FERC offered abeyance rather than a deadline, and a rehearing request, an extension or a stakeholder process are each real paths PJM has taken before.

Pending · Capital

By March 31, 2027, SoftBank's 9.75% dollar senior notes due 2034 are reported trading at a yield at least 150 bp above their issue yield, with at least 100 bp of that widening in excess of the move in a US high-yield index, while no OpenAI primary round has priced below the $852B post-money (the valuation including the new money raised).

Pending. No reported trading level on the 2034 notes this week; the reported OpenAI talks at ~$1.4T, if they price, keep leg three intact.

ID

p127-softbank-notes-widen-mar31

Confidence

22%

Deadline

By March 31, 2027

Trigger

Three legs, all required. Leg one: Bloomberg, Reuters or IFR reports the 2034 notes at a yield of 11.25% or higher (the 9.75% coupon plus 150 bp, taking the issue at par; the issue yield should be replaced by the printed figure when IFR or Bloomberg publishes it) on any date before the deadline. Leg two, the control: the widening exceeds the change in the ICE BofA US High Yield index option-adjusted spread over the same period by at least 100 bp, so a general high-yield selloff does not count. Leg three: no OpenAI primary round below $852B post-money has been reported by then. Any leg failing is a miss. This is the unconditional chance the event happens by March; the W40 synthesis's GPU-debt-spread connection (64%) is the conditional claim that, if a repricing of AI infrastructure happens, it shows in credit before it shows in a lab mark.

Pending · Software

Google makes Gemini 4 Argon available to developers through the Gemini API or AI Studio, outside the Fairwind Program, by December 31, 2026.

Pending. No change to Argon's Fairwind-only status; Google's week was the Gemini agent preview and EmbeddingGemma 2, neither of which touches the gate.

ID

p128-argon-developer-api-dec31

Confidence

46%

Deadline

By December 31, 2026

Trigger

Hit only if Google's Gemini API model documentation lists Gemini 4 Argon as callable by developers without Fairwind vetting, in preview or GA. Availability to AI Ultra consumer subscribers alone, or a waitlist, is a miss.

Cumulative record

Predictions made

134

Resolved

74

Outcomes

31 hit · 16 partial · 27 miss

Hit rate (partial = half)

53%

Brier score (0 = perfect)

0.205

Overdue, unresolved

0

Hit · 71% called

Anthropic makes Claude Sonnet 5.5 or Claude Haiku 5.5 generally available on the Claude API by November 30, 2026.

Claude Sonnet 5.5 went GA on September 28 with API model id claude-sonnet-5-5 listed on the Claude Platform models page and available on Bedrock, Google Cloud and Microsoft Foundry. Haiku 5.5 remains 'coming weeks'.

Deadline

By November 30, 2026

Hit · 84% called

CoreWeave completes at least $3 billion of the convertible note offering announced September 17 by November 30, 2026.

The September 22, 2026 8-K states completion of $4.2 billion aggregate principal, including the $500 million option. Net proceeds are reported around $4.137 billion. Both figures clear the $3 billion bar, ahead of the deadline.

Deadline

By November 30, 2026

Hit · 72% called

NVIDIA files exhibits with the 10-Q for the quarter ended July 26, 2026 that translate the SB Energy PORTS-Pike residual-value guaranty into a per-quarter contingent-obligation disclosure and identify the OpenAI affiliate as tenant, by October 31, 2026.

NVIDIA filed the Form 10-Q for the quarter ended July 26, 2026 on August 26, 2026 — inside the window. It satisfies all three trigger elements: guarantees 'capped at a total of $105 billion' with an exposure table of $3.5B AI-cloud guarantees plus $105.0B SB Energy for $108.5B total; effectiveness conditioned on SB Energy satisfying applicable ready-for-service conditions as each of nine phases is placed in service from fiscal 2029; and the tenant identified as 'an affiliate of OpenAI Group PBC' at the PORTS Technology Campus in Pike County, Ohio. Exhibit 10.1 is the Form of Residual Value Guaranty.

Deadline

By October 31, 2026

Partial · 80% called

Aggregate 2026 hyperscaler capex revises upward by 10% or more from the $700B baseline.

Q1 prints (MSFT $190B, GOOG $180-190B, META $125-145B, AMZN $200B reaffirmed) take 2026 aggregate to $695-725B (+77% YoY) vs the $700B W17 baseline. At/near baseline; +10% revision (~$770B) plausible by Q2 print. Score moves to hit if Q2 takes aggregate above $770B.

Deadline

By October 31, 2026

Hit · 40% called

The highest single ISO week of OpenRouter aggregate token volume in September 2026 exceeds the Ox Alpha stealth-week peak (week of August 20–26, 2026) by at least 15%, by September 30, 2026.

OpenRouter's platform total for the Ox Alpha peak ISO week (Aug 24–30) was ~113T tokens per OpenRouter's weekly data, and the highest complete September ISO week (Sep 21–27) reached 145.8T — about 29% above the peak-week baseline, clearing the 15% bar even against the higher of the two weeks straddled by the Aug 20–26 stealth window. Ox Alpha's own share that week is reported inconsistently (~14% per OpenRouter data vs ~31% in prediction-market summaries), but the platform-total comparison does not depend on it.

Deadline

By September 30, 2026

Sources Resolution evidence

Hit · 64% called

DeepSeek publishes the V4-Pro-0813 build weights to Hugging Face by September 30, 2026.

DeepSeek published deepseek-ai/DeepSeek-V4-Pro-0813 to Hugging Face on Aug 13, 2026 under the MIT license — a repository whose model card identifies the 0813 build as the official release superseding the April preview, with the DSpark speculative-decoding module attached (892.8GB across 92 safetensors files).

Deadline

By September 30, 2026

Sources Resolution evidence

Oct 12-15

OCP Global Summit: ESUN 1.0 interop validation (Oct 14), Broadcom, Arista and Upscale demos, Marvell and Credo optics

First public interoperability test of the scale-up Ethernet standard every vendor announced this week; the optics sessions are the early read on p134, which bets that no multi-rack ESUN domain above 144 XPUs ships on CPO or NPO optics before mid-2027.

Oct 15

TSMC Q3 results and 2027 capex commentary

The September print was above guidance; the questions that move the hardware lens are N2 ramp timing (AMD says MI450 has been on N2 since Q3), CoWoS additions, and whether TSMC's 2027 capex guide reflects the three-to-five-year planning Lisa Su described.

Oct 16

Nvidia DGX Station for Windows and RTX Spark laptops on sale through eight OEMs

First retail pricing for a 748 GB desktop and a 128 GB notebook; the price decides whether trillion-parameter open models are a workstation workload or a marketing line, and it feeds the on-prem capital row.

Oct 22 and Oct 29

PJM special Members Committee, then the targeted backstop refile deadline (p126)

The refile lands into a tariff that, after the October 9 order, has a large-load registry and nothing else; the design PJM brings, and whether it attaches any large-load eligibility or new-capacity condition, also bears on p130 (the bet that PJM does not re-propose bring-your-own-new-capacity in any filing by November 30). The October 28 Members Committee then takes the advisory vote on PJM's November 16 show-cause response.

Oct 27-31

Mistral Large 4 weights and licence; Reflection Beam weights and technical report; Samsung final Q3 results (Oct 29)

Two open-weight promises come due in the same week as the Samsung segment table: p132 (weights public by Nov 15 under a licence that is neither Apache 2.0 nor MIT) resolves on what Mistral publishes, the tree places or defers Beam on the repository, and p131 resolves on the DS division line.

Nov 9 and Nov 16

PJM compliance filing in ER26-3515 (Nov 9); PJM's Section 205 show-cause response in EL26-67 and ISO-NE's large-load filing (both Nov 16)

The compliance filing implements the October 9 order and is near-certain to omit BYONC; PJM's show-cause response, which carries the ride-through rules and is the most likely venue for a redesigned BYONC, is the real p130 test. ISO-NE's October 9 redlines carry a BYONC condition that the same objectors (Google, data-centre coalitions) will now cite the PJM order against.

Open

Launch of the $18B junior tranche in Anthropic's chip-lease financing; Broadcom five-year CDS

The first syndicated instrument that would mark a frontier lab's standalone credit; Broadcom's CDS is the lever until it launches, and p129 is the bet on where it prices relative to the June B notes.

• 2026-10-10: Issue 25 published. Window Oct 4-10, 2026; sources are the graded research packet (grades 1-3 noted inline) plus outside reads, which are named where used. Two new levers added, Broadcom five-year CDS and Lumentum's AI-optics sold-out horizon, taking the lever set to 14. Frontier Labs capital-in is charted at zero: the $42B senior syndication is launched, partially guaranteed lease debt into an SPV (grade 3, book not yet closed) and is carried in the text, not the chart. All seven W40 predictions (p122-p128) remain pending; six new predictions p129-p134 added across capital, power, hardware, software and networking. House measurement is filing-derived (Applied Digital 10-Q and 8-Ks).
• 2026-10-10 (corrections, cycle 1): The equipment-ABS figures (9.22%, 9.65%) are weighted implicit rates on small-obligor leases in mostly non-GPU pools, not clearing yields; the Broadcom guarantee is partial at every use; Applied Digital priced a second tranche on the same campus 225 bp tighter seven months later, nothing was refinanced; H1 is graded strained (W40 was the first supported week on the canonical framing, in which each hypothesis is tested as written, since W37, so there is no streak), H3 and H5 strained on contradicted clauses; Counterpoint's reversal is 8-Hi to 12-Hi; the mathematicians' letter is from the Association for Human Mathematics; p134 is re-pointed to the multi-rack ESUN-on-optics test. Removed as unsourced: a Lumentum 'no cancellation' statement, a Dell'Oro $20B/2029 figure, a 30% QoQ HBM step, a Haiku/Sonnet cyber-safeguard parity claim, a Fitch reference and a Sierra CEO/OpenAI board line.
• 2026-10-10 (corrections, cycle 2): The $42B is described as launched at stated spreads, not priced; the claim that no market has priced a frontier lab standalone is narrowed to no syndicated or public market, and the June private package's $4.5B of unguaranteed B notes at a reported 8.5% is carried as the one prior print, with p129's threshold moved from 8.0% to 8.5% to match; Meta's Vistra PPAs are dated to January, with DOE's October 5 conditional loan as the in-window event, so the week has one new nuclear deal, not two; Haiku 5.5's cut is stated as 90% below 100K tokens and 50% above, the tokenizer effect as 20-23% sooner, and the decision-tier connection is relabelled abductive; PJM's November 16 show-cause response in EL26-67 is added to p130, the ISO-NE connection and the watchlist; the ISO-NE claim is narrowed to the undue-discrimination showing and its confidence lowered; connection 3 scores allocation, not Gilder's falsifier; Mistral Large 4 is 52B-active, matching The Model Pulse; Artificial Analysis's Haiku cost figure is described as provisional and untiered everywhere; the Willison and IFR entries in 'where we differ' are rewritten from the sources; an unverified Cyber Index score for Mistral Large 4 is removed. The published thesis verdicts supersede the pre-review synthesis. Figures used without a graded citation, each labelled at point of use: the June B-note print (PitchBook, The Inference); Microsoft-Decision-1's disclosures and TypeSafe Jev's $0.042 (vendor and outlet pages); Cloudflare Clef-flash at $0.038 (Dbggr, Willison); Artificial Analysis's provisional ~$0.21 Haiku cost per task (via Capital and Compute); Anthropic's 75% blended Haiku estimate and 20% cache-read saving; Willison's 1.25x tokenizer measurement.