AI-capex complex — cluster pass 2: S6 for 41 names, S0/S4/S7 for three, 2026-09-21

AI-capex complex — cluster pass 2: S6 for 41 names, S0/S4/S7 for three, 2026-09-21

Slot cluster-ai-capex. Drafted by Claude in an unattended cloud routine. Nothing here is human-verified. Every figure carries its source; anything from background knowledge is marked [background]. Search-result evidence is marked [search-summary] — it is evidence of the summary kind, not a primary document, exactly as the skill's rung 6 requires.

This pass does three things:

1. S6 (scuttlebutt) for all 41 members. Gate: one piece of evidence per member that did not come from the company. Met for 41 of 41; the thin ones are named in §3.
2. S0, S4 and S7 for the three members the 2026-09-20 pass did not score: 9984.T, IREN, CIFR. The other 38 carry forward S0/S4/S7 from _clusters/ai-capex-2026-09-20.md.
3. One correction to that file, and one of its two next-actions closed.

The product test and the expert call stay ⏳ OPEN (user); S7 conviction stays ⏳ OPEN (user).

---

§1 — S6, part one: the board is the cheapest scuttlebutt in the building

Ten of the 41 already carry independent commentary inside this repo. ledger/pipeline/gold/ holds 371 claim rows extracted from tracked podcasts; every row is a named outside investor or analyst speaking about the name on the record, with a transcript locator. Under the skill's own rule ("podcast rows already on the board — a claim row is free scuttlebutt") this is S6 evidence that cost nothing and was located rather than typed.

Quotes copied from the claim files, not transcribed. Each row carries its own source.url, timestamp_start and speaker id.

Cluster-level, same source. ai-capex-cycle carries 31 rows, semiconductors 12, us-megacap-tech 10, ai-frontier-labs 10. They are S6 for the group rather than for a name, and they are notably two-sided, which matters because the 2026-09-20 pass concluded the complex is one position:

• Paul Kedrosky, 2026-08-28: "whenever these things become financialized and it becomes divorced from the underlying application, you need to pay a lot of attention. And that's the moment we're in right now." (candidates/meb-faber-2026-08-28-paul-kedrosky-ai-first-01-ai-capex-cycle.yaml)
• Jared Dillian, 2026-09-03: "In my lifetime, this is the first time I've seen tech being financed with debt." (candidates/monetary-matters-2026-09-03-i-m-insanely-bullish-on-bonds-jared-dillian-09-ai-capex-cycle.yaml)
• Cameron Dawson, 2026-09-06, on the circular-financing denials: "If things were as great as you say that they are, you should not have to be playing the games that you are playing." (candidates/excess-returns-2026-09-06-nvidia-is-betting-on-its-customers-gen-z-is-01-ai-capex-cycle.yaml)
• Against them, Brian Singer (Goldman), 2026-09-01: "Hyperscaler spending for 2029 was projected at 1.5 trillion. It's today at 2.1 trillion." (candidates/exchanges-2026-09-01-the-outlook-for-data-center-power-demand-as-01-ai-capex-cycle.yaml)

The outside-in view independently reached this pass's own conclusion, from different evidence: the financing, not the demand, is the argument. That is a convergence worth recording because the 2026-09-20 file derived it from XBRL and these speakers derived it from watching the capital markets.

---

§2 — S6, part two: the other 31, from outside the company

Rung 6 (WebSearch). Each row is one piece of evidence that is not the company talking. All rows here are [search-summary] unless marked otherwise.

§3 — What the body of scuttlebutt says, and where it is thin

• Four of the 41 carry a published short thesis: CRWV and OKLO (Kerrisdale), TE and EOSE (Fuzzy Panda). All four sit in the debt-funded or pre-revenue half of the complex that the 2026-09-20 pass separated out. None of the ten self-funding names at the top has one. The outside world is short the same layer this pass's balance-sheet test flagged.
• The independent evidence bites hardest one tier below the famous names. The findings with real information content this pass are MPWR (Edgewater, socket reallocation), MRVL (Benchmark, Trainium 3/4 to Alchip), CBRS (86% of revenue from two UAE entities) and SNDK (no moat at 27x). Each is a specific, checkable claim about a company's position — worth more than another view on whether AI is a bubble.
• Employee-side evidence is thin across the cluster. Only three Glassdoor readings were obtained (AVGO 3.4, APH 3.9, AEIS 3.5). No Blind, G2 or Reddit sweep was run for the semis names; G2 was only informative for DOCN, which is the one member selling to developers.
• Thinnest rows, stated plainly: AEIS, LFUS, SLRE.XC — analyst opinion and a rating only, no independent operational evidence. APH is close behind. These four pass the gate on the letter of it and should be re-run when a name pass reaches them.
• Nothing found outside-in contradicts the top of the chain. The bear case on MSFT/GOOG/AMZN is about rate of change and depreciation, not about the businesses funding themselves. That is consistent with the 2026-09-20 funding test rather than a new finding.

---

§4 — 9984.T SoftBank Group: S0, S4, S7

Cluster tag is holding_co, not japan_*, so the standard AI-surplus Stage 0 applies, per the skill §4 ("Names outside japan_* clusters … SoftBank … use the standard test").

Primary evidence: EDINET DB, E02778, FY2026 annual (IFRS) and the event feed 2026-06-01 → 2026-09-21.

S0 — Universe & fit ❌ FAILED, 2/4

Surplus question — who keeps it? Not SoftBank. It is a levered claim on other companies' surplus, and the market says so: a ~46% discount to stated NAV, read by outside analysts as pricing the debt rather than the assets [search-summary]. On the company's own IFRS numbers the leverage is not disputed: equity ratio 29.0%, D/E 1.46, free-cash-flow yield −24.1%, EDINET health score 40/100 (FY2026, get_company E02778).

Falsifier. A markdown of the private AI positions. The NAV is a mark-to-model on OpenAI and SVF2 holdings funded with public bonds; OpenAI ruling out a 2026 listing took ~11% off the shares in a day [search-summary]. One adverse revaluation moves both sides of the leverage at once.

Liquidity. ADV ¥348bn/day (mean of close × volume, 21 sessions to 2026-09-18, ROIC TSE:9984; mean 57.7m shares). No constraint at any plausible size.

Verdict: ❌ FAILED at 2/4. The failure is informative rather than disqualifying: SoftBank is not a surplus capturer, it is the complex's funder, and the pipeline should hold it as that.

S4 — where it sits on the chain

The 2026-09-20 file mapped supplier → company → customer and found every link concentrated in the same counterparties. SoftBank sits one layer above that map, on the capital side, and its own filings show the funding pattern the cluster thesis turns on. From the EDINET event feed:

FY2026 scale, for context (get_company E02778): revenue ¥7.80tn, net income ¥5.00tn, total assets ¥60.7tn, net assets ¥17.6tn, PER 4.1. The net income is dominated by investment revaluation, which is why the P/E is 4 and the free-cash-flow yield is −24%.

Chain position in one line: the hyperscalers fund the buildout from operating cash flow; SoftBank funds its share of the same buildout in the bond market, and the asset it is funding is an equity stake in the buyer — OpenAI — whose failure is the named kill condition for Oracle in the 2026-09-20 file.

S7 — thesis, bear case first, kill criteria

Bear case first. SBG is the single most levered expression in the cluster of the one counterparty risk the rest of the complex already carries. Its net asset value is marked on private AI positions; its funding is public, serial and increasingly structured (a hybrid with deferrable coupons is not ordinary debt). The market's ~46% NAV discount is not a valuation anomaly to be arbitraged — it is the price of that mismatch.

Variant perception. The watchlist plausibly holds 9984.T as Japan exposure or as an Arm proxy. Measured, it is neither: it is the cluster's funder-of-last-resort with a 29% equity ratio and a −24% free-cash-flow yield, and roughly 70% of its NAV in one listed asset whose repricing power was just curtailed in court [search-summary]. Holding it alongside NVDA, AVGO, ARM and ORCL is not diversification within the AI trade; it is the same trade with leverage on top.

Kill criteria (specific, testable).

1. A write-down of the SVF2 / OpenAI carrying value in any quarterly report.
2. A fourth and fifth bond issue in a single quarter, or a second hybrid with deferred interest — the funding cadence, watchable monthly in get_events.
3. Arm royalty guidance reset downward after the Qualcomm ruling, which would hit the ~70% pillar.
4. The NAV discount widening past ~55% while the asset marks are unchanged — the market repricing the leverage rather than the assets.

Time horizon. 2–3 years, matched to the bond maturities being created now. Framework used: the lens table from _clusters/ai-capex-2026-09-20.md (Baker-style), not the account baker-thesis skill, which is not available in routine runs. On "seller vs buyer of the shortage," SoftBank is the purest buyer, funded with debt — the category that file scored Weak. Conviction: ⏳ OPEN (user).

---

§5 — IREN: S0, S4, S7

Primary evidence: 10-K for FY ended 2026-06-30, filed 2026-08-27 (https://www.sec.gov/Archives/edgar/data/0001878848/000187884826000052/iren-20260630.htm), located by exact substring in the filing text.

S0 — Universe & fit ❌ FAILED, 2/4

Surplus question. The customer keeps it. IREN rents GPU-hours under fixed multi-year contracts; if inference costs fall, the saving accrues to Microsoft and NVIDIA at renewal, and IREN's own filing lists hyperscalers and CoreWeave, Nebius, Crusoe, Lambda, Nscale and SpaceX as competitors in the same sentence. Pricing power sits with whoever holds the scarce input, and today that is power, not the reseller of it.

Falsifier. The 2026-09-20 file's finding, unchanged by this pass: 100% of FY2026 revenue was bitcoin mining (10-K, 2026-08-27). The AI revenue is contracted, not yet earned. The specific break is a Horizon 2-4 delivery slipping past the stated grace periods, which converts the contract from an asset into a liability — and the 10-K already records equipment delays having "resulted in delayed delivery of certain equipment required to meet our obligations with respect to the delivery of Horizon 1."

Liquidity. ADV $1.73bn/day (mean close × volume over 25 sessions to 2026-09-18, ROIC NASDAQ:IREN; mean 41.3m shares). No constraint.

Verdict: ❌ FAILED at 2/4 — and the two points it does score are the two that a hyperscaler could replicate by signing a different landlord.

S4 — chain position

IREN is not in the same box as CoreWeave and Nebius, though the 2026-09-20 table grouped them by cash burn. The neoclouds buy GPUs and sell compute; IREN sells contracted capacity with acceptance milestones, and its counterparties are the two largest names in the complex.

The new finding this pass produces: NVIDIA is now IREN's supplier and its customer at once. The May 2026 contract is NVIDIA buying $3.4bn of cloud services "to support its internal AI and research workloads" from a company whose entire fleet it sells into. The names-ai-capex-1 slot found the three big sellers underwriting their own buyers (NVDA's capped guarantee and take-or-pay, AMD's warrants, AVGO's XPV co-sponsorship). This is the next turn of the same wheel: the seller is now renting capacity back from a buyer it supplies. Whatever the commercial logic, it means the revenue line and the supply line on this name point at the same counterparty.

S7 — thesis, bear case first, kill criteria

Bear case first. H.C. Wainwright's Sell on execution risk and "irrational exuberance" [search-summary]; the co-CEO RSU grants of $800m–$1.1bn with limited performance hurdles [search-summary], which is a governance flag on a company whose equity is its currency; and the plain fact that not one dollar of FY2026 revenue was AI.

Variant perception. The bull case is the $13.1bn of contracted value. The variant reading is that this is a construction company with a delivery schedule, not a cloud business: the contract value is real, and so is the grace-period clock, and the two are the same object viewed from opposite ends. The market prices the contract; the filing prices the risk of missing it.

Kill criteria.

1. Horizon 2 slipping past its grace period, or any disclosed amendment to the Microsoft delivery schedule.
2. FY2027 Q1/Q2 revenue mix still showing AI below a third of revenue.
3. A further equity raise to fund the "$5.8 billion in GPU purchases still ahead" [search-summary] rather than drawing on prepayments.
4. NVIDIA not renewing or extending beyond the initial five-year term.

Time horizon. 18 months — this resolves on a delivery schedule, not a cycle. Conviction: ⏳ OPEN (user).

---

§6 — CIFR: S0, S4, S7

Primary evidence: 10-Q for the quarter ended 2026-06-30, filed 2026-08-04 (https://www.sec.gov/Archives/edgar/data/0001819989/000181998926000041/cifr-20260630.htm), located by exact substring.

S0 — Universe & fit ❌ FAILED, 1/4

Surplus question. Not Cipher. It earns a contracted lease spread on land, power and shells, and it paid for the tenant's creditworthiness with equity. If AI compute gets cheaper, the lease is unchanged and the value accrues to Fluidstack's customers.

Falsifier. The revenue that services the debt is still mining. The 2026-09-20 file recorded Foundry USA Pool at 70% of 2025 revenue (10-K, 2026-02-24), and the AI lease commences September 2026. Against that, the 10-Q shows the capital structure already built for the pivot: "$172.5 million of principal amount of convertible notes due in 2030 with an interest rate of 1.75%", "$1,300.0 million of principal amount of convertible notes due in 2031 with an interest rate of 0.00%", plus a revolver whose maturity springs if the 2030 notes are still outstanding 91 days before they mature. Zero-coupon converts are equity risk wearing a bond's clothes.

Liquidity. ADV $0.58bn/day (21 sessions to 2026-09-18, ROIC NASDAQ:CIFR; mean 35.1m shares). Adequate, and the thinnest of the three.

Verdict: ❌ FAILED at 1/4 — the weakest Stage 0 in the cluster.

S4 — chain position

Cipher is a landlord whose tenant's rent is guaranteed by a hyperscaler. That is a genuinely different risk shape from CoreWeave's and it should not be grouped with it: the credit exposure runs to Google, not to Fluidstack, for $1.73bn of the obligations [search-summary, confirmed in direction by the $544.5m backstop asset in the 10-Q]. What Cipher retains is construction risk, power-delivery risk and the mining business that currently pays the bills.

This also completes a pattern across the three new members: at every layer of this complex, the party with the strong balance sheet is underwriting the party with the weak one — NVIDIA and AMD for their customers (slot names-ai-capex-1), Google for Cipher's tenant, SoftBank for OpenAI. The 2026-09-20 file's split into "self-funding sellers" and "debt-funded buyers" is now better described as one credit chain with the hyperscalers at the root of it.

S7 — thesis, bear case first, kill criteria

Bear case first. 70% of revenue is one mining pool; the AI lease has not commenced; $1.47bn of converts sit against it; and 5.4% of the equity was spent buying a tenant's credit rating.

Variant perception. The market reads the Fluidstack deal as Cipher becoming an AI company. The filing reads as Cipher becoming a contractor to a Google-backstopped obligation — a better business than mining, and a much smaller one than the announced contract values imply, because the party taking the AI risk is the party that gets paid for it.

Kill criteria.

1. Barber Lake commencement slipping past the stated September 2026 date.
2. Any amendment reducing or terminating the Google backstop.
3. Bitcoin revenue concentration unchanged at the FY2026 10-K — i.e. the pivot still not in the revenue a year after it was announced.
4. A further convertible issue before the lease commences.

Time horizon. 12–18 months. Conviction: ⏳ OPEN (user).

---

§7 — Corrections to _clusters/ai-capex-2026-09-20.md

1. AVGO customer concentration is stale in that file and understates the position by 18 points. It records "One customer = 32% of FY25 revenue." The Q3 FY26 10-Q (quarter ended 2026-08-02), read in slot names-ai-capex-1 and quoted with its locator in ledger/research/AVGO/dossier-2026-09-21.md, puts one distributor customer at 50% of net revenue for the quarter and the top five end customers at ~55%. The arc is 21% → 28% → 32% → 46% → 50%. Use the AVGO dossier, not the 2026-09-20 table, for this number.
2. The IREN and CIFR cluster tags are already corrected. That file's next action asked for it; watchlist.yaml now carries bitcoin_miner_ai_pivot for both. Closed.
3. TSM is still not on the watchlist. That file's second next action. It is an owner decision (the 122 names are confirmed, §17), so this pass does not add it — it is restated as an open item below.
4. Its S6 ⬜ — not attempted this pass line is superseded by this file for all 40 of its members.

---

Cluster scorecard — cluster-ai-capex, 41 members

S0 Universe & fit        ⚠️ SPLIT (38 carried forward: _clusters/ai-capex-2026-09-20.md)
                            + 9984.T ❌ 2/4, IREN ❌ 2/4, CIFR ❌ 1/4 scored this pass.
                            Owner checks dropped per WATCHLIST_PLAN.md §16 — not asked, not open.
S4 Industry/supply map   ✅ — 2026-09-20 chain extended: capital layer (9984.T) and the
                            backstopped-landlord layer (IREN, CIFR) added; NVIDIA found on both
                            sides of IREN's P&L.
S6 Scuttlebutt           ✅ — 41 of 41 carry one piece of non-company evidence. 10 from board
                            claim rows, 31 from WebSearch (summary-kind, marked). Four names
                            carry published short theses. Thin: AEIS, LFUS, SLRE.XC, APH.
                            Product test / expert call ⏳ OPEN (user).
S7 Thesis + framework    ✅ (partial) — three new members given variant perception, bear case
                            first and testable kill criteria. Cross-sectional valuation still
                            UNKNOWN (ROIC free tier is annual only). Conviction ⏳ OPEN (user).

➡️ Next action: the cross-sectional valuation test has now been deferred twice for the same reason. Either accept that it cannot run on this source ladder and say so permanently, or spend Equibles GetValuationMultiples calls on a fixed basket during a name pass.

⚠️ Open for the owner: TSM is still absent from the watchlist. The 2026-09-20 file called its capacity discipline the best single bubble indicator for this group, and the list holds three of its equipment suppliers (AMAT, LRCX, 8035.T) without the customer whose orders they depend on. The list is owner-confirmed, so nothing was added; this is a one-line decision only the owner can make.

Process support, not investment advice.