AI-capex complex — cluster pass, 2026-09-20

AI-capex complex — cluster pass, 2026-09-20

Drafted by Claude in a supervised session. Nothing here is human-verified. Every figure is sourced; anything from background knowledge is marked [background]. Framework: baker-thesis. Cluster pass carries pipeline stages S0, S4, S6 and S7 across the group; per-company stages (S2, S3, S5, S8, S9) stay with the name passes.

40 of the 122 watchlist names. ai_semis 10, ai_datacenter 8, megacap_tech 6, semis_equipment 4, power_nuclear 4, semis_analog 3, semis_memory 2, energy_storage 2, electronics_mfg 1. Four are stranded from primary documents: 8035.T, IFXD.XC, ORSTED.CO, SLRE.XC.

---

Variant perception

The watchlist holds these 40 as diversified exposure to AI. Measured, they are one position with one point of failure, and the layer most likely to be owned for upside is the only layer that cannot fund itself.

The interesting question is not whether AI capex continues. It is that the complex has quietly split into two populations with opposite balance-sheet physics, and the split is visible in filings rather than in narrative.

---

The funding test — the one that matters

Baker's central reason for thinking this cycle is not 2000 is that the buildout is funded from operating cash flow rather than debt. That is testable. Operating cash flow minus capex, from EDGAR XBRL, $bn:

It holds for three of the four, and breaks completely at Oracle. Capex over five years, same source, $bn:

Microsoft's capex is up ~4.8x in four years and free cash flow has fallen only 10%. Alphabet grew capex 2.9x and free cash flow rose. That is a business paying for its own buildout.

Oracle is not. Capex went 6.9 → 21.2 → 55.7 while operating cash flow went 18.7 → 20.8 → 32.0. Buybacks went 1.2 → 0.6 → 0.1. Baker names the sequence to watch as "issue debt → cancel dividends → stop buybacks." Oracle has done the third. Meta has done the first: long-term debt issuance 8.5 → 10.4 → 29.9.

Amazon is absent above because it does not tag the capex concept used here; the figure needs a different tag and was not resolved this session. That is a gap, not a finding.

---

The layer that cannot fund itself

Same test, ai_datacenter plus the power and storage names, latest tagged annual period, $bn:

Ten of eleven burn cash. DigitalOcean is the only one that does not, and it is the one with no AI-buildout story attached. Roughly $31bn of long-term debt sits across this group against about $13bn of combined annual free-cash-flow burn.

CoreWeave carries 4.7x its cash in long-term debt. Cipher carries 4.3x its cash with negative operating cash flow. This is precisely Baker's NII Holdings rule: leverage converts a survivable thesis error into a permanent one, and "sometimes not everything goes right."

---

Who depends on whom

Customer concentration, structured XBRL from each issuer's own filing:

Marvell tags only channel mix, not named customers. Hut 8 tags concentration outside the supported dimensions. Neither is evidence of no concentration.

The chain, in one line

A handful of hyperscalers fund the whole thing from cash → four of them are 60% of Nvidia and one is a third of Broadcom → the neoclouds resell that compute back to the same buyers, on debt, with 73% to 83% of their revenue or receivables in two or three names.

Every link is concentrated in the same few counterparties. Holding 40 of these names is not diversification. It is the same bet expressed forty ways, and the concentration is multiplicative rather than additive.

---

A seam error in this watchlist's own clustering

Baker's §12 warns about names sorted into the wrong basket. Two sit in this repo's own ai_datacenter cluster on the strength of their narrative:

• IREN: 100% of FY2026 revenue is bitcoin mining. The AI pivot is not in the revenue.
• CIFR: 70% of revenue is one mining pool.

They are bitcoin miners with an AI story and a datacenter footprint, not AI-datacenter businesses. Their revenue correlates with the bitcoin price, not with hyperscaler capex, while their capex correlates with the AI story. That is a genuinely different risk shape and the cluster tag currently hides it. watchlist.yaml should be corrected, and the name passes on both should test the pivot against revenue rather than against announcements.

---

Baker lenses, scored

---

What kills it

One variable: hyperscaler capex. Everything downstream is derivative. The neoclouds cannot refinance $31bn into a slowdown, and their revenue is concentrated in the same buyers whose spending would be slowing. A hyperscaler capex pause does not hurt the neocloud layer proportionally — it hits revenue and refinancing at the same moment.

Unknown unknowns, acknowledged as a category rather than dismissed: an export-control change, a Taiwan event, or a Bitter-Lesson violation would each hit every one of these 40 names simultaneously, because there is no genuine diversification inside the group to absorb it.

---

The single most important thing to watch

Oracle. Not Nvidia, and not the hyperscalers that are still self-funding.

Oracle is the first large spender to cross into sustained negative free cash flow and to stop buying back stock, which is two of the three steps in Baker's stated bubble sequence. It is the test case for whether a company that must finance the buildout rather than self-fund it can keep pace. If Oracle's capex holds at 55.7 while free cash flow stays deeply negative, that is the prisoner's dilemma overriding economics, which is his named condition for the cycle turning into a bubble. If Oracle blinks, the neocloud layer loses its proof of concept.

Gap worth closing: TSMC is not on this watchlist. Baker calls its capacity discipline the single best bubble indicator, and the watchlist holds three of its equipment suppliers (AMAT, LRCX, Tokyo Electron) without holding the customer whose decisions determine their orders. TSM is an SEC filer and would be reachable today.

---

Scorecard

S0 Universe & fit (cluster)   ⚠️ SPLIT — sellers of shortage pass; the debt-funded buyers do not.
                                 Owner exclusion list and competence check still ⏳ OPEN (user).
S4 Industry/supply map        ✅ — chain mapped end to end with filing-sourced concentration at each link
S6 Scuttlebutt                ⬜ — not attempted this pass
S7 Thesis + framework test    ✅ (partial) — baker-thesis applied; valuation cross-section UNKNOWN
                                 (ROIC free tier is annual only); proxies/incentives not examined

➡️ Next action: correct the IREN and CIFR cluster tags in watchlist.yaml, and add TSM.

⚠️ Open: cross-sectional valuation needs a source with current multiples. Amazon capex needs the right XBRL tag. Proxy and incentive analysis not started.

Process support, not investment advice.

---

Corrections issued after this file was written

Recorded here rather than by rewriting the text above, per CLAUDE.md's rule on correcting rather than silently amending.

• 2026-09-21, slot names-ai-capex-1 — the AVGO row in "Who depends on whom" (one customer = 32% of FY25 revenue) is a year stale. The Q3 FY26 10-Q puts one distributor customer at 50% of net revenue for the quarter and the top five end customers at ~55%. Sourced with its locator in ledger/research/AVGO/dossier-2026-09-21.md.
• 2026-09-21, slot cluster-ai-capex — the S6 ⬜ — not attempted this pass line is superseded for all 40 members by _clusters/cluster-ai-capex-2026-09-21.md, which also scores S0/S4/S7 for 9984.T, IREN and CIFR and records the IREN/CIFR cluster-tag correction this file asked for as closed.

• 2026-09-21, slot names-ai-capex-2 — two additions to "Who depends on whom", which this file left open for both names:
• MRVL. "Marvell tags only channel mix, not named customers" is right, and the channel mix is now the finding: one distributor took 44% of Q2 FY27 net revenue against 34% a year earlier (45% vs 35% for the half), plus one direct customer at 16%; the top ten were 82% of FY26 revenue and four customers were 73% of gross receivables. Sourced with locators in ledger/research/MRVL/dossier-2026-09-21.md.
• ARM is absent from this table and should not be. Its top five customers — a list that includes Arm China and SoftBank Group themselves — were 57% of FY26 revenue, and revenue from related parties was 30% of the total, supplying 74% of the year's growth. Qualcomm, a litigation adversary, was another 9%. Sourced with locators in ledger/research/ARM/dossier-2026-09-21.md.
• The chain sentence ("a handful of hyperscalers fund the whole thing from cash") holds for the US names, but ARM is the one member of the 40 whose largest counterparty is its own controlling shareholder rather than a hyperscaler. A future cluster pass should carry that as a separate row rather than folding it into customer concentration.