Cross-system contagion & interconnection — the channels between the bubbles

Built 2026-06-14. Synthesis of the cited blocks plus primary reconciliations. All edges here are structural overlay (directional transmission links, no cycles) — they enrich the bubble map and connector analysis but are excluded from the SCC / Z3 / TLA+ proofs, which run on the financial layer only.

Every other block maps one subsystem. This one draws the edges between them — the channels along which a shock in one bubble propagates to the rest. Each channel's links are documented facts; the firing of the cascade is contingent (and, for the Fed channel, interruptible). It also canonicalizes three figures that appear in more than one form across blocks, and states the highest-value open questions plainly so they aren't mistaken for settled facts.

1. Crypto → Treasury → banks → AI marks (the missing cascade)

Stablecoins, Treasuries, banks, and AI capex are one system, not four. Mechanism: stablecoin reserves (Tether ~$140B+, USDC) are T-bill-concentrated → a peg break / redemption run forces a T-bill fire-sale → front-end yields spike → the rate move deepens banks' unrealized HTM losses (already ~$300B+ and re-widening, macro-bank-htm-marks) → stressed banks pull credit / raise equity → the marginal lender to the AI/data-center buildout retrenches, pressuring the very marks the core depends on (fin-ai-depreciation-debttrap). Each link is documented; the cascade firing depends on a peg break. Fact (links) + contingent (firing). New edges: Stablecoins→US_Treasuries, Tether→US_Treasuries, US_Treasuries→US_Banks, US_Banks→SINK_lenders (the WLFI↔Binance ~87% USD1 custody concentration is already edged Binance→World_Liberty_Financial in spec-exchanges-asia).

2. Yen carry → the AI core (the hidden funding leg)

Cheap yen is the world's funding currency; as JGB yields escape yield-curve control (0% → ~2.66% in 2026), crowded carry-funded longs unwind (macro-carry-trades). SoftBank — a major OpenAI/Stargate financier — is yen-funded and yen-refinancing-exposed, so a Tokyo rates move is an AI-core funding shock arriving exactly as its commitments peak. Fact (mechanism) + contingent (magnitude). New edges: SoftBank→Stargate, SoftBank→OpenAI (carry-exposure overlay).

3. Chip ↔ private credit (a second circular structure)

Chip vendors are now infrastructure financiers. Broadcom (custom XPUs for Google/Meta/OpenAI/Anthropic + two unnamed customers — analyst-flagged Apple/ByteDance) anchors a ~$35B+ XPU compute-financing platform with Apollo/Blackstone (already edged Apollo_Blackstone→Broadcom in fin-google-amazon-anthropic-meta); Meta funds Hyperion via a Blue Owl SPV. Private credit (macro-cre-privatecredit / macro-private-credit-marks) thus finances the data centers that buy the chips — a second circular structure beside the equity loop, and a second place a private-credit mark-down would bite. New edge: Apple→Broadcom (unconfirmed, weak) ties Apple capex into that stack.

4. China leverage, made legible (node merge)

Merging the fragmented China nodes makes one actor's concentration of chokepoints visible: the same China node gates rare earths/magnets + antimony to the defense-industrial base (spec-rare-earth-statecraft), is the prime suspect in Baltic/Taiwan undersea-cable cuts and the Royal Mint Court comms-siting exposure (geopolitics-cables-space-layer), and runs the export-licensing faucet (geopolitics-chip-chokepoint-war). Fact (each chokepoint) + labeled interpretation (concentration as leverage). New edge: China→Defense_Primes.

5. The Fed circuit-breaker (the missing node)

Every cascade above can be interrupted or amplified by Fed policy, so any unwind thesis that assumes policy inaction is incomplete. If the crypto→Treasury→bank cascade fires, the playbook is a BTFP-style facility (banks borrow at par, suppressing the HTM forced-sale spiral) plus QE / cuts to cap the yield spike. This does not prevent the loss — it socializes and delays it — but it changes the timing and form of any unwind. Read the trigger panel (spec-unwind-timing) against this reaction function. New edges: Federal_Reserve→US_Banks, Federal_Reserve→US_Treasuries.

The Fed reaction-function decision-tree (added 2026-06-16, #57)

The circuit-breaker node, made formal: given which cascade fires, what the Fed does, in what order, gated by what. Tools + precedents are fact; thresholds + branch model are labeled. The binding constraint is the inflation gate — the Z3 "no single rate fits the targets" (macro-fed-trap-regional) rendered as a tree.

The inflation gate (the binding constraint). With CPI ~4.2% (macro-official-data-integrity), easing to backstop the cascade reignites inflation while holding to fight inflation lets the cascade run — the Fed can't cap the yield spike, protect bank capital, and hold inflation with one rate (the Z3 UNSAT result). So the reaction function is asymmetric: liquidity tools B1–B3 fire fast (framed as "plumbing, not stimulus"); the rate-cut lever B4 is inflation-gated and fires slow.

Terminal insight. The Fed never prevents the loss in any branch — it socializes and delays it (B2/B3) or trades inflation for it (B4). So the unwind's timing/form is set by which constraint binds first: a plumbing break (fast, inflationary), a solvency break (par-lending, moral hazard), or a labor break (cuts, gated). Any unwind thesis assuming policy inaction is incomplete; any assuming a clean rescue ignores the inflation gate.

Cross-file reconciliations (canonical values)

• NVIDIA → OpenAI. The "$100B" was a milestone-linked letter of intent (22 Sep 2025), later called "never a commitment" by Huang; the closed figure is $30B direct equity, part of OpenAI's ~$110B round at a $730B pre-money valuation (27 Feb 2026; SoftBank +$30B, Amazon +$50B). Canonical in fin-nvidia-openai.
• Oracle RPO. $455B at end of Q1 FY2026 (quarter ended 31 Aug 2025; reported 9 Sep 2025), +359% YoY (vs ~$99B a year earlier). The $317B is the prior quarter (Q4 FY2025) — so $317B→$455B is the QoQ step and +359% is the YoY rate; no discrepancy once periods are labeled.
• Anthropic Series G. Closed 12 Feb 2026 at $350B pre / $380B post; the round expanded from a ~$20B target to $30B on demand (led by GIC and Coatue). "$350B/$380B" = pre vs post; "$10–30B" = target vs final — one event, not conflicting rounds.

Open questions (genuine gaps, with what resolves each)

• Cancelable-deal terms (SpaceX↔Google $920M/mo; OpenAI↔Oracle/CoreWeave) — contracts / commitment disclosures; sets how fast the robust SCC sheds nodes.
• Google's contingent $30B in Anthropic — what milestones gate it (a round disclosure); a miss could collapse the second funding ring.
• CoreWeave debt maturities (~$21.4B; $6.7B current) — 10-K MD&A; the 2026–27 bullets as the Microsoft revenue share falls below 50%.
• Broadcom's two unnamed XPU customers — Apple and/or ByteDance? (export-control question) — earnings/capex confirmation.
• SoftBank's yen refinancing need 2026–27 vs its AI commitments — debt schedule; sizes the carry-unwind transmission.

What is NOT asserted

• These transmission edges are not new money-flows and do not enter the formal proofs.
• No cascade is asserted to fire — each is a documented mechanism with a contingent (and possibly Fed-interrupted) trigger.
• China-concentration is leverage/exposure, not an assertion of intent to use every chokepoint.
• The reconciliations correct period/stage/labeling confusion; they allege no error by any source.

Sources: NVIDIA–OpenAI 10GW partnership, Fortune — Huang: "never a commitment", Oracle Q1 FY2026 results, Anthropic — $30B Series G at $380B post, TechCrunch — Anthropic Series G.

← Research index · structured data: spec-cross-system-contagion.json · spec-cross-system-contagion.md