p(doom) evidence — Node 7: compute & cloud chokepoints

Each section: Claim · Why · Evidence · Analogue · Would update if · Conf (H/M/L).

Parent timeline: node7 compute cloud
Master: Shared Ci spine
Date: 2026-07-04
Settings: Hybrid time (C); modal + tail branches

1. Where this fits in the repo

File	What’s there
This file	Claim \| Why \| Evidence for every major Node 7 probability (~22 sections)
node7 compute cloud	Full node — definition, timing, modal/tail, actors, p(doom), falsifiers
node3	Cloud KYC P≈0.55; DPA seizure; EU sovereign AI — overlapping levers
research_ai_pause_advocacy_playbook.md	Cloud registration near-term win; HEM 3–5 yr; DPA hard version
HEM_hardware_enabled_mechanisms.md	Verified compute attestation; VEU datacenter mandate
Anthropic_Fable_Mythos_export_ban_2026_深度解读.md	Live kill switch on frontier API
RAND_2024_Securing_AI_Model_Weights_导读.md	SL1 = cloud default; concentration = attack surface
correlation matrix	Cluster A race — do not double-count with N3/N4/N6

Claim: Conditional on C2–C5 window, ~58% of worlds follow modal: AWS/Azure/GCP retain ≥70% US frontier FLOPs; cloud KYC/export rules tighten; no durable training stop.

Why: Same playbook pattern as Nodes 2–4 — physical/export layer moves, compute cap dead. Hyperscaler capex ($200B+ AWS, Stargate) creates too-big-to-pause coalition. Fable shows gov can cut access but chose narrow export frame, lifted partially in 2 weeks.

Evidence:

Synergy Q1 2026: triopoly ~63–68% global IaaS — frontier more concentrated than average
research_ai_pause_advocacy_playbook.md §1.2 — compute export viable; training cap not
Fable Jun 2026 partial restore — not sustained global ban
Node 4 modal 0.58 — same human-response reference class

Analogue: Post-9/11 airline security — infrastructure rules tighten; industry keeps flying.

Would update if: ≥2 labs public halt >90d citing cloud/gov gate (not just rhetoric).

Conf: M

P = 0.82 — US frontier ≥70% on AWS/Azure/GCP through 2028

Claim: At least 70% of US frontier training FLOPs run on the triopoly through 2028-12.

Why: Labs signed multi-year GW deals (Anthropic–AWS 5GW Mar 2026); OpenAI Stargate still uses hyperscaler + Oracle; self-build partial (Abilene) not full migration by 2028. Neocloud captures overflow, not majority.

Evidence:

SQ Magazine / AWS stats Mar 2026 — Anthropic 5GW Trainium commitment
CSA 2025 report — triopoly overwhelming majority of AI-grade enterprise compute
(internal note) — Stargate 2 buildings op / 8 building
OpenAI multi-cloud pattern (Azure primary + Oracle Stargate)

Analogue: Mobile OS duopoly — alternatives exist but not for frontier scale.

Would update if: Meta or xAI announces >50% training on non-triopoly owned DC with verified FLOPs.

Conf: M–H

P = 0.55 — Federal cloud training registration / KYC by 2028

Claim: US requires registration or KYC for large cloud training runs (BIS threshold or ~10²⁵ FLOP class) by 2028-12.

Why: Playbook lists as near-term win #3; BIS May 2025 explicitly flags IaaS training for D:5 parties; Jan 2026 chip rule requires KYC procedures from exporters; House Remote Access Security Act (Jan 2026) expands cloud under EAR. Trump EO 2026 acceleration frame still natsec-hawk on CN access.

Evidence:

BIS AI training policy statement — IaaS + knowledge standard
Mayer Brown Jan 2026 — remote end-user disclosure
node3 §10 — cloud KYC P≈0.55 (same lever)
research_ai_pause_advocacy_playbook.md §1.3 items 3, 126

Analogue: Bank Secrecy Act for large wire transfers — reporting, not prohibition.

Would update if: Explicit Trump admin repeal of IaaS guidance + no replacement by 2028.

Conf: M

P = 0.60 — BIS enforces IaaS / training catch-all (knowledge-based)

Claim: ≥1 major enforcement action or licensing denial for cloud-facilitated AI training to D:5 entities by 2028.

Why: May 2026 guidance re-confirms Nov 2023 D:5 HQ rule; DOJ/BIS diversion push on chips signals will to enforce. Gap: proof of knowledge hard — expect case-by-case, not blanket ban.

Evidence:

BIS May 31, 2026 guidance PDF
MoFo Feb 2026 — compliance expectations ↑ across AI chip ecosystem
Node 3 export ctrl P≈0.70 — same policy cluster

Analogue: Huawei Entity List enforcement — selective, high-impact.

Would update if: Zero enforcement actions despite public evidence of CN training on US cloud through 2028.

Conf: M

P = 0.75 — China blocked from US cloud frontier training (effective)

Claim: Effective block on PRC-headquartered entities using US hyperscaler for frontier-scale training by 2028.

Why: Already legal requirement; hyperscalers implement D:5 screening; smuggling tail not modal. Jan 2026 certifications require no prohibited remote access.

Evidence:

BIS 2025–26 policy stack (above)
node3 — chip-hawk bipartisan baseline
AI 2027 CDZ — CN domestic training assumption

Analogue: Semiconductor equipment embargo — leakage exists but not at frontier scale modal path.

Would update if: Documented DeepSeek-scale run on US cloud with no enforcement response.

Conf: H

P = 0.35 — Repeatable gov API / deployment recall (Fable-class)

Claim: US executive uses EAR / EO to recall or gate a frontier deployed model (API) ≥1 more time by 2028, beyond Fable/Mythos.

Why: Precedent live; industry pushback strong (freefable.org, partial lift); selective not systematic. 35% = more than once-off, less than default.

Evidence:

Anthropic_Fable_Mythos_export_ban_2026_深度解读.md — full timeline
Nate Soares / Yudkowsky — gov can act “out of nowhere”
Lawfare / Dean Ball — industry calls controls “cartoonish”

Analogue: FDA drug withdrawal — rare, high-profile, not industry shutdown.

Would update if: Second recall within 2026; or explicit no-more-recalls Commerce policy.

Conf: M

P = 0.45 — DPA for grid / transformers (not datacenter seizure)

Claim: DPA Section 303 or successor used to prioritize domestic grid equipment for AI power stack by 2028.

Why: Apr 20, 2026 Presidential Determination already identifies transformers, substations, etc. as defense-critical; binding constraint for GW DCs is power, not GPUs. Distinct from seizing OpenAI DC.

Evidence:

Atlas Peak Research — Apr 2026 DPA determination
CSA report — grid + cooling as systemic risk
Azure May 2026 outage — cooling/power not cyber

Analogue: COVID DPA for PPE — supply chain, not seizing factories.

Would update if: DPA determination rescinded or grid crisis resolves without further action.

Conf: M

P = 0.06 — DPA frontier datacenter seizure / mandatory training pause

Claim: US government seizes or halts a frontier AI datacenter under DPA or emergency authority for >30 days by 2028.

Why: Node 3/4 tables: P(DPA seizure)=0.03–0.08; Yudkowsky Only Law path not mainstream DC. Stargate = political asset; seizure triggers market + ally crisis. Slightly higher than Node 4 alone because concentration makes target legible.

Evidence:

research_ai_pause_advocacy_playbook.md §1.2D — DPA as hard pause lever
node4 — P(DPA seizure)=0.03 modal
Fable — gov chose export recall, not DC seizure

Analogue: Truman steel seizure 1952 — struck down; high bar.

Would update if: Pentagon public contingency plan for DC seizure; or actual seizure attempt.

Conf: L–M

P = 0.10 — Tail T7-A: Outage delays frontier training >30 days

Claim: ≥1 hyperscaler regional incident causes ≥1 US frontier lab to delay a major training run >30 days (2026–2028).

Why: Oct 2025 + May 2026 AWS, May 2026 Azure show multi-day regional failures; SageMaker/GPU pools affected. 30d requires bad luck (wrong region + no failover + grid compound). CSA: 90% enterprises lose >$300k/hr downtime — labs multi-region mitigate.

Evidence:

Analogue: TSMC earthquake delay — weeks, not years; rare.

Would update if: 2026–2027 second >14d outage in same region with documented frontier delay.

Conf: L–M

P = 0.12 — Tail T7-B: Gov gate on frontier training runs (beyond export)

Claim: US requires prior approval for frontier training runs above FLOP threshold on US soil (not just foreign-national access) by 2028.

Why: Higher than DPA seizure (0.06) because registration is softer — EO 2026-06 30-day pre-brief is proto-gate. Full gate needs statute or EO expansion; industry fights. 12% = tail, not modal.

Evidence:

Fable trigger: Fable released 7 days after policy, not 30 — gap invites gate
Anthropic_Fable_Mythos_export_ban_2026_深度解读.md §3 — 8/1 covered-model framework deadline
Playbook — training licensing medium horizon, not dead

Analogue: FDA IND for clinical trials — gate without banning research field.

Would update if: 8/1 2026 framework mandates pre-approval for training; or GAAIA adds cloud training license.

Conf: L

P = 0.15 — Tail T7-C: EU/India materially decouple >25% frontier spend off US hyperscalers

Claim: By 2029, >25% of EU + India government-linked frontier AI compute spend runs on non-US-hyperscaler sovereign stacks.

Why: CADA mandates sovereignty assessment; €180M tender Apr 2026; India G42 deal May 2026. Commercial labs may stay on AWS; threshold is government-linked + material (25%, not 100%).

Evidence:

Analogue: GDPR → local cloud — partial decouple, not autarky.

Would update if: CADA withdrawn; or EU explicitly allows US hyperscaler for all GPAI training.

Conf: M

P = 0.50 — EU sovereign cloud >30% public-sector AI compute by 2028

Claim: EU institutions + member states procure >30% of public-sector AI compute from SEAL-certified EU sovereign providers by 2028.

Why: Commission already awarded €180M / 6yr tender; mandate for institutions; member state adoption slower. 50% = institutional more than commercial.

Evidence:

IP/26/833 — four providers, diversification explicit
Proximus–S3NS (Google JV) — sovereign wrapper on US tech
node3 — EU sovereign AI spend P≈0.60 (broader)

Analogue: EU Galileo vs GPS — parallel, not replacement.

Would update if: Tender cancelled or undersubscribed.

Conf: M

P = 0.40 — India national AI off US hyperscaler for sovereign projects

Claim: Indian government/national frontier projects (CDAC-class) run primarily on non-US sovereign DC (G42, domestic) by 2028.

Why: G42–Cerebras May 2026; parallel to commercial AWS/Azure ($17B+ MS, AWS investments). DPDP data residency pressure.

Evidence:

Rest of World May 2026 — CDAC + Core42, data under Indian rules
Paris 2025 / India AI Summit — sovereign AI frame
research_ai_pause_advocacy_playbook.md — Global South anti-pause, pro access

Analogue: India UPI — domestic rail, foreign cards still exist.

Would update if: National project hosted on AWS Mumbai exclusively with no sovereign duplicate.

Conf: M

P = 0.08 — Tail T7-D: HEM / verified compute pilot deployed

Claim: ≥1 US export-controlled or VEU datacenter runs FlexHEG-class attestation on frontier training workload by 2028.

Why: HEM R&D not product (HEM_hardware_enabled_mechanisms.md); playbook puts hardware licensing 3–5 yr. 8% = pilot, not mandate. FlexHEG + gov interest on export scenarios.

Evidence:

HEM_hardware_enabled_mechanisms.md — VEU datacenter mandate scenario
research_ai_pause_advocacy_playbook.md §1.3 item 8 — HEM kill switches
AI 2027 Appendix S — compute pause immature 2026

Analogue: Smart meter pilots — years before national roll-out.

Would update if: Commerce mandates HEM for licensed chip exports to allies.

Conf: L

P = 0.12 — Tail T7-E: Cloud controls loosen for allies / accelerate build

Claim: Post-2027 policy explicitly eases cloud/chip access for allied sovereign AI while tightening CN — net acceleration of Western frontier capacity.

Why: Jan 2026 case-by-case China chip review = dual track; Stargate + DPA grid = build more; Trump AI Action Plan acceleration frame. 12% as distinct tail from modal (modal already includes mild acceleration).

Evidence:

Mayer Brown Jan 2026 — case-by-case H200 to China under certs
(internal note) — competition frame won
Node 3 T3 race accelerates P≈0.15–0.25 — subset is cloud-specific

Analogue: Marshall Plan + COCOM — allies in, adversaries out.

Would update if: Unified global cloud training freeze proposed seriously in DC.

Conf: M

P = 0.85 — Frontier labs maintain multi-cloud portfolio

Claim: All major US frontier labs use ≥2 of {AWS, Azure, GCP, CoreWeave, Oracle} for training or inference by 2028.

Why: Capacity crunch; Fable single-provider risk; Anthropic AWS + Google; OpenAI Azure + Oracle + CoreWeave. Risk management, not antitrust.

Evidence:

CSA — concentration risk cited by CISOs; labs respond with multi-cloud
Tech Insider 2026 — 89% enterprise multi-cloud

Analogue: Airlines multi-hub — still oligopoly suppliers.

Would update if: Lab exclusive single-cloud contract leaked with no fallback.

Conf: M–H

P = 0.55 — Partial self-build (Stargate / lab-owned DC) for ≥1 major lab

Claim: OpenAI, Anthropic, Meta, or xAI operates owned or dedicated GW-scale DC capturing >15% of that lab’s training FLOPs by 2028.

Why: Stargate Abilene; Meta $125–145B/yr DC capex; partial not majority by 2028. 55% = any major lab meaningful self-build.

Evidence:

(internal note) — Stargate structure
AWS stats — Anthropic still 5GW on Trainium (self-build complements)

Analogue: Apple private CDN — partial bypass of carrier oligopoly.

Would update if: Stargate cancelled; or lab 100% cloud by 2028.

Conf: M

P = 0.70 — Neocloud (CoreWeave / Oracle) wins ≥1 frontier training contract

Claim: By 2028, ≥1 of {OpenAI, Anthropic, Google DeepMind, Meta AI} has public frontier training on CoreWeave or Oracle Stargate-adjacent cluster.

Why: OpenAI–CoreWeave public; capacity bottleneck on hyperscaler GPU; Oracle Stargate explicit for OpenAI.

Evidence:

(internal note) §CoreWeave
CSA — neoclouds “inch higher” in share

Analogue: TSMC vs in-house fab — overflow foundry.

Would update if: All frontier runs confirmed triopoly-only through 2028.

Conf: M

P = 0.18 — Mandatory hyperscaler resilience / stress-test standard

Claim: US binding rule requiring frontier customers’ cloud providers to meet AI-specific outage resilience standard by 2028.

Why: Outages repeated but no FAA-for-cloud yet; industry prefers voluntary; 18% needs Tail T7-A fire + DC hearing.

Evidence:

AWS Oct 2025 — 70+ services; no regulation
Node 1 P(hearings)=0.75 — could extend to cloud

Analogue: NERC CIP for grid — after big blackout.

Would update if: GAAIA or EO adds cloud BC title.

Conf: L

P = 0.35 — Congress passes cloud training reporting law

Claim: Statute requires reporting large cloud training runs to Commerce/DoC (not necessarily cap) by 2028.

Why: Lower than executive KYC (0.55) because Congress slow; Remote Access Security Act passed House Jan 2026 — possible. Reporting ≠ cap — Cruz coalition might accept.

Evidence:

Mayer Brown — House Remote Access Security Act
GAAIA draft — audit/reporting theme
Node 4 P(GAAIA-like passes House)=0.25

Analogue: SAR bank reporting — transparency without ban.

Would update if: GAAIA dies with no cloud title substitute.

Conf: M

P = 0.25 — EU bans US hyperscaler for sensitive GPAI training

Claim: EU member state or EU-wide rule prohibits GPAI frontier training on non-SEAL-3+ cloud for defined sensitive use cases by 2028.

Why: CADA enables risk assessment → mandate; not yet full ban; 25% = sectoral (defense, gov), not Llama fine-tunes.

Evidence:

CADA COM(2026) 503 — member states must assess sovereignty risk
EU AI Act + GPAI — enforcement 2025–26

Analogue: EU Schrems II — sectoral data localization.

Would update if: CADA softens to voluntary labels only.

Conf: M

P = 0.65 — Hyperscalers offer EU “sovereign SKU” (S3NS-class)

Claim: ≥2 of {AWS, Azure, GCP} market EU sovereignty-compliant AI cloud offering with EU data boundary by 2028.

Why: Proximus–S3NS already in tender; commercial response to CADA; keeps US hyperscaler in EU market.

Evidence:

EU IP/26/833 — S3NS (Thales–Google) awarded
Microsoft EU cloud sovereignty initiatives (2024–25 pattern)

Analogue: AWS GovCloud — segmented product.

Would update if: EU rejects US-JV sovereign wrappers as non-compliant.

Conf: M

P = 0.08–0.12 / yr — Regional hyperscaler outage >7 days (any cause)

Claim: Annual probability of ≥1 AWS/Azure/GCP region >7 days impaired for GPU/training services.

Why: 2025–26 had multiple multi-hour to multi-day events; 7d ** rare** (May 2026 AWS still recovering EC2 days). 0.08–0.12/yr = ~20–30% over 3yr window for any region.

Evidence:

AWS US-EAST-1 history — Oct 2025 15h; May 2026 multi-day
Azure May 2026 ~22h with 2 AZ

Analogue: Major airline ground stop — 1–2×/decade national visibility.

Would update if: Zero regional outages >48h through 2027.

Conf: L–M

Claim: Node 7 modal branch adds +0.5–1.5pp to P(extinction by ~2050) vs distributed-compute counterfactual.

Why: Concentration increases frontier duty-cycle efficiency (Stargate, GW deals) → faster C9–C10; gov chokepoint used for CN containment, not alignment pause. Partial offset: kill switch could slow if used — 2026 track record says won’t.

Evidence:

Node 3 modal +0.5–1.5pp — export without pause
Node 4 — transparency not kill switch
Fable — 2 week partial restore

Analogue: Financial centralization → faster crises and faster growth.

Would update if: Gov sustained training halt >90d via cloud lever.

Conf: L–M

p(doom) — T7-A outage −0.5–2pp (temporary)

Claim: Major outage tail reduces extinction risk temporarily by slowing training weeks–months.

Why: Not structural — labs resume; no treaty. −0.5–2pp = speed channel only, low confidence.

Evidence:

AWS outages — no policy pause
Node 4 — incidents rarely durable

Would update if: Outage triggers binding federal training moratorium.

Conf: L

p(doom) — T7-C fragmentation +0.5–1.5pp

Claim: Sovereign cloud fragmentation increases coordination failure and multipolar race contribution to extinction.

Why: Weakens US-led HEM/verification (node3 T4 allied fracture); CN + EU + India parallel stacks → harder mutual pause.

Evidence:

Paris 2025 US unsigned; India sovereign frame
correlation matrix — race cluster

Would update if: EU–US joint cloud verification treaty.

Conf: L–M

p(doom) — T7-D HEM pilot −0.5–2pp if scales

Claim: Verified compute pilot reduces extinction if it becomes template for caps — conditional on scale beyond pilot.

Why: Only path where concentration helps pause — metered FLOPs. Pilot alone minimal; −0.5–2pp if adopted by 2030.

Evidence:

HEM_hardware_enabled_mechanisms.md — workload verification
Playbook item 8 — 3–5 yr horizon

Would update if: HEM mandated on Stargate clusters.

Conf: L

Confidence summary

Claim	Conf
Triopoly ≥70% US frontier through 2028	M–H
Modal = KYC/export not halt	M–H
Cloud registration ~0.55	M
DPA DC seizure ~0.06	L–M
Outage >30d delay ~0.10	L–M
EU/India fragmentation material	M
HEM pilot ~0.08	L
p(doom) lifts	L

External sources

Repo sources

(internal note)
(internal note)
(internal note)
(internal note)
(internal note)
(internal note)
(internal note)
(internal note)

1. Where this fits in the repo

P = 0.58 — Modal branch (triopoly + incremental rules, no halt)

P = 0.82 — US frontier ≥70% on AWS/Azure/GCP through 2028

P = 0.55 — Federal cloud training registration / KYC by 2028

P = 0.60 — BIS enforces IaaS / training catch-all (knowledge-based)

P = 0.75 — China blocked from US cloud frontier training (effective)

P = 0.35 — Repeatable gov API / deployment recall (Fable-class)

P = 0.45 — DPA for grid / transformers (not datacenter seizure)

P = 0.06 — DPA frontier datacenter seizure / mandatory training pause

P = 0.10 — Tail T7-A: Outage delays frontier training >30 days

P = 0.12 — Tail T7-B: Gov gate on frontier training runs (beyond export)

P = 0.15 — Tail T7-C: EU/India materially decouple >25% frontier spend off US hyperscalers

P = 0.50 — EU sovereign cloud >30% public-sector AI compute by 2028

P = 0.40 — India national AI off US hyperscaler for sovereign projects

P = 0.08 — Tail T7-D: HEM / verified compute pilot deployed

P = 0.12 — Tail T7-E: Cloud controls loosen for allies / accelerate build

P = 0.85 — Frontier labs maintain multi-cloud portfolio

P = 0.55 — Partial self-build (Stargate / lab-owned DC) for ≥1 major lab

P = 0.70 — Neocloud (CoreWeave / Oracle) wins ≥1 frontier training contract

P = 0.18 — Mandatory hyperscaler resilience / stress-test standard

P = 0.35 — Congress passes cloud training reporting law

P = 0.25 — EU bans US hyperscaler for sensitive GPAI training

P = 0.65 — Hyperscalers offer EU “sovereign SKU” (S3NS-class)

P = 0.08–0.12 / yr — Regional hyperscaler outage >7 days (any cause)

p(doom) — Modal +0.5–1.5pp (concentration → speed, not pause)

p(doom) — T7-A outage −0.5–2pp (temporary)

p(doom) — T7-C fragmentation +0.5–1.5pp

p(doom) — T7-D HEM pilot −0.5–2pp if scales

Confidence summary

External sources

Repo sources