← Back to all writing

p(doom) evidence — physical AI & economy limits

July 5, 2026

Evidence index · 中文 · Main post

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


Covers: node physical ai embodied · Physical economy limits
Date: 2026-07-04
Format: Claim | Why | Evidence | Analogue | Would update if | Conf


Embodied deployment

P = 0.75 — Humanoids remain pilot-scale through 2029

Claim: <5,000 humanoids worldwide doing paid multi-shift industrial work by 2029-12.

Why: Only Agility + Figure have verified shift KPIs; global productive base ~100–500 units (Robonaissance May 2026); integration timelines 12–24 mo per site; Unitree volume ≠ factory deployment.

Evidence:

  • (internal note) — master KPI table
  • Robonaissance deployment gap — ~13k shipped, ~100 sustained productive
  • Figure F02: 1 robot, 1,250 hrs over 11 mo — not fleet scale

Analogue: Autonomous trucking — years between “pilot” and “remove safety driver at scale.”

Would update if: Two additional OEMs publish >200 shift-work units each with KPIs by 2028-H1.

Conf: M–H


P = 0.12 — T-PA1 humanoid scale surprise by 2028

Claim: ≥3 OEMs each >500 paid shift-work units + replicated throughput KPIs (tote/parts/hr).

Why: BotQ 12k/yr nameplate + Unitree 10–20k 2026 target + Hyundai 25k roadmap — capacity exists; integration is bottleneck; tail requires simultaneous customer pull.

Evidence:

  • Figure BotQ: 12,000/yr line capacity (Figure blog Apr 2026)
  • Unitree IPO coverage: 10k–20k 2026 shipment target
  • Hyundai Atlas 25k/yr roadmap (JPM session May 2026, media coverage)

Analogue: iPhone ramp — supply chain can exceed expectations once product-market fit hits.

Would update if: 2027 shows zero fleet expansion beyond GXO/Figure/BMW — collapse tail to <5%.

Conf: L–M


P = 0.65 — Digital Node 1 salience peaks before physical labor coalition

Claim: Node 1 policy peak (~2027 Q2) precedes manufacturing/UAW mass mobilization by ≥12 mo.

Why: White-collar cuts (Block, Canaries) already live; factory humanoids still pilot; union cycles slower; Yale macro null on national unemployment.

Evidence:

  • (internal note) — partially live digital layer
  • (internal note) — Canaries lead macro
  • Agility/Figure KPIs — thousands of units max, not millions of workers displaced

Analogue: ATM adoption — bank teller political salience lagged technology by decade.

Would update if: UAW or AFL-CIO embodied-AI campaign launches 2026-H2 with major strike threat.

Conf: M


P = 0.10 — T-PA2 physical catastrophe → GUARD vetting (incremental on Node 1 T3)

Claim: +~4–6pp on Node 1 T3 (8% → 12–14%) from embodied Trigger E2 vs software-only.

Why: Injury/death legible; Boeing 737 MAX precedent; Hawley GUARD frames deployment vetting; warehouse humanoid > API agent for media cycle.

Evidence:

  • (internal note) §P=0.08 T3, §P(E2)≈0.10
  • (internal note) — GUARD Act watch list
  • OSHA workplace robot incident base rate low but fat-tail salience

Analogue: Uber AV pedestrian death (2018) — immediate regulatory response vs years of software-only debate.

Would update if: Major software-only agent loss >$100M without physical component passes vetting bill first.

Conf: L–M


P = 0.70 — Battlefield autonomy expands without US domestic robot moratorium

Claim: Ukraine/Middle East autonomy ; US no broad ban on industrial humanoids through 2029.

Why: Separate policy lanes (DoD vs OSHA/DOL); Replicator under-delivered but continues; LAWS treaty stalled.

Evidence:

Analogue: Drone warfare 2010s — expanded abroad, domestic FAA separate.

Would update if: Autonomous strike kills >50 civilians + attributed to US AI module → domestic moratorium bill in 90 days.

Conf: M


P = 0.15 — T-PA4 embodiment data flywheel accelerates VLA (not full Ci)

Claim: Log-linear scaling past 20k hr continues~2× real-world task improvement rate 2027–2029 for manipulation.

Why: EgoScale Feb 2026; capital forcing deployment-for-data; no saturation proof yet.

Evidence:

  • (internal note) §EgoScale
  • Figure Helix 02 whole-body loco-manipulation @ BMW F03

Analogue: ImageNet → CV revolution — data scale preceded algorithmic breakthroughs.

Would update if: Published scaling study shows plateau at <50k hr.

Conf: L–M


Energy / GW

P = 0.60 — Modal: constraint absorbed via on-site gen + delays (0.85× Ci)

Claim: Energy binds as 6–18 mo project delays, not multi-year global training halt; effective Ci pace ~0.85× 2027–2029.

Why: Hyperscaler capex $200B+ class; Microsoft/OpenAI Stargate; gas PPAs; secondary markets; idle compute episodic.

Evidence:

Analogue: 2000s fiber overbuild → utilization lag, not permanent stall.

Would update if: Two frontier labs publicly cite power (not chips) as primary reason for skipping next OOM run in 2027.

Conf: M


P = 0.15 — Tail: energy as hard cap (0.5–0.7× Ci 2027–2030)

Claim: ≥43% of announced GW cannot energize by 2028; 10²⁸ runs slip 2+ yr; Ci 0.5–0.7×.

Why: 2.3 TW interconnection queue; 6–8 yr waits; transformer shortages; NIMBY shrinking pipeline; 5× >1 GW sites by 2026.

Evidence:

  • CBRE H2 2025: construction pipeline first time since 2020
  • The Register NIMBY datacenter
  • Enkiai: 267% wholesale price spikes near DCs; 3+ yr connection delays

Analogue: 1970s oil shock — real growth constraint, not narrative.

Would update if: PJM/FERC reforms cut interconnection to <3 yr median by 2027; pipeline >8 GW under construction.

Conf: L–M


P = 0.55 — On-site / behind-the-meter generation becomes default for >100 MW AI sites

Claim: Majority of new frontier-scale campuses include owned or dedicated generation (gas, SMR MOU, solar+storage).

Why: Grid queue failure; Microsoft, Amazon, Meta, xAI Colossus patterns; CBRE trend list.

Evidence:

  • Creative Strategies — on-site gen “routine” in large planning
  • xAI Colossus, Stargate, Prometheus references in industry reports

Analogue: Aluminum smelters — co-located with hydro/nuclear historically.

Would update if: Federal grid fast-track reduces need; new sites 100% grid-connected without PPA in 2027–28.

Conf: M


Data wall + synthetic

P = 0.55 — HQ text exhaustion effective 2028–2030 (modal, post-Epoch revision)

Claim: Next OOM pretrain (post-Llama-4 class) hits data-limited regime before compute-limited; calendar 2028–2030.

Why: Epoch ~300T token stock; overtraining common; filtering gains pushed back 2022 estimate.

Evidence:

Analogue: Oil proven reserves — exhaustion date moves with technology.

Would update if: New public text stock >1P tokens verified (e.g. massive proprietary corpora released).

Conf: M


P = 0.50 — ~30% rephrased synthetic + verification extends runway 3–5 yr

Claim: Frontier labs sustain scaling via curated synthetic mix without benchmark regression through ~2030–2032.

Why: EMNLP 2025 systematic study; 5–10× loss speedup; Altman “high-quality synthetic” endorsement; MGT detection.

Evidence:

  • arXiv:2510.01631 / ACL Anthology EMNLP 2025
  • arXiv:2510.16657 — verification prevents collapse
  • EMNLP 2025 MGT detection paper

Analogue: Audio compression — synthetic augmentation standard in ML once quality controlled.

Would update if: Frontier model >50% synthetic shows WMDP/MMLU regression vs prior gen at equal compute.

Conf: M


P = 0.05 — S3 data-collapse tail (textbook-style synthetic dominates)

Claim: ~5% labs over-index textbook synthetic → measurable capability regression / diversity collapse 2027–28.

Why: Tail of bad practice; open-weight race incentives; web AI-slop feedback loop.

Evidence:

  • arXiv:2510.01631 — textbook-style collapse patterns
  • arXiv:2404.05090 — recursive synthetic-only collapse theorem

Analogue: Financial model overfitting — institutions occasionally blow up.

Would update if: No regression observed through 2029 despite >40% synthetic in major releases.

Conf: L


P(Δ Node 1 timing | physical AI modal) ≈ +12–24 mo for second labor wave

Claim: Physical displacement salience follows digital peak unless T-PA1 or T-PA2.

Evidence: Sections above; Node 1 hybrid rule

Conf: M


P(energy hard cap) × P(C7 by 2029) — interaction

Claim: Conditional on energy tail, P(C7 by 2029) ≤0.35 vs ~0.55 unconditional (rough).

Why: C7 requires sustained RSI infrastructure; power binds training not just inference.

Evidence: Physical economy limits scenario table; cross node C5–C7 Emerging

Conf: L (explicit model not yet calibrated)


Claim: Datacenter GW constraint is Tier 0 manifestation of Landauer + finite grid — same physics family as cosmic limits, different scale.

Evidence: (internal note) §2.1 Tier 0, §4 Landauer

Conf: H (physics); M (economic binding timing)