AI futures evidence — U4: physical deployment

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

Parent: Utopia timeline §U4
Ceiling: superintelligence physical limits
Cross-cuts: Physical economy limits · node physical ai embodied · c8 society snapshot by ci §bottlenecks
Path family: C — Radical abundance tail (primary); enables U1 bucket conditional on deployment, not discovery alone
Date: 2026-07-04
Format: Claim | Why | Evidence | Analogue | Would update if | Conf

Executive summary

Node U4 question: P(lunar ISRU / fusion pathway / GW clean energy deployed at scale — not merely designed — by horizon ~2035–2050)?

Load-bearing crux — deployment vs discovery:

Layer	What AI accelerates (bits)	What stays human/physics-bound (atoms)	Typical lag
Energy	Tokamak control, site selection, grid modeling	Interconnection queue, transformer fab, NRC licensing, concrete	5–15 yr post-design
Fusion	Plasma physics simulation, materials candidates	FOAK plant build, tritium supply, regulatory	10–20 yr post-Q>1
Space	Trajectory optimization, autonomy, lightweight structures	Δv, launch cadence, vacuum testing, ISRU TRL climb	5–12 yr post-bench demo
Embodied AI	VLA policies, factory scheduling	Robot-hour data, OSHA, integration, liability	3–8 yr post-capability

Modal (@C7–C10): AI compresses design cycles 2–5× for reducible engineering; does not bypass Landauer heat, c latency, interconnection queues, or clinical/GMP analogues in space hardware. c8 society snapshot by ci: fusion still not modal grid contributor at C8; no Kardashev jump before C9+.

Path family C (radical abundance tail): P(any U1-scale physical deployment — fusion >1 GW grid + lunar ISRU operational + Dyson-class not required) ~0.12–0.18 by 2050 conditional on alignment + distribution elsewhere. Discovery-only (papers, demos, Q>1 in lab) ~3–5× more likely than deployed scale.

U1 bucket mapping: U4 node feeds U1 — Radical abundance only through deployed watts, kg propellant, and factory throughput — not through internal lab genius alone. P(U1 partial credit from U4 alone | C9 reached) ~0.25–0.35; full U1 requires U1/U2/U3 conjunction.

Amodei vs physical-limits branches:

Branch	Claim	P(modal)
Amodei optimistic	”Compressed 21st century” — 10× progress in energy, bio, space outcomes within a decade of powerful AI	Design: ~0.55; deployed abundance: ~0.20
Physical-limits conservative	Superintelligence = optimal scheduler under c, Landauer, Bekenstein; logistics dominate	~0.70 modal through C9
Synthesis	Both can be true: AI shrinks search time; physics/institutions set deployment floor	—

Energy — GW datacenter growth

P = 0.70 — US datacenter load reaches 50–60 GW by 2030 (deployed)

Claim: US datacenter connected load ≥50 GW by 2030-12, from ~42 GW (2026 est.).

Why: IEA US DC ~180 TWh (2024) → +240 TWh incremental to 2030; hyperscaler capex $320B+ (2025); Stargate-class campuses; load factor and regional buildout support ~2× in 4 yr despite bottlenecks.

Evidence:

IEA Electricity Mid-Year Update 2025 — US DC major demand driver; >2% annual US demand growth 2025–26
IEA Energy and AI executive summary — global DC 485 TWh (2025) → ~950 TWh (2030)
Physical economy limits — US ~42 GW (2026), GPU incremental ~20 GW (2027)

Analogue: Mobile data traffic 2010s — demand forecasts repeatedly low.

Would update if: US EIA actualized DC load <45 GW in 2028 with pipeline cancellations >30%.

Conf: M

Claim: Energy binds as project delays and regional shortfalls; no global frontier training halt through 2029.

Why: On-site gas, behind-the-meter nuclear PPAs, Gulf/China less constrained; hyperscaler balance sheets absorb idle compute episodically.

Evidence:

node physical ai §P=0.60
Creative Strategies AI Infrastructure 2026
c8 society snapshot by ci — Physical-economy modifier S1 P≈0.55

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

Would update if: Two frontier labs skip next OOM run citing power (not chips) as primary reason in 2027.

Conf: M

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

Claim: ≥43% of announced US AI GW cannot energize by 2028; largest 10²⁸ runs slip 2+ yr.

Why: ~2.3 TW interconnection queue; 6–8 yr waits; transformer >3 yr lead; NIMBY shrinks pipeline (CBRE H2 2025).

Evidence:

Physical economy limits §Tail S2
The Register — NIMBY datacenter pushback

Analogue: 1970s oil shock — real growth constraint.

Would update if: PJM/FERC reforms cut median interconnection to <3 yr by 2027.

Conf: L–M

P = 0.55 — Behind-the-meter / dedicated generation default for >100 MW new AI sites

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

Why: Grid queue failure; Talen-Amazon, Stargate, xAI Colossus patterns; FERC friction pushes front-of-meter deals.

Evidence:

node physical ai §P=0.55
Utility Dive — FERC rejects Talen-Amazon ISA — regulatory friction on co-location
Presenc AI Hyperscaler Nuclear PPA Tracker 2026 — ~9.8 GW disclosed nuclear commitments

Analogue: Aluminum smelters co-located with hydro historically.

Would update if: Federal grid fast-track makes 100% grid-connected sites modal without dedicated gen 2027–28.

Conf: M

P = 0.40 — AI-focused DC electricity grows 50% YoY through 2027 then moderates

Claim: AI-specific DC load maintains ~40–50% annual growth 2025–2027, then ~25–35% 2028–2030 as bottlenecks bind.

Why: IEA reported 50% AI-DC surge 2025; bottlenecks “reduce likelihood of more aggressive near-term scenarios” per IEA 2025 update.

Evidence:

IEA Energy and AI — executive summary — 17% overall DC growth 2025; AI-focused 50%
IEA Global Energy Review 2026 — electricity demand — DC ~17% of global demand growth increment

Analogue: Smartphone unit growth curve — steep then logistic.

Would update if: 2027–28 actual AI-DC growth stays >50% with no reported energization delays.

Conf: M

Clean energy coupling

P = 0.35 — ≥70% of incremental US AI DC capacity carbon-free by 2030 (deployed)

Claim: Of new GW added 2026–2030, ≥70% matched to nuclear restart, PPA solar/wind, or certified clean firm power.

Why: Hyperscaler net-zero pledges; ~9.8 GW nuclear PPAs alone; gas still fills gap for speed — not 100% clean.

Evidence:

Presenc AI nuclear tracker — Meta ~6.6 GW, Microsoft TMI restart 835 MW, Google Kairos 500 MW path
IDTechEx — SMR datacenter timelines — FOAK SMR ~2030; restarts 2027

Analogue: Corporate renewable PPAs 2015–2020 — real but partial.

Would update if: >50% of 2027–28 new US AI GW is gas-only with no clean PPA attached.

Conf: L–M

P = 0.65 — Nuclear restart / existing-plant PPAs deliver ≥5 GW to hyperscalers by 2030 (deployed)

Claim: ≥5 GW continuous carbon-free power from restarts + existing plant deals (not greenfield SMR) energizing AI campuses by 2030.

Why: TMI restart 2027; Talen Susquehanna; Meta Constellation Clinton; deals already signed — deployment horizon 2–4 yr.

Evidence:

Presenc AI tracker — first-power 2027 concentration for restarts
IRECruit SMR/datacenter 2026 — AWS-Talen 1.92 GW PPA; operational spring 2026

Analogue: Dedicated industrial power contracts — slow but reliable.

Would update if: TMI restart slips past 2028 and two other restart PPAs cancelled.

Conf: M

P = 0.25 — First FOAK SMR energizes hyperscaler campus by 2031 (deployed)

Claim: At least one SMR (≥50 MW) delivers power to a disclosed hyperscaler site by 2031-12.

Why: Kairos Hermes 2 2030 target; Oklo Aurora pilot 2027–28 aspirational; HALEU fuel chain slow; FOAK delay base rate high.

Evidence:

Google-Kairos PPA — 50 MW initial, 500 MW scale
IDTechEx SMR article — FOAK ~2030 even without delays

Analogue: First commercial solar farms — decade from demo to GW.

Would update if: Kairos or Oklo achieves grid connection before 2029 with public meter data.

Conf: L–M

Fusion pathways

P = 0.08 — Grid-connected fusion electricity ≥100 MW deployed by 2035

Claim: ≥100 MW fusion power selling to grid (any confinement concept) by 2035-12.

Why: NIF G>1 is ICF lab demo, not power plant; CFS SPARC net energy late 2026 target; ARC early 2030s; Helion 2028 PPA contingent — slip history (Polaris net elec not confirmed Feb 2026).

Evidence:

PRL — NIF G>1 (Feb 2024) — 3.1 MJ yield; not power plant
CFS ARC physics basis papers (Jun 2026) — 400 MW net elec design; grid early 2030s
DIW Weekly Report 2026 — Helion Polaris net elec not publicly confirmed; commercial plant 2028 stated

Analogue: Fission commercialization — ~15 yr from demo to grid after physics proof.

Would update if: SPARC achieves documented Q_engineering >1 and grid PPA for ARC signed with <5 yr delivery.

Conf: L

P = 0.25 — SPARC (or equivalent) demonstrates net fusion energy gain by 2028

Claim: CFS SPARC or comparable private tokamak achieves documented net energy gain (fusion power > input power, plant definition) by 2028-12.

Why: CFS late 2026 public target; $3B+ raised; Milestone-Based DOE program; delay base rate for FOAK magnets.

Evidence:

CFS news Jun 2026
NatureTechMemos fusion startups 2026 — no private net gain yet

Analogue: Wright Flyer — flight demo before airline.

Would update if: SPARC slips past 2029 with public magnet or plasma failure.

Conf: M (for attempt); L (for success timing)

P = 0.55 — P(fusion designed at scale | C7) >> P(fusion deployed ≥100 MW | C7) — gap ≥12 yr

Claim: Conditional on C7 internal genius, P(validated fusion plant design + materials list) ~0.50–0.65 vs P(operational ≥100 MW grid fusion by C7+10 yr) ~0.08–0.15.

Why: Discovery = reducible plasma/materials search; deployment = NRC-equivalent licensing, tritium, heat exchangers, concrete — superintelligence physical limits §9 “logistics not algorithm alone”; Amodei: trials “not infinitely fast.”

Evidence:

c8 society snapshot by ci C8 — fusion still not modal grid contributor
Dwarkesh — Amodei interview — economy 10–20% growth, not 300%; compute capped by physical buildout
superintelligence physical limits §2.3 — Dyson ~10²–10³ yr engineering

Analogue: mRNA vaccine design (days) vs rollout (months–years).

Would update if: AI-designed fusion plant achieves grid connection <7 yr after first Q>1 demo.

Conf: M

P = 0.03 — Dyson swarm or Kardashev Type II energy capture before 2050 (deployed)

Claim: ≥10²⁴ W (order-of-magnitude stellar capture) operational by 2050-12.

Why: Tier 1 transition ~10¹–10³ yr per superintelligence physical limits §2.1; even C9 superhuman researcher does not repeal c or mass logistics; Armstrong–Sandberg galaxy fill 10⁶–10⁸ yr.

Evidence:

superintelligence physical limits §2.2, §6.1 — Dyson ~10²⁶ W requires stellar engineering at scale
Armstrong & Sandberg 2013 — eternity in six hours assumes existing replicator infrastructure; bootstrap still bound by launch mass

Analogue: 1950s Project Orion — physics OK, deployment absent.

Would update if: Detected mid-IR excess from Sol-scale partial Dyson structure (SETI) — would raise to ~0.30 and re-evaluate entire timeline.

Conf: H (physics); M (timeline conditional on ASI)

Physical limits ceiling (Earth + cosmic)

P = 0.95 — Earth AI compute remains Landauer-limited; no “infinite scale on Earth”

Claim: Even at C9–C10, Earth-bound AI ensembles cannot exceed ~10¹²–10¹³ W practical sustained dissipation without ecological/economic collapse — far below Bekenstein 1 kg laptop ~10⁵⁰ ops/s theoretical bound.

Why: Landauer: E_erase ≥ k_B T ln 2 per irreversible bit; datacenter 1 GW ≈ 3×10⁹ W = 0.3% US peak — cosmic Tier 1 ~10²⁶ W is 10¹⁷× higher.

Evidence:

Lloyd, Ultimate physical limits to computation, Nature 406 (2000) — arXiv:quant-ph/9908043
Landauer (1961) — irreversibility heat; IBM J. Res. Dev. 5(3)
Cosmic-scale evolution and game theory

Analogue: CPU clock scaling stopped on heat, not ideas.

Would update if: Room-temperature reversible computing demonstrably cuts practical datacenter heat >100× at scale (not lab only).

Conf: H

P = 0.90 — Krauss–Starkman ~600 yr Moore’s-law civilization ceiling applies to AI-designed chips

Claim: Superintelligence does not indefinitely extend exponential compute-per-watt on Earth; ~600 yr order-of-magnitude ceiling (Krauss–Starkman 2004) remains binding for civilization-scale ensembles.

Why: De Sitter horizon + finite harvestable energy ~3.5×10⁶⁷ J; Λ>0 kills Dyson infinite subjective time.

Evidence:

Krauss & Starkman, Universal Limits on Computation, arXiv:astro-ph/0404510
Krauss & Starkman, Knowledge Decreases with Time, arXiv:astro-ph/9902189
superintelligence physical limits §4.3–4.4

Analogue: Thermodynamics limits on engine efficiency — design hits asymptote.

Would update if: Peer-reviewed demonstration of sustained >2×/yr economy-wide ops/J beyond 2040 with no heat crisis.

Conf: H (physics); M (economic manifestation)

P = 0.85 — No FTL, no horizon escape, no global entropy reversal even with C10 ASI

Claim: Items in superintelligence physical limits §7 CANNOT table remain impossible through C10.

Why: Relativity, second law, Bekenstein, Gögel/irreducibility — intelligence is scheduler, not lawbreaker.

Evidence:

superintelligence physical limits §7 CANNOT table
Wolfram (2024) — computational irreducibility limits “solve science” shortcuts

Analogue: Chess engine cannot move pieces faster than clock allows.

Would update if: Reproducible FTL communication peer-reviewed (would overturn entire doc).

Conf: H

P = 0.40 — Post-C9 Tier 1 solar-system engineering begins (probe factories, partial Dyson studies) by ~2100

Claim: If C9+ aligned expansionist trajectory, first self-replicating probe launch systems and partial stellar-harvest studies begin ~2075–2125 — not full Kardashev II.

Why: Armstrong–Sandberg: interstellar probe launch could start in “foreseeable future” with automation; full swarm ~10²–10³ yr; conditional on alignment + instrumental convergence.

Evidence:

superintelligence physical limits §2.3 — probe launch ~10¹–10² yr startup

Analogue: ISS → Artemis — multi-decade institutional ramp.

Would update if: C9 not reached by 2040 or alignment prevents any expansionist infrastructure.

Conf: L (highly conditional)

Space — Starship, lunar ISRU, autonomy

P = 0.72 — Starship delivers first commercial payload to orbit by 2027 (deployed)

Claim: Paid commercial orbital payload on Starship by 2027-12 (SpaceX S-1 H2 2026 target).

Why: Flight 12 (V3) May 2026 suborbital success; mishap investigation ongoing; commercial inflection explicit in IPO materials; history — milestones slip but arrive.

Evidence:

Analogue: Falcon 9 — years from demo to commercial, then rapid.

Would update if: No commercial payload contract flown by 2028-H1.

Conf: M

P = 0.30 — Starship achieves ≥12 orbital flights/year by 2029 (deployed cadence)

Claim: ≥12 Starship orbital missions in calendar 2029.

Why: FAA authorized 25/yr Starbase + 44/yr LC-39A; 2025 flew 5 vs 25 target; airline cadence not yet demonstrated; mishap cycles add downtime.

Evidence:

The Next Web — 5 vs 25 in 2025
New Space Economy — milestone completion record — propellant transfer, Starlink V3, HLS all incomplete Jun 2026

Analogue: Shuttle — ~5–9 flights/yr at peak, not 25.

Would update if: 2027 calendar shows ≥8 Starship orbital flights.

Conf: L–M

P = 0.50 — Artemis IV crewed lunar landing by early 2029 (deployed)

Claim: NASA Artemis IV lands crew on lunar surface by 2029-Q2 (NASA target early 2028 + slip buffer).

Why: Artemis III re-scoped to LEO demo late 2027; first landing Artemis IV early 2028 official; SLS/HEOMD delay history; Starship HLS uncrewed demo incomplete.

Evidence:

Analogue: Apollo — schedule slips measured in years, not months.

Would update if: Starship HLS uncrewed lunar demo before 2027 + SLS cadence ≥2/yr.

Conf: M

P = 0.18 — LIFT-1 lunar oxygen extraction demo operational ≥1 kg/hr by 2032 (deployed)

Claim: NASA/industry LIFT-1 class demo produces ≥1 kg/hr O₂ from regolith on lunar surface by 2032.

Why: LIFT-1 RFI Nov 2024; CaRD ground demo validated chemistry; NextSTEP-3 Jun 2026 targets TRL 5–6 on ground — flight 5+ yr after TRL-6 typical; Artemis base late 2020s start.

Evidence:

Analogue: MOXIE Mars — kg-scale demo, not industrial plant.

Would update if: LIFT-1 flight manifest before 2029 with funded hardware.

Conf: L–M

P = 0.48 — P(lunar ISRU plant designed | C7) vs P(kg O₂/hr operational | C7) — gap ≥8 yr

Claim: C7 internal genius produces validated ISRU flowsheet + mass/energy ~0.45–0.55 by C7; operational lunar plant ~0.10–0.20 within 10 yr of C7.

Why: TRL climb, space qualification, launch mass, surface ops — same deployment crux as fusion; c8 C7 space: “hardware still human-paced.”

Evidence:

NASA ISRU overview — analog tests Earth-only; “long mission durations” in space not yet
c8 society snapshot by ci §Space row — ISRU lead times bottlenecked

Analogue: Offshore oil platform — years from engineering to barrels/day.

Would update if: Artemis V base includes operational ISRU before 2030.

Conf: M

P = 0.45 — AI compresses autonomous mission design cycle 2–3× by C8 (deployed in ops, not just studies)

Claim: NASA/DoD/commercial flight-certified mission design iterations (requirements → validated design) 2–3× faster than 2020 baseline by C8.

Why: Reducible subsystem in superintelligence physical limits §6.1 — orbit mechanics, autonomy, materials for specific missions; still need vacuum test.

Evidence:

c8 society snapshot by ci C8 — LEO traffic management AI-operated; C9 — probe design months not years
c8 §bottlenecks — Space row

Analogue: CAD/CFD revolution in aerospace 1990s — 2×, not 100×.

Would update if: Independent audit shows no measurable cycle compression 2026–2029 despite C6–C7 tools.

Conf: M (speculative at C8)

Embodied AI — positive manufacturing paths

P = 0.75 — Humanoids remain <5k shift-work units through 2029 (deployment ceiling)

Claim: <5,000 humanoids on paid multi-shift industrial work globally by 2029 — positive productivity localized, not sector-wide.

Why: Only Agility + Figure verified KPIs; integration 12–24 mo/site; capital ≠ deployment.

Evidence:

node physical ai §P=0.75
node physical ai embodied §Verified deployment baseline

Analogue: Autonomous trucking — pilot years before scale.

Would update if: ≥3 OEMs each >500 shift units with KPIs by 2028-H1.

Conf: M–H

P = 0.38 — AI-driven factory productivity +15–25% in early adopter plants by 2030 (deployed KPI)

Claim: ≥10 public disclosed factories report ≥15% output/labor-hour gain attributable to AI+VLA+scheduling stack by 2030.

Why: BMW Figure hours, GXO totes real but small; digital twin + scheduling easier than humanoid scale; Ghost GDP margin channel.

Evidence:

Figure @ BMW — 1,250+ hr, 90k parts
Agility 100k totes
node physical ai embodied — modal no sector employment step-change

Analogue: Industrial robotics 2010s — 10–20% cell productivity in adopters.

Would update if: Zero new industrial KPI disclosures 2027–28.

Conf: L–M

P = 0.15 — Embodiment data flywheel (T-PA4) doubles physical deployment improvement rate 2027–2029

Claim: EgoScale-like log-linear scaling holds → ~2× VLA task improvement rate vs 2024–2026 baseline.

Why: Capital forcing deployment-for-data; no saturation proof; does not accelerate LLM reasoning spine.

Evidence:

node physical ai §P=0.15 T-PA4
Physical economy limits — robot-hour scarcity separate from text data

Analogue: ImageNet → CV — data scale preceded algorithms.

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

Conf: L–M

Deployment vs discovery — cross-cutting cruxes

P = 0.65 — Amodei “compressed 21st century” accurate for design; not for deployed watts/kg by 2035

Claim: By 2035, ≥50% of Amodei Machines of Loving Grace energy/space design targets reachable in lab/sim; ≤25% of corresponding deployment targets (grid fusion MW, lunar ISRU kg/hr, GW clean DC).

Why: Amodei hedges on energy/compute caps; Dwarkesh interview — economy 10–20% growth not 300%; clinical/physical “not infinitely fast.”

Evidence:

Amodei — Machines of Loving Grace — energy tech listed under positive applications; explicit stall risks
Dwarkesh Amodei 2026 — compute purchase bankruptcy risk if off 1 yr
superintelligence physical limits §9

Analogue: Human Genome Project — sequence fast, therapy slow.

Would update if: Two Amodei-class deployment targets (fusion >100 MW, or lunar O₂ >100 kg/hr) operational before 2032.

Conf: M

P = 0.70 — Physical-limits conservative branch dominates atoms timeline by ≥10 yr vs Amodei implied calendar

Claim: Median gap between Amodei-implied ** societal deployment** dates and physics/institution-limited dates ≥10 yr for energy + space hardware.

Why: Lloyd/Krauss ceilings + interconnection + NRC/NASA + c — superintelligence shifts reducible subspace, not queue physics.

Evidence:

superintelligence physical limits §1, §9
c8 society snapshot by ci — RTSC economically irrelevant through C9; fusion multi-decade modal

Analogue: Nuclear fission — physics 1938, grid scale 1950s–60s.

Would update if: Post-C7 world shows ≥3 atom-scale deployments at Amodei cadence (<5 yr idea→GW).

Conf: M

P = 0.35 — AI compresses historical deployment lag 30–50% (optimistic branch)

Claim: For reducible infrastructure (datacenter energization, factory automation, mission design), AI tools cut idea → operational time to 50–70% of 20th-century analogue — not 10% (Amodei) or 100% (status quo).

Why: Middle ground between branches; software/control reducible; regulatory/nuclear not.

Evidence:

c8 C9 — probe design months vs years
SMR FOAK still 2030 despite AI chip design acceleration — partial compression only

Analogue: BIM + modular construction — ~30% schedule cut on buildings, not 90%.

Would update if: Median FOAK fusion or SMR timeline unchanged 2026–2035 despite C7 tools.

Conf: L–M

Path family C — radical abundance tail → U1

P = 0.14 — Path C tail: deployed radical abundance physical layer by 2050

Claim: By 2050: (a) fusion ≥1 GW grid or (b) lunar ISRU ≥100 kg/hr O₂ or (c) AI DC ≥200 GW globally with ≥60% clean firm power — any two of three.

Why: Each sub-threshold alone ~0.25–0.40 by 2050; conjunction discounted; requires C8+ capability and institution stability and no energy tail S2 permanent bind.

Evidence:

Sections above — fusion ~0.08 by 2035 for 100 MW scales to ~0.15–0.25 by 2050; ISRU ~0.18 by 2032 scales slowly; DC clean GW trajectory from IEA
my putopia — U1 horizon ~2050+; historical tech optimism downward prior

Analogue: Post-WWII electrification + Green Revolution — multi-decade compound deployed infrastructure.

Would update if: Any single sub-condition met before 2035 → raise Path C to ~0.25.

Conf: L

P = 0.28 — U1 partial credit: GW clean compute deployed enables abundance without fusion/space

Claim: P(U1 partial — energy/compute half of abundance story) ~0.28 by 2050 from deployed GW alone, conditional on U2/U3 distribution.

Why: Cheap inference + clean firm GW = radical digital abundance; U1 full post-scarcity needs bio/longevity/space — U4 alone insufficient.

Evidence:

IEA — DC 950 TWh 2030
my putopia — U1 definition post-scarcity + longevity + space; overlap note

Analogue: Internet bandwidth abundance — real, but not U1 alone.

Would update if: Energy prices ↑ near DC clusters despite buildout (failure mode for digital abundance).

Conf: L–M

P = 0.45 — Discovery-without-deployment dominates Path C narrative (anti-utopia trap)

Claim: ~45% of worlds reaching C8+ show impressive demos (Q>1, MOXIE-class, humanoid KPIs) but fail Path C deployment thresholds by 2050 — feeds modal c8, not U1.

Why: Base rate of demo→scale in energy/space; Amodei branch risk; PE1 energy tail 0.15.

Evidence:

Helion Polaris slip pattern (DIW 2026)
Starship cadence miss 2025
c8 society snapshot by ci — modal not utopia

Analogue: Segway — invention without transformation.

Would update if: ≥2 atom technologies at scale before 2035.

Conf: M

Impossible even with superintelligence (synthesis estimates)

P = 0.99 — No FTL communication/coordination by 2100

Claim: Relativity-bound latency remains for all human/AI civilization.

Evidence: superintelligence physical limits §7; Tier A causal bucket.

Conf: H

P = 0.98 — No infinite subjective time / Dyson eternal intelligence (Λ>0)

Claim: de Sitter T_min + Landauer → finite thoughts.

Evidence: Dyson (1979) refuted by Krauss–Starkman; superintelligence physical limits §4.4.

Conf: H

P = 0.95 — No shortcut past computational irreducibility for chaotic/nonequilibrium systems

Claim: Weather, turbulence, many-body nonequilibrium — still require simulation at scale.

Conf: H

P = 0.92 — Bekenstein density cannot be exceeded for information storage

Claim: Saturating bound → black hole; no denser stable memory.

Evidence: Bekenstein (1981); Lloyd (2000) ultimate laptop.

Conf: H

Falsifiers (node-level)

Observation	Effect
≥3 frontier 10²⁸+ runs on schedule 2027–28, no power delays	PE1 weakened; U4 energy deployment ↑
Grid fusion ≥100 MW before 2032	Raise fusion sections +10–15pp; Path C ~0.25
LIFT-1 ≥10 kg/hr O₂ on Moon before 2030	ISRU deployment crux broken upward
Starship ≥20 orbital flights in 2028 calendar	Cadence sections +15pp; space Path C ↑
Zero hyperscaler nuclear PPAs energized by 2029	Clean coupling ↓; behind-meter ↓
Detected partial Dyson technosignature	Re-open Tier 1 timeline; entire U4 ceiling revised

Sources (external)

IEA: Energy and AI · Electricity Mid-Year 2025 · Global Energy Review 2026
Fusion: NIF G>1 PRL 2024 · CFS ARC papers Jun 2026 · DIW fusion startups 2026
Space: NASA Artemis architecture Feb 2026 · Ars Artemis III/IV · New Space Economy Starship milestones
Nuclear/DC: Presenc AI nuclear tracker 2026 · IDTechEx SMR datacenters
Amodei: Machines of Loving Grace · Dwarkesh interview 2026

Changelog

Date	Change
2026-07-04	Initial research pass — 32 evidence sections; deployment vs discovery crux; Path C / U1 mapping; Amodei vs physical-limits branches

Executive summary

Energy — GW datacenter growth

P = 0.70 — US datacenter load reaches 50–60 GW by 2030 (deployed)

P = 0.60 — Modal: GW constraint = 6–18 mo delays, Ci ~0.85× (not training halt)

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

P = 0.55 — Behind-the-meter / dedicated generation default for >100 MW new AI sites

P = 0.40 — AI-focused DC electricity grows 50% YoY through 2027 then moderates

Clean energy coupling

P = 0.35 — ≥70% of incremental US AI DC capacity carbon-free by 2030 (deployed)

P = 0.65 — Nuclear restart / existing-plant PPAs deliver ≥5 GW to hyperscalers by 2030 (deployed)

P = 0.25 — First FOAK SMR energizes hyperscaler campus by 2031 (deployed)

Fusion pathways

P = 0.08 — Grid-connected fusion electricity ≥100 MW deployed by 2035

P = 0.25 — SPARC (or equivalent) demonstrates net fusion energy gain by 2028

P = 0.55 — P(fusion designed at scale | C7) >> P(fusion deployed ≥100 MW | C7) — gap ≥12 yr

P = 0.03 — Dyson swarm or Kardashev Type II energy capture before 2050 (deployed)

Physical limits ceiling (Earth + cosmic)

P = 0.95 — Earth AI compute remains Landauer-limited; no “infinite scale on Earth”

P = 0.90 — Krauss–Starkman ~600 yr Moore’s-law civilization ceiling applies to AI-designed chips

P = 0.85 — No FTL, no horizon escape, no global entropy reversal even with C10 ASI

P = 0.40 — Post-C9 Tier 1 solar-system engineering begins (probe factories, partial Dyson studies) by ~2100

Space — Starship, lunar ISRU, autonomy

P = 0.72 — Starship delivers first commercial payload to orbit by 2027 (deployed)

P = 0.30 — Starship achieves ≥12 orbital flights/year by 2029 (deployed cadence)

P = 0.50 — Artemis IV crewed lunar landing by early 2029 (deployed)

P = 0.18 — LIFT-1 lunar oxygen extraction demo operational ≥1 kg/hr by 2032 (deployed)

P = 0.48 — P(lunar ISRU plant designed | C7) vs P(kg O₂/hr operational | C7) — gap ≥8 yr

P = 0.45 — AI compresses autonomous mission design cycle 2–3× by C8 (deployed in ops, not just studies)

Embodied AI — positive manufacturing paths

P = 0.75 — Humanoids remain <5k shift-work units through 2029 (deployment ceiling)

P = 0.38 — AI-driven factory productivity +15–25% in early adopter plants by 2030 (deployed KPI)

P = 0.15 — Embodiment data flywheel (T-PA4) doubles physical deployment improvement rate 2027–2029

Deployment vs discovery — cross-cutting cruxes

P = 0.65 — Amodei “compressed 21st century” accurate for design; not for deployed watts/kg by 2035

P = 0.70 — Physical-limits conservative branch dominates atoms timeline by ≥10 yr vs Amodei implied calendar

P = 0.35 — AI compresses historical deployment lag 30–50% (optimistic branch)

Path family C — radical abundance tail → U1

P = 0.14 — Path C tail: deployed radical abundance physical layer by 2050

P = 0.28 — U1 partial credit: GW clean compute deployed enables abundance without fusion/space

P = 0.45 — Discovery-without-deployment dominates Path C narrative (anti-utopia trap)

Impossible even with superintelligence (synthesis estimates)

P = 0.99 — No FTL communication/coordination by 2100

P = 0.98 — No infinite subjective time / Dyson eternal intelligence (Λ>0)

P = 0.95 — No shortcut past computational irreducibility for chaotic/nonequilibrium systems

P = 0.92 — Bekenstein density cannot be exceeded for information storage

Falsifiers (node-level)

Sources (external)

Changelog