2050: P(SI) ≈ 67%, P(abundance | SI) ≈ 25%, P(doom | SI) ≈ 7%

November 1, 2026 / 32 min read

Assessment dated 1 November 2026. All three events have a 2050 deadline. Rounded model outputs, not measured frequencies.

Hypothesis: intelligence can accelerate the factories, experiments and infrastructure that spread its benefits. Several energy routes could support global material abundance. Reliable production, clinical validation, ecological protection and access must succeed together. The same capabilities can amplify destructive power. I first test the physical requirements, then calculate probabilities using shared timing and institutional assumptions.

The headline numbers are outputs. Every numerical probability below is a disclosed judgment. Physical calculations and observed results constrain those judgments; they do not uniquely determine them. The purpose of the formulas is to make the reasoning coherent and auditable, not to turn uncertainty into an apparent measurement.

Define the events

S — superintelligence by 2050: a system exceeds strong human teams on unfamiliar, multi-week projects in mathematics, software, experimental science and engineering. Require at least 80% independently accepted completions in each domain, fixed budgets, limited corrective supervision and replication by two evaluators. These are proposed criteria. SI needs neither consciousness nor a humanoid body. AGI, broadly competent-human cognitive ability, is a weaker condition.
A — material abundance in 2050: a common group comprising at least 90% of humanity has reliable, affordable access to the entire basket: adequate food, housing, clean water, low-carbon electricity and essential healthcare. Its quality-adjusted real cost is at least 50% below 2025. The basket and regional baseline remain to be specified. This is a strict global target, not merely substantial progress or greater GDP.
D — doom by 2050: human extinction or irreversible loss of humanity’s capacity to recover a flourishing civilisation. A genuinely independent, viable off-Earth civilisation would preserve that capacity; an Earth-dependent outpost would not. A recoverable catastrophe is not D. If abundance was reached and then destroyed by doom, the endpoint counts as D, so A and D exclude one another.

Needs concern a decent life. Desires include comfort, novelty and status; some are unlimited or mutually incompatible. Everyone can have adequate housing, but everyone cannot own the most prestigious house. Abundance therefore cannot mean satisfying every desire. Nor does material abundance at one date prove ecological sustainability.

The quantities sought are P(S), P(A | S) and P(D | S). The vertical bar means conditional on. It does not mean “caused by”.

First establish the physical budget

Energy enables production. Machines, materials, ecosystems and institutions determine which services reach people. Jancovici’s starting point is useful: account for physical flows before attaching money or intelligence to them. Then distinguish a conservation law from a manufacturing problem that better tools can change. Physical framing.

Here, 1 W = 1 J/s; 1 GW = 10⁹ W; 1 TW = 10¹² W. One continuously operating GW supplies 8.76 TWh/year; 1 TWh = 3.6 × 10¹⁵ J. An exajoule, EJ, is 10¹⁸ J. Power measures a rate; energy accumulates over time. Nameplate capacity becomes output only after utilisation, losses and downtime are counted.

The following are scale tests using technologies already under development. That is a conservative selection rule, not a claim that we can enumerate inventions a century ahead. All future service budgets and engineering sensitivities are stated assumptions. They identify requirements for the probability factors; they do not supply empirical probability distributions.

Energy abundance means tens of terawatts

The world’s 2025 primary supply was about 655 EJ/year, or 21 TW. Final consumption was approximately 450 EJ in 2024, or 14 TW. Gross electricity generation reached 32,132 TWh in 2025, averaging 3.7 TW. These overlap: primary energy includes conversion losses; final energy reaches users; electricity is a carrier. Electrification can eliminate combustion losses, while storage and synthetic fuels introduce others. IEA energy accounts; final consumption.

2050 service-budget sensitivity: ten billion people. These are scenarios, not minimum needs or demand forecasts.
Final power per person	World final power	Annual final energy
1 kW	10 TW	316 EJ
2 kW	20 TW	631 EJ
4 kW	40 TW	1,262 EJ

Use the middle case for scale: if half is electricity, users receive 10 TW continuously, or 87,600 TWh/year. Generation must cover network and storage losses too. Capacity factor means average output divided by nameplate power. No single invention is required to meet the budget. These alternative portfolio tests expose different construction tasks; their capacities are not quantities to add:

Test	Calculation	What must scale?
All-solar electricity	10 TW / 20% capacity factor = 50 TW nameplate. At assumed 200 W/m² module rating and 80% packing: about 310,000 km².	Modules, sites, networks and balancing. A new 50-TW tranche over 25 years averages 2 TW/year, roughly three times PVPS’s 0.69-TW 2025 addition rate. Existing capacity and retirements belong in a full plan. PVPS
Storage across a full-load deficit	10 TW × 12 h = 120 TWh; seven days = 1,680 TWh, before losses/reserves.	Storage duration and chemistry. Battery-factory nameplate exceeded 4 TWh/year in 2025, serving vehicles and other uses too. Wind, firm supply, interconnection and flexible demand reduce the deficit. IEA
All-fission electricity	10 TW / 90% availability ≈ 11,000 one-GW reactors. A new fleet over 25 years averages 440 GW/year.	Factories, fuel and qualified construction. In 2025, 3 GW came online and 3 GW retired; 12.2 GW started construction. A two-order increase in commissioning requires a changed production system. IEA ledger

These tests are neither recommended portfolios nor proofs of impossibility. Current annual build rates can change. The IEA’s goal-aligned grid requirement—80 million km added or refurbished by 2040—averages nearly five million km/year from 2023, including distribution. SI could improve manufacturing, routing, installation and system operation. Actual connections and reliable output determine success. Grid assessment.

Several sources can serve the same budget

The principal route families. SI can improve designs, experiments, controls and manufacturing; each must still deliver a complete system.
Route	SI opportunity	Physical or deployment test
Solar PV and concentrated solar heat	Cells, structures, process yield, thermal storage and dispatch	Area, weather, network capacity and reliable service. Concentrated heat can serve industry directly. DOE
Onshore/offshore wind	Structures, forecasting, controls, installation and maintenance	At assumed 40% capacity factor, 10 TW average needs 25 TW nameplate: 2.5 million hypothetical ten-MW turbines. Spacing, weather and marine durability matter. DOE
Hydro and pumped storage	Turbines, construction and reservoir operation	Water, elevation and ecosystems. At 500 m head and 90% generating efficiency, 12 GWh needs about 10 million m³ transferred. Pumping consumes energy. DOE
Geothermal	Drilling, subsurface inference, exchange geometry and materials	Net heat extraction, pumping, reservoir lifetime, corrosion and seismicity. IEA
Fission, including smaller and advanced reactors	Standardisation, construction, inspection and fuel processing	Economic repeated builds, qualified components, fuel and cooling. Size alone does not establish lower cost. DOE
Fusion	Plasma control, magnets, breeding, materials and target production	Net electricity, fuel self-sufficiency, lifetime and economical replication; calculation below.
Biomass and waste	Catalysts, conversion and collection	Land, food, water and carbon accounting. A hypothetical 100 EJ/year is 3.2 TW thermal, about 1 TW electric at 30% conversion. IEA scenario
Tides, waves, ocean heat and salinity	Converters, membranes and durable marine equipment	Sites, construction and ecology. A heat engine between 298 K and 278 K has maximum efficiency 1−278/298 ≈ 7%, before real losses. EIA
Fossil energy with carbon capture	Separation chemistry, integration and monitoring	Fuel supply, energy penalty, durable storage and residual emissions. At the stated capture boundary, 95% capture still leaves 5%. IEA
Geological hydrogen	Exploration and extraction	A highly uncertain underground stock must become recoverable annual output. USGS

Hydrogen made from electricity and synthetic fuels are carriers. At assumed 50 kWh/kg H₂ production, 33.3 kWh/kg chemical energy and 50% electrical reconversion, the round trip returns about one third before transport and storage. Heat pumps work differently: a coefficient of performance of four delivers four units of heat from one electrical unit plus three from the environment. Thermal storage can supply heat without converting it back to electricity. Count the required service and temperature. Hydrogen reference; heat pumps.

A twenty-kilometre well needs more than hot rock

A deliberately simple 25 K/km gradient and 15°C surface give 515°C at 20 km. Real gradients vary; that extrapolation does not predict a site. At assumed rock density 2,700 kg/m³, vertical overburden stress is ρgh ≈ 0.53 GPa. Fluid and casing pressures differ. Drilling, circulation, materials and heat exchange must work together.

Heat in 1 km³ of rock cooled by 100 K:
Q = ρVcΔT = 2,700 × 10⁹ × 800 × 100 ≈ 2.2 × 10¹⁷ J = 60 TWh thermal.

Here 800 J/(kg·K) is assumed rock heat capacity. A plant producing 100 MW gross electric for 30 full-power years at 15% conversion needs 175 TWh thermal: 3 km³ at perfect recovery, or 15 km³ at assumed 20% recovery, before pumping and other loads. The accessible exchange volume and rate matter as much as depth. For example, 100 kg/s extracting 0.8 MJ/kg yields 80 MW heat, then 12 MW electricity at 15%, before other loads. Reservoir models.

Earth naturally loses about 47 TW of heat. That does not prohibit faster short-term heat mining; extraction then draws down a stock. SI could improve branching, stimulation, fluids, corrosion control and replacement-well schedules. Evidence already spans stages: Fervo reports 33 MW net commercially operating at Cape Station. Quaise’s July 2026 update described millimetre-wave drilling as approaching 1 km; 20 km remains an objective. Its superhot volcanic project seeks hot rock much shallower. Earth heat flow; Fervo; Quaise.

Fusion must close electrical, fuel and materials balances

The leading reaction is D + T → helium-4 + neutron + 17.6 MeV; D is deuterium and T tritium, hydrogen isotopes. One electronvolt, eV, is 1.602 × 10⁻¹⁹ J; keV and MeV multiply it by 10³ and 10⁶. About 80% of the energy leaves in a 14.1-MeV neutron that magnetic confinement cannot stop. The blanket must collect heat, protect components and breed tritium from lithium. For magnetic D–T near 8–20 keV, temperature is expressed as particle energy kBT, with kB Boltzmann’s constant. High gain requires particle density n, temperature T and energy-confinement time τ to combine at roughly nTτ = 4–10 × 10²¹ keV·s/m³. Achieving that plasma condition leaves the rest of the plant to build. ITER; National Academies.

NIF’s April 2025 shot yielded 8.6 MJ from 2.08 MJ of laser energy at the target: gain 4.13 and 2.4 kWh of fusion heat. Target gain excludes laser and facility electricity. ITER’s target Q = 10 similarly compares fusion power with heating delivered to plasma; ITER will not generate electricity. NIF; ITER.

Approximate net export:
Pnet ≈ ηthermal Pfusion − Pfusion/(Qηdriver) − Paux.

Use steady or cycle-averaged powers. Q = Pfusion/Pdelivered heating; ηthermal is heat-to-electricity efficiency; ηdriver converts driver electricity into delivered heating; Paux covers other electrical loads. The balance omits recovered driver heat and blanket multiplication. Illustrative efficiencies of 40% thermal and 50% driver leave 20% of fusion power at Q = 10, or 33% at Q = 30, before auxiliaries. The 50% is a generic heating assumption, not a demonstrated laser efficiency.

Plant requirement	Order-of-magnitude test	SI opportunity
Fuel and replication	17.6 MeV/reaction implies 56 kg tritium/GW fusion/full-power year. At 40% conversion, 1 GW gross electric needs at least 140 kg, before internal loads. Tritium’s half-life is 12.3 years.	Breed and recover enough fuel for current plants, losses, decay and new startup inventories. Better separation or dedicated breeders can change the constraint. NRC
Materials and heat rejection	At illustrative 1 MW/m² neutron loading: 4.4 × 10¹⁷ neutrons/m²/s. A 40%-efficient plant producing 1 GW gross electric rejects 1.5 GW heat.	Relevant irradiation tests, longer-lived walls, remote maintenance and better cooling. A 10°C once-through water rise would require about 36 m³/s; that is flow, not unavoidable freshwater consumption.
Fleet construction	Starting at 1 GW in 2035: two-year capacity doubling gives 180 GW by 2050; five-year doubling gives 8 GW.	Earlier starts, parallel factories, stronger magnets and faster replication can change the trajectory. Fuel and component supply must keep up.

Why a breeding ratio just above one may not replicate the fleet

Assume burn B = 56 kg/year, operating inventory I = 5 kg and recovered breeding ratio b. Let J be accumulated fuel for another plant and λ = ln(2)/12.3 year⁻¹. Then dJ/dt = (b−1)B−λI−λJ. At b = 1.05, another 5 kg accumulates in about 2.1 years. At b = 1.01, decay limits J to 4.94 kg: it never reaches 5 kg under these assumptions. Startup inventory and breeding efficiency are design variables, not constants imposed on every reactor.

Inertial fusion has another factory problem: ten shots/second requires about 3 × 10⁸ targets/year. At an assumed $1 each, targets alone cost about $36/MWh for a continuously operating one-GW-electric plant; at $0.10, about $3.6/MWh. LLNL’s driver programme targets order-ten-Hz repetition and billion-shot optics. SI could improve the target factory as well as the plasma. LLNL.

Fission’s fuel chain deserves the same scrutiny. Holding a dated requirement of 150 tonnes uranium/GW-capacity-year fixed, the 11-TW endmember needs about 1.7 million tonnes/year. The Red Book’s 2023 identified resource category of 7.9 million tonnes would cover fewer than five years if frozen. It is not fixed: exploration, prices and closed fast-reactor cycles can expand usable supply greatly. Mines, enrichment, fabrication and recycling must scale alongside reactors. NEA/IAEA baseline; closed-cycle potential.

Convert gigawatts into useful work

The IEA estimates all data centres used 485 TWh in 2025, projecting 950 TWh in 2030: 108 GW average facility power. Company announcements use other boundaries. OpenAI reported more than 10 GW secured in April 2026; that is a pipeline, not metered operation. Crusoe’s March release documented two energised 100-MW buildings at Abilene, alongside larger future targets. Do not add campus headlines to a global aggregate that already includes them. IEA; OpenAI; Crusoe.

A token is a model-specific encoded unit. A FLOP is a floating-point operation. Neither measures a scientific result. For a short-context transformer, arithmetic per generated token is roughly 2Nactive FLOP, where Nactive counts active parameters. A 100-billion-parameter dense model needs order 2 × 10¹¹ FLOP/token. Attention, memory movement, serving and verification add work.

The memory calculation explains a wide energy range. At two bytes/parameter, those weights occupy 200 GB. Three H100s provide 240 GB memory and about 10 TB/s aggregate peak bandwidth: streaming the weights takes order 20 milliseconds, versus 67 microseconds at peak arithmetic. Assuming 700 W per chip and facility/accelerator power ratio two gives about 84 J per batch-one token. Sharing weight reads across sixteen sequences takes that component toward 5 J/token; their context memory and other work still have to fit. These are engineering assumptions, not fleet measurements or physical lower bounds. H100 specifications.

Tokens/person/year = EDC × 3.6 × 10¹⁵ × f / (eN).
EDC is annual facility electricity in TWh; f its allocated inference fraction; e full-facility joules per reference token; N population.

For 2030, choose f = 0.5 × 0.7 = 35%: half of data-centre energy for AI, 70% of that for inference. At e = 10 J/token and 8.5 billion people, the result is 38,600 tokens/person/day. At 1–100 J/token it spans roughly 386,000–3,860. The assumed e already includes utilisation, hosts and cooling; subtracting those overheads again would double count.

A bounded reference-workload scenario, not a forecast of SI demand or quality. Only the 2030 electricity figure is an IEA projection; later budgets, allocation and efficiency are author assumptions.
Year	DC electricity	Inference share f	People	J/token	Tokens/person/day	Cumulative since 2025
2030	950 TWh	35%	8.5 billion	10	38,600	53.7 million
2040	2,000 TWh	45%	9.2 billion	2	482,000	724 million
2050	3,000 TWh	56%	10 billion	1	1.66 million	4.26 billion

The 2050 allocation is 3.4% of the 87,600-TWh central electricity budget. Keeping e fixed at 10 J gives only 166,000 tokens/day in 2050, rather than 1.66 million. Cumulative tokens become 0.77 billion rather than 4.26 billion. That tenfold endpoint difference comes from assumed efficiency, not extra power. Cumulative values integrate linear electricity, allocation and population paths with geometric efficiency between checkpoints.

A job requiring an assumed 10⁸ token equivalents, including failed searches, permits 0.14 accepted jobs/person/year from the 2030 budget, or about six in the improving-efficiency 2050 case. The job cost is hypothetical. Larger models and longer reasoning can absorb hardware gains; better verifiers and experimental design can reduce them. The decisive metric is joules per independently accepted task at fixed quality.

A ten-GW facility also needs chips, memory and replacement capital

Assume facility/IT power ratio 1.2, accelerators consuming 60% of IT power and 700 W per H100-equivalent. Ten GW then powers about 7.1 million accelerators, with 570 PB of memory at 80 GB each. An 814-mm² die yields order 63 gross dies on a 300-mm wafer; at assumed 70% usable yield, the initial logic requirement is about 160,000 wafers. Memory, packaging, networks and leading-edge tool capacity remain separate requirements. Die specification.

At assumed $30,000 installed server/network capital per accelerator-equivalent and five-year turnover, replacement is order $43 billion/year. Electricity at assumed $50/MWh costs $4.4 billion/year. Cheap electricity helps without eliminating capital or supply-chain costs. These prices and lifetimes are scenario inputs, not vendor quotations.

Commissioning also matters: a ten-GW project waiting three years and ramping over the next two supplies only ten GW-years in five years, rather than fifty if counted at full power throughout. A one-billion-person group receiving 20% of the global inference pool gets 1.7 times the 2030 global per-capita budget, not 8.5 times; hardware allocation must accompany the smaller denominator.

Test the space-abundance thesis by mass and lifetime

Musk’s strongest space argument is that cheaper reusable launch can open a much larger energy domain. SpaceX’s June 2026 prospectus proposes solar-powered orbital computing, ultimately adding 100 GW/year at about 100 kW per tonne. The same document describes unproven technology and commercial uncertainty. The FCC accepted an application for up to one million satellites for review; that is not approval or operating capacity. SpaceX prospectus, pp.82/97; FCC notice.

Our deductions from stated physical laws and explicit engineering assumptions. These are system requirements, not impossibility arguments.
Requirement	Calculation	Implication
Solar collection	1 GW / (1,361 W/m² × assumed 30% conversion) = 2.45 km² of cells.	Assumes continuous normal illumination. Losses, degradation, eclipse and structure increase requirements. NASA irradiance
Heat rejection	A = P/(εσT⁴). For 1 GW heat and emissivity ε = 0.9: 2.42 km² at 300 K, or 1.31 km² at 350 K.	Emitting surface, assuming unobstructed cold sky. Two-sided panels can have half that panel area. Sun/Earth heating, view geometry and coolant limits matter. Vacuum eliminates convective cooling. NASA thermal design
Annual launched mass	100 GW/year × 10 kg/kW = one million tonnes/year; at 100 tonnes delivered/launch: 10,000 launches/year, about 27/day.	Uses the proposed specific power as an optimistic full-system assumption. Launch, factories, propellant, assembly and operations must reach this flow.
Lifetime launch charge	At 10 kg/kW and five-year full utilisation: $100–1,000/kg contributes 2.3–23 cents/kWh of compute power.	Launch only. Hardware, financing, communications, losses and servicing add costs. Prices and life are sensitivity inputs.

σ is the Stefan–Boltzmann constant, 5.67 × 10⁻⁸ W/(m²·K⁴); T is radiator temperature in kelvin, with 300 K ≈ 27°C. Raising temperature nearly halves radiator area here. Doing so while retaining efficient, reliable computation is an engineering question. SI could improve chips, thermal structures, optical networks and robotic servicing. GAO identifies large-scale cooling, communications, radiation damage and orbital congestion as unresolved issues. GAO assessment.

The favourable mechanism is real: near-continuous sunlight and valuable bits returned to Earth. The full economic test is useful compute-years per launched kilogram and dollar. With five-year hardware life, a 500-GW fleet needs 100 GW/year merely to replace retirements. Modular chip replacement could separate computing life from the solar/radiator platform’s life.

Using power in orbit differs from beaming it to Earth. At assumed 15% overall sunlight-to-grid conversion, one delivered GW needs 4.9 km² collecting area, plus transmission and ground reception. NASA’s 2024 study placed two 2050 point designs at $0.61 and $1.59/kWh in its baseline, but $0.04 and $0.08 under combined favourable sensitivities. Those are conditional designs, not universal limits or orbital-compute prices. NASA study.

Local lunar oxygen, propellant or shielding can avoid Earth launch mass. Importing bulk commodities faces a further extraction, processing and transport chain while competing with terrestrial supply. A growing space economy therefore does not automatically halve Earth’s food, housing or medicine costs. A self-sufficient settlement could protect humanity from some Earth-specific catastrophes; an outpost dependent on Earth cannot. Independence requires replacing essential food, tools, electronics and medicines after Earth supply stops. NASA resource-use research.

There is a genuine long-run thermal limit. Spread 20 TW of entirely additional heat over Earth’s 5.1 × 10¹⁴ m²: 0.04 W/m²; at 10,000 TW it becomes 20 W/m². Terrestrial solar redirects an existing flow; imported space power and nuclear energy add heat on this timescale. Orbital computing rejects heat off Earth. This supports space expansion at extreme scales, without making it necessary for the article’s tens-of-TW 2050 service budget. Energy-budget geometry.

The ocean buffers damage and prolongs recovery

The ocean contains about 1.335 billion km³ of water. At rounded density 1,025 kg/m³ and heat capacity 4,000 J/(kg·K), its mass is 1.4 × 10²¹ kg and its heat capacity 5.5 × 10²⁴ J/K. The IPCC estimates 396 zettajoules gained during 1971–2018; one ZJ is 10²¹ J. That is 270 TW historical average uptake, driven by the planetary energy imbalance. Divided across the whole ocean it gives only 0.07°C equivalent warming. Neither figure measures today’s uptake or sea-surface temperature. NOAA volume; properties; IPCC heat inventory.

The apparent smallness of that average is misleading. An illustrative upper 100 m contains only 2.7% of ocean volume; organisms experience local temperatures and marine heatwaves. Surface exchange can be rapid, while deep-water return takes decades to millennia. Pumping heat from ocean to air redistributes it and adds machinery’s waste heat. Durable cooling requires changing Earth’s radiative balance. Ocean exchange times.

Carbonate chemistry has no shortcut through free protons

pH = −log₁₀(aH⁺), with aH⁺ hydrogen-ion activity. A change from 8.2 to 8.1 means 26% higher activity; the water remains alkaline. Added CO₂ shifts carbonate toward bicarbonate, reducing material available for many calcium-carbonate structures. Aragonite saturation Ω compares the calcium/carbonate ion product with its solubility product. Ω = 1 is chemical saturation; the assessment’s precautionary global surface boundary is 2.86. Ecological stress can occur before exposed mineral dissolves. Species responses vary, and warming-driven bleaching has a separate mechanism. NOAA pH; boundary definition; biological effects.

Near pH 8, free hydrogen ions are order 10⁻⁸ mol/kg; dissolved inorganic carbon is order 10⁻³ mol/kg. Removing H⁺ causes those much larger carbonate/bicarbonate reservoirs to re-equilibrate. Multiplying ocean volume by the apparent free-proton increase therefore badly mis-sizes a neutralisation programme. A defensible intervention needs alkalinity, dissolved carbon, temperature, salinity and gas exchange. NOAA carbon chemistry; National Academies.

Carbon stocks and repair throughput. These are arithmetic tests, not recovery forecasts.
Test	Calculation	Consequence
Atmospheric inventory	1 ppm = 2.12 Gt carbon × 44/12 ≈ 7.8 GtCO₂. A 100-ppm reduction is about 780 GtCO₂.	At hypothetical 10 Gt/year net removal: 80 throughput-years, before ongoing emissions and carbon-cycle exchanges. Current emissions are order 40 Gt/year. NOAA; carbon budget
Mineral alkalinity	Mg₂SiO₄ + 4CO₂ + 4H₂O → 2Mg²⁺ + 4HCO₃⁻ + H₄SiO₄. Ideal pure-mineral ratio: 176/141 ≈ 1.25 tCO₂/t rock.	Ten GtCO₂/year needs at least 8 Gt rock/year under this route’s idealisation. Impurities, dissolution, precipitation and air–sea exchange change actual net uptake. Reaction and limitations
Processing energy	8 Gt × assumed 100–1,000 kWh/t grinding = 800–8,000 TWh/year, excluding other operations.	UNEP estimates 50 Gt/year of sand/gravel already used: bulk scale is conceivable, but the material, processing and ecological controls differ. UNEP

SI could improve mineral selection, reactors, low-carbon supply, shipping and monitoring. Demonstrate net, additional, durable carbon removal and acceptable ecological effects. Avoiding emissions reduces the industry needed for later repair. Surface chemistry can improve through CO₂ removal; deep recovery is slower. Restoring chemistry does not recreate an extinct species. IPCC reversibility; Intervention research guidance.

Nine boundaries describe interacting risks

The September 2026 assessment classifies seven boundaries as breached. These are precautionary risk zones, with different indicators and uncertainties. They neither mark nine independent switches nor supply extinction probabilities. A classification can change through better observations or revised baselines without a sudden physical jump. Current assessment; framework.

Process	What can fail?	2026 global classification
Climate	Heat, rainfall patterns, ice and sea level	Breached; high risk
Biosphere integrity	Diversity and ecological functions	Breached; high risk
Land systems	Forests, habitats, carbon and water regulation	Breached; increasing risk
Freshwater	River flows and soil moisture	Breached; increasing risk
Nitrogen/phosphorus flows	Nutrient pollution, oxygen loss and ecosystem disruption	Breached; high risk
Novel entities	Persistent/toxic substances outpacing assessment and control	Breached; high risk
Ocean acidification	Carbonate chemistry and dependent organisms	Breached; increasing risk
Stratospheric ozone	Protection against ultraviolet radiation	Within global boundary; recovering
Aerosols	Particles altering climate and rainfall; regional pollution	Within global boundary; regional harms remain

Recovery is possible: restrictions phased out 99% of controlled ozone-depleting substances. Recovery also takes time. Energy and intelligence can improve protection, adaptation and repair, while ecological function and permanent biological loss require their own accounts. UNEP.

Cheap energy buys services through real production chains

Use ten billion people again. Water, calories and material throughput expose different requirements. These calculations hold their stated technology and efficiency assumptions fixed; better processes can change them.

Service	Scale calculation	What SI could change
Domestic water	Assumed 50 litres/person/day gives 183 km³/year. At 3 kWh/m³, desalination needs 550 TWh/year. Lifting water one kilometre adds at least 2.7 kWh/m³ before losses.	Membranes, reuse, pumps, leak detection and networks. Reliable inland delivery, brine and capital remain. This excludes agriculture, which accounts for 72% of global withdrawals. IEA; FAO
Food	Assumed 2,500 kcal/person/day = 1.2 TW edible chemical energy. At hypothetical 5% electricity-to-edible conversion, all-electric food needs 24 TW electric.	Breeding, field robots, greenhouses, microbial protein and fermentation. The efficiency is a sensitivity, not a ceiling. Field crops use sunlight directly; artificial lighting changes the energy budget. Count balanced diets, soil, water and price. Microbial-food system model
Nitrogen fertiliser	Hydrogen for assumed 200 Mt ammonia/year needs 1,765 TWh/year at 50 kWh/kg H₂, before air separation and synthesis: 200 × 10⁹ kg × 3/17 × 50.	Low-carbon ammonia, biological fixation, precise application and nutrient recovery. Cheap production does not prevent runoff, air pollution or N₂O from overapplication. IEA; FAO
Phosphorus	73 Gt phosphate-rock reserves / 0.25 Gt annual output ≈ 300 throughput-years.	Wastewater/manure recovery and better placement. Phosphorus has no biological substitute; the ratio is not a depletion date. Economic reserves, demand and access change. USGS 2026
Metals and recycling	If annual installations double every five years and products last fifteen, retiring cohorts supply only 2⁻³ = 12.5% of current material demand, even at perfect recovery and fixed material intensity.	Substitution, collection, disassembly, chemistry and lower material intensity. Rapidly growing stocks still need new inputs; mature stocks can recycle a much larger share. IEA scenarios

Land and ecosystems remain separate accounts. FAO reports 4.6 billion hectares of agricultural land in 2024. Better yields and microbial foods could reduce pressure; water savings at one field can disappear through expanded irrigation or reduced downstream flows. Restoration needs biological recovery, connectivity and lasting protection. Counting hectares started or calories manufactured does not measure those functions. FAO land; basin accounting; restoration assessment.

Materials chemistry also matters. Limestone calcination, CaCO₃ → CaO + CO₂, releases 44 g CO₂ per 56 g CaO. Illustrative clinker containing 65% carbonate-derived CaO emits 0.51 t process CO₂/t clinker even with clean heat. Hydrogen reduction, Fe₂O₃ + 3H₂ → 2Fe + 3H₂O, needs at least 54 kg H₂/t iron; at assumed 55 kWh/kg, hydrogen alone needs 3 MWh/t. SI can develop alternative feedstocks, low-clinker cement, electrochemical routes and better recycling. These are constraints of specified processes, not immutable limits on every future material. DOE cement alternatives; ironmaking example.

Medicine needs evidence and a delivery architecture

A useful treatment must alter a disease mechanism, reach the right cells, produce meaningful benefit, remain acceptably safe, and reach patients. SI can improve every link. Generative chemistry and proteins expand candidates; mRNA platforms reuse production methods; gene editing, engineered cells and organ replacement change what can be treated. Diagnostics, prevention and existing medicines can improve essential health before new cures arrive.

Developing route	Evidence today	Remaining test
AI-designed medicines/proteins	Rentosertib: 71 patients, twelve weeks, phase 2a; safety/tolerability primary, efficacy exploratory. Designed binding proteins also have laboratory evidence.	Meaningful durable benefit, specificity, tissue exposure and manufacturability. Trial; protein experiments
Gene editing and reusable delivery	Approved patient-specific Casgevy; a personalised liver-editing case went from diagnosis to treatment in six months.	Safe delivery across tissues, durable effect, production and access. One patient does not establish a universal platform. FDA label; NIH case
Replacement cells and organs	Early engineered-islet human evidence and formal gene-edited pig-kidney trials.	Immune compatibility, infection, long-term function and reproducible supply. Islet report; organ trial
Vaccines, diagnosis and prevention	Working products and expanding manufacturing platforms.	Prospective health benefit, quality and coverage. Aging biomarkers also require validation against meaningful human outcomes. WHO platforms; aging research

The information requirement has a scale. For two equal groups, a continuous outcome, 80% power and 5% two-sided significance, n per group ≈ 15.7/d², where d is effect size divided by the outcome’s standard deviation. An effect d = 0.5 needs about 63/group; d = 0.1 needs 1,600. Separately, a one-in-10,000 independent adverse event needs about 30,000 exposures for a 95% chance to observe at least one: n ≈ −ln(0.05)/p. These simplified tests neither prove causation nor reveal a delayed event before it occurs. NIH methods.

Timelines can move through stronger treatments, better experiments and validated surrogate endpoints. FDA’s follow-up guidance of up to fifteen years for genome editing concerns surveillance after administration, not a universal fifteen-year preapproval wait. Manufacturing can change too: ten million hypothetical patient-specific treatments/year occupying a position for fourteen days require 384,000 concurrent positions, before downtime. Off-the-shelf cells or in-vivo treatment could change that architecture. Surrogates; follow-up guidance.

Factories can accelerate the construction of factories

History supports considerable industrial elasticity. Tesla and Panasonic combined cell expertise, buildings, equipment and suppliers; Tesla reported more than 37 GWh/year of Nevada cells in 2023. China holds over 80% of global lithium-ion manufacturing capacity. The first Liberty ship took 244 days from keel laying to delivery; by December 1943 the average was 39 days. That clock excludes upstream factory creation and component production. Standardisation, prefabrication and coordination changed assembly, while turbine shortages constrained faster ships. Tesla; IEA; first ship; later average.

Five million industrial robots operated in 2025. A hypothetical 100 million by 2050 requires 12.7% net annual growth. At assumed 5% annual retirement, a stock near that target needs about 18 million installations/year, thirty times the current flow. Count robot stock × productive hours × accepted output/hour, including supervision, maintenance and rejected work. A robot count alone has no fixed human-worker equivalence. IFR baseline.

A simple replication model makes the time assumption explicit: dK/dt = αK(t−τ)−δK. K is productive capacity, α the rate at which it finances new capacity, τ construction delay, and δ retirement. At assumed α = 0.40/year and δ = 0.05/year, zero delay permits two-year doubling; a three-year delay stretches it to 3.8 years. SI could lower construction time, improve utilisation or reduce equipment cost. Complementary supply and demand still constrain the result.

Learning also follows accumulated production. A hypothetical 20% unit-cost reduction per cumulative-output doubling makes tenfold output reduce cost to 48% of its starting value. It does not promise automatic annual cost decline. The strongest cross-cutting candidates are machines that construct and commission equipment, autonomous laboratories that produce verified evidence, and maintenance that extends useful life. They improve several routes together. Historical PV learning model; laboratory robots; NIST experiments.

Turn requirements into conditional hypotheses

The calculations support a specific hypothesis: 2050 abundance has several physically credible routes, while deployment and access remain demanding. Fusion, deeper wells and orbital computing expand the options. Solar, wind, existing nuclear, efficiency and ordinary infrastructure can deliver gains before those routes mature. A portfolio can substitute between technologies; its common dependence on factories, networks and institutions remains.

For the energy condition E below, the relevant event is sufficient annual supply and reliable regional delivery at the required service level. In each surviving SI scenario, P(E | S,z,¬D) = P(supply ≥ requirement, reliability ≥ requirement | S,z,¬D). Here z denotes the shared timing/institutional scenario; U separately requires affordable access. Numerical probabilities need distributions for demand, construction, cost and failures. A large resource-to-demand ratio cannot provide them.

Food, water and material throughput constrain M; clinical validation and delivery constrain H; actual coverage and prices constrain U. Shared energy, finance and manufacturing make their uncertainties dependent. The strict abundance basket still needs a regional 2025 price and quality baseline. The service examples above do not settle that resolution rule. With those limits exposed, the following model makes my judgments and their consequences explicit.

1. P(SI): three necessary conditions

Let G mean broad competent-human research capability, X superiority over strong expert teams, and R a dependable integrated system. S = G ∩ X ∩ R. The chain rule gives:

P(S) = P(G) × P(X | G) × P(R | G, X)
= 0.98 × 0.85 × 0.80 = 66.64% ≈ 67%.

Conditional judgments for 2050. Evidence supports the reasoning, not the exact percentages.
Necessary condition	Evidence in favour	Evidence against certainty	Input
G: broad research competence	Current models work across coding, science, images, CAD and unfamiliar interactive environments. Twenty-four years remain.	Unfamiliar experimental work still exposes gaps. Benchmarks cover only part of the target.	98%
X, given G: broad expert superiority	New mathematical results, verified reinforcement learning, better algorithms and increasing compute. Human ability is not a demonstrated physical ceiling.	Superhuman results in some domains do not establish broad superiority. Reliable feedback and fresh data can be expensive.	85%
R, given G and X: dependable integration	Verification, retries, tools, specialist controllers and laboratory automation offer concrete engineering paths.	Long projects compound errors. Integrity, cost, experimental throughput and real-time control remain unresolved.	80%

These inputs allocate 2%, 15% and 20% residual failure probabilities at successive conditions. The 80% probability assigned to R is distinct from its 80% task-completion criterion. They are judgments, not benchmark pass rates. I give the most remaining uncertainty to making high capability dependable. The current evidence makes an early-AGI interpretation of Astra and Opus reasonable; it does not already establish this stronger SI event. Current evaluations and sources.

The electricity, memory and token budgets constrain available attempts and their cost. They do not establish the probability that learning reaches the required quality.

Reliability check: 100 critical stages, each independently correct with probability 99%, yield 0.99¹⁰⁰ ≈ 37% project success. At 99.9%, the result is about 90%. Actual errors are correlated and recovery changes the calculation. The example explains why local brilliance does not settle R.

Use the same worlds for abundance and doom

Arrival timing changes the time available to build. Apply the same SI chain at three deadlines:

Deadline	P(G)	P(X \| G)	P(R \| G, X)	P(SI by deadline)
2030	85%	45%	50%	19.1%
2040	95%	70%	70%	46.5%
2050	98%	85%	80%	66.6%

These are cumulative judgments, not independent events. Conditional on SI by 2050, the chance it arrives by 2040 is 0.4655 / 0.6664 ≈ 70%. For tractability, represent early arrivals by 2035 and later arrivals by 2045: roughly fifteen versus five years for deployment. This compresses real variation, including potentially much earlier arrival.

Next distinguish a deployment-capable world, where finance, infrastructure and institutions can coordinate large projects, from a fragmented world. Capable does not mean safe. I assign P(capable | early SI) = 65% and P(capable | late SI) = 55%. This assumes early arrival somewhat favours strong industrial conditions, while later arrival faces more accumulated environmental stress. The sign is contestable: later arrival could instead allow better preparation.

Shared scenario	Weight, conditional on SI	Calculation
Early / capable	45.4%	69.85% × 65%
Late / capable	16.6%	30.15% × 55%
Early / fragmented	24.4%	69.85% × 35%
Late / fragmented	13.6%	30.15% × 45%

Both outcome calculations use these weights w_z. Timing and institutions are associated explicitly. We do not give abundance an efficient, peaceful world and then calculate doom in an unrelated world.

2. P(doom | SI): a shock is not extinction

I calculate doom first because abundance requires survival. The shock probabilities cover 1 November 2026 through the end of 2050, conditional on SI arriving in the stated scenario. For each pathway i, let qᵢ be the probability of a qualifying shock and rᵢ the probability of irreversible doom given that shock. Both are conditional on the scenario and on no preceding listed pathway having caused doom. The ordering below is bookkeeping, not a temporal sequence. It prevents overlap from being added twice.

Early / capable reference case. q and r are elicited judgments; source links establish mechanisms, not these rates.
Pathway	For higher risk	Against irreversibility	q	r	q × r
Loss of AI control	Strategically capable systems could defeat oversight or gain persistent control over critical operations.	Capability is not harmful intent. Containment, verification and limits on deployment can prevent a failure from becoming irreversible.	20%	20%	4.000%
Nuclear war	Roughly 10⁴ nuclear weapons exist; conflict, automation and false confidence can amplify escalation.	Deterrence, diplomacy, safeguards and geographic diversity can prevent or limit escalation. Even a terrible war need not end recovery.	10%	5%	0.500%
Severe pandemic	Natural emergence, connected populations and misuse can outpace defensive deployment.	Surveillance, medicine, manufacturing and adaptation improve response. Mass mortality and extinction are different events.	20%	1%	0.200%
Ecological and food-system shock	Warming, ocean chemistry, water and ecosystem losses can damage several essential systems together.	Geographic diversity, substitution, adaptation and lower emissions preserve recovery paths. A crossed boundary is not an extinction probability.	35%	0.2%	0.070%
Critical supply breakdown	Concentrated energy, mineral, refining and trade networks can transmit persistent shortages.	New supply, substitution, recycling and demand adjustment can restore services. Rate constraints are not exhaustion of all material.	25%	0.1%	0.025%

A shock is an intermediate event, defined below; it does not automatically meet D. The r inputs are especially uncertain: history contains catastrophes, not repeated trials of human extinction. The chosen thresholds make the assumptions explicit; they do not provide a dataset that calibrates the probabilities.

Add r₀ = 0.5% for omitted doom mechanisms, conditional on surviving the listed pathways. This is a disclosed model-completeness allowance, not a measured background rate. Then, within scenario z:

d_z = 1 − (1 − r₀) ∏ᵢ(1 − qᵢ,z rᵢ,z)
P(D | S) = Σ_z w_zd_z = 7.38% ≈ 7%.

This product is the chain rule for conditional survival, not an assumption that war, disease, climate and resources are independent. If the inputs were marginal risks, the formula would generally be wrong. Combined cascades belong in the conditional inputs. Reordering pathways requires reassessing them.

Shock definitions, all doom inputs, and the four scenario calculations

Use these working severity conventions when interpreting q:

Control: a strategically capable AI system retains consequential operational access outside effective human oversight despite attempts to stop or contain it, with access sufficient to affect critical systems.
War: at least one nuclear weapon is used in interstate war.
Disease: a pandemic causes excess deaths of at least 1% of the global population within five years.
Ecology: a new multi-region ecological shock deprives at least 10% of humanity of adequate food or water for at least one year.
Resources: supply interruptions reduce delivered energy or essential material inputs by at least 20% across economies containing at least 25% of the world population, for at least one year.

These are model definitions, not universal scientific thresholds. They can overlap; the no-prior-doom conditioning handles overlap in D. A shock can occur without being the pathway that ends recovery. Changing its threshold requires revisiting both q and r.

Every cell is q × r: this pathway’s doom probability conditional on no preceding pathway having caused doom. The exact one-third entry is displayed as 33.3%.
Pathway	Early / capable	Late / capable	Early / fragmented	Late / fragmented
Loss of AI control	20% × 20.0% = 4.000%	7% × 20.0% = 1.400%	36% × 33.3% = 12.000%	14% × 33.3% = 4.667%
Nuclear war	10% × 5.0% = 0.500%	6% × 5.0% = 0.300%	19% × 9.5% = 1.805%	13% × 9.5% = 1.235%
Severe pandemic	20% × 1.0% = 0.200%	12% × 1.0% = 0.120%	36% × 2.0% = 0.720%	24% × 2.0% = 0.480%
Ecological and food-system shock	35% × 0.2% = 0.070%	30% × 0.2% = 0.060%	58% × 0.4% = 0.232%	55% × 0.4% = 0.220%
Critical supply breakdown	25% × 0.1% = 0.025%	22% × 0.1% = 0.022%	44% × 0.2% = 0.088%	40% × 0.2% = 0.080%
Other doom, after surviving all above	0.5%	0.5%	0.5%	0.5%
Total d_z	5.24%	2.38%	14.91%	7.04%

Late arrival reduces the assumed exposure to AI loss of control. Ecological and resource shocks remain relatively high because accumulated pressures do not wait for SI. All q values are full-horizon probabilities conditional on SI arrival in that cell, including the selection that this entails. They are not annual hazards multiplied by fifteen or five.

Fragmentation raises both shock likelihood and the chance that recovery fails. The values are explicit stress judgments, not measured governance multipliers. In the early/capable case, for example, the calculation is 1 − 0.995 × 0.96 × 0.995 × 0.998 × 0.9993 × 0.99975 = 5.24%.

What drives this result? Under the chosen first-pathway ordering, AI control contributes 5.6 percentage points of the 7.4% total. This is a bookkeeping allocation, not a uniquely identified causal share. The physical analysis does not calibrate that control-failure or irreversibility judgment. Historical collapses inform propagation and recovery, rather than an extinction frequency. Historical evidence.

3. P(abundance | SI): four conditional requirements

For each scenario, condition first on survival through 2050. Let E be sufficient reliable low-carbon energy, M productive material services, H clinically valid essential healthcare, and U actual affordable access. E, M and H require sufficient supply; U requires the entire basket reach a common group comprising at least 90% of humanity at the stated quality and cost. Thus A = ¬D ∩ E ∩ M ∩ H ∩ U. Later factors condition on all preceding ones. Recoverable wars and ecological damage can still depress them:

a_z = P(E | S,z,¬D) × P(M | E,S,z,¬D)
× P(H | E,M,S,z,¬D) × P(U | E,M,H,S,z,¬D).

Evidence for and against the necessary service conditions. Technology choices are routes to a condition, not independent tickets in a lottery.
Condition	In favour	Against
E: Reliable low-carbon energy	Existing technologies and large annual deployment; SI can improve the whole system.	Annual generation is not reliable regional supply. Grids, storage, competing demand and commissioning matter.
M: Productive material services, given E	Factories and robots can produce better production equipment. Fast capacity growth is already observed in batteries and solar.	Housing, food and water infrastructure require complementary equipment, sites, labour and maintenance. A prototype is not delivered output.
H: Validated essential healthcare, given E and M	Existing prevention and treatments leave large delivery gains. AI can improve trials, diagnosis, manufacturing and care.	Biology remains uncertain. Clinical follow-up and real delivery take time. This gate does not require curing every disease.
U: Affordable global access, given E, M and H	Lower production costs, competition, public provision and effective distribution can spread benefits.	Exclusion, rents, conflict and weak institutions can prevent cheaper technical supply from reaching 90% of people.

Conditional probabilities after survival and all preceding conditions. Global access is the largest institutional discount.
Condition	Early / capable	Late / capable	Early / fragmented	Late / fragmented
E	90%	65%	80%	45%
M	85%	60%	70%	40%
H	80%	65%	65%	45%
U	75%	60%	35%	25%
Product a_z	45.9%	15.2%	12.7%	2.0%

Early SI in capable institutions makes each supply condition fairly likely. The target remains demanding: 90% of humanity, reliable services, and half the real cost. In fragmented worlds, access stays low even after supply succeeds. These probabilities express that judgment. They are not inferred directly from solar watts or robot counts.

Smil writes that “the key transformations have to unfold on enormous scales.” His 2019 accounting puts annual cement consumption at about 4.5 billion tonnes. That constrains the scale of the task, not the future rate at which factories can expand. Book reference.

The energy portfolio, factory replication and clinical/production requirements constrain these inputs. The gates concern delivered outcomes; fusion, orbital computing or humanoids need not all succeed.

P(A | S) = Σ_z w_z(1 − d_z)a_z
= 25.12% ≈ 25%.

Scenario	w_z	Survival	a_z after survival	Contribution to P(A \| S)
Early / capable	45.4%	94.8%	45.9%	19.75%
Late / capable	16.6%	97.6%	15.2%	2.46%
Early / fragmented	24.4%	85.1%	12.7%	2.65%
Late / fragmented	13.6%	93.0%	2.0%	0.26%

Because the same worlds and survival term appear in both calculations, P(A | S) + P(D | S) ≤ 1 automatically. The remaining 68% is neither abundance nor doom. It includes substantial progress, inequality, stagnation and recoverable disasters.

Which assumptions decide the answer?

Stress tests, not confidence intervals. Each row changes the stated inputs while holding the others fixed.
Changed assumption	Computed consequence
SI reliability R: 60% → 95%	P(S) = 50% → 79%. This row concerns SI alone; outcome mixtures must be recomputed if its timing distribution changes.
Capable institutions: 35% → 80%, in both arrival bins	P(A \| S): 17% → 30%; P(D \| S): 10% → 6%.
All four abundance conditions at their low → high scenario values below	P(A \| S): 7% → 51%.
Irreversibility after AI-control failure: half → twice the reference values	P(D \| S): 5% → 13%.
Omitted-mechanism allowance: 0% → 2%	P(D \| S): 6.9% → 8.8%.

Exact abundance stress-test inputs

Each cell is low → high. These are joint stress cases; no independent random draws or statistical coverage is claimed.
Condition	Early / capable	Late / capable	Early / fragmented	Late / fragmented
E	80% → 97%	40% → 85%	60% → 90%	20% → 70%
M	65% → 95%	35% → 80%	45% → 85%	20% → 65%
H	60% → 95%	40% → 85%	45% → 85%	20% → 70%
U	50% → 90%	30% → 80%	15% → 60%	5% → 45%

The result is conditional on the model. The arithmetic is exact given its inputs; the inputs are not calibrated frequencies. Splitting one intuition into several percentages does not create evidence. The tables make disagreements specific: a different view of reliable autonomy, factory replication, clinical validation, access or irreversibility changes a named factor and a reproducible result.

Nor is 7% the probability that AI causes doom. It is P(D | S). Overall risk needs the no-SI branch: P(D) = P(S)P(D | S) + [1 − P(S)]P(D | ¬S). A causal comparison needs further assumptions about otherwise comparable worlds. This assessment does not supply them.

The most useful observations are therefore independently validated long-project reliability; useful robot output per unit capital and energy; commissioned supply rather than announced capacity; delivered clinical benefit; affordable access; and evidence that control and institutions remain effective as capability rises. They determine whether the conditional inputs deserve to move.

Technical evidence and sources

The physical calculations are visible above. These notes retain evaluation details, methodological limits, historical evidence and source provenance.

Terms: intelligence, abundance, desires and physical quantities

AGI: Ability to learn and perform a broad range of unfamiliar cognitive tasks at roughly competent-human level. This working definition requires neither consciousness nor human-shaped embodiment. SI demands reliable superiority over strong expert teams.
Capability, autonomy, reliability: Capability is what a system can do. Autonomy is how little corrective human involvement it needs. Reliability is its probability of meeting a defined standard under specified conditions. One impressive success does not establish any of these universally.
Need and desire: A need is a condition for survival or a chosen standard of decent life. A desire is a preference. Some desires are positional: everyone cannot have highest status. Material abundance cannot mean satisfying every possible desire.
Scarcity and abundance: Scarcity is a constraint relative to demand. Our abundance target concerns essential services, access and cost. It does not mean infinite matter, free energy, equal wealth or unlimited consumption. Material abundance at one date does not itself certify ecological sustainability.
Real, quality-adjusted cost: Cost after accounting for inflation and changes in the service provided. Cheaper but unreliable electricity is not an equal-quality improvement. A lower global average cost does not guarantee access.
Power and energy: Power is a rate: 1 watt = 1 joule/second. Energy accumulates over time. 1 GW = 10⁹ W; 1 TWh = 10¹² watt-hours = 3.6 × 10¹⁵ joules. Battery GWh measures stored energy; a battery factory's GWh/year measures annual production of that storage capacity.
Capacity, output, utilisation: Capacity is rated production capability. Output is actual production. Utilisation compares output with capacity over a defined period. A generator's capacity factor compares average and nameplate power. Factory product mix and uptime also matter.
Token, inference, FLOP: A token is a model-specific unit of encoded information. Inference runs a trained model. A FLOP is one floating-point arithmetic operation; FLOP/s measures arithmetic throughput. None measures a completed scientific result.
Boundary and tipping point: A planetary boundary is a proposed precautionary limit on pressure on an Earth-system process. A tipping point marks a qualitative change driven by feedbacks. Local harm can occur inside a global boundary.
Doom and collapse: Doom here is extinction or irreversible global loss of recovery capacity. Collapse can mean a regional reduction in population, political organisation or economic activity. A fallen empire is not automatically human extinction.

SI: scaling, new science and the reliability gap

Current frontier: GPT-6 Astra was released on 3 September 2026; Opus 5.5 on 22 September. Reported abilities extend across software, scientific workflows, images, CAD and computer interfaces. A broadly cognitive AGI interpretation is reasonable; reliable superiority across expert research is a stronger claim. Robotics is evidence about generality and deployment, not a requirement silently added to move the AGI threshold. Astra, Opus.

Selected independent measurements. Astra: max effort; Opus 5.5: max with default fallbacks. ARC configurations are specified separately.
Evaluation	Astra	Opus 5.5	Interpretation
AA Intelligence Index v5, 27 October; max effort	57	58	Composite index, not a percentage of AGI
AA scientific terminal tasks, same release	63%	59%	Success on this suite, not arbitrary research
AA FrontierCode 2, same release	42%	46%	Unsolved coding work remains
ARC-AGI-3, verified September results	62.7% standard harness, max; 99.9% provider adapter, high	Not compared here	Retain harness and effort with each score

Artificial Analysis uses nine evaluations, 60% private/held-out, in v5. A redesign changes the measuring instrument: compare version-matched results. ARC's results show substantial ability to learn unfamiliar interactive rules, alongside the importance of preserving reasoning state. No cited evaluator certifies a universal AGI threshold. These selected models are not claimed to be the current top two.

A fitted loss law such as L ≈ L∞ + a/N^α + b/D^β describes a tested training regime. It is not a universal law relating compute to autonomous scientific productivity. Under the illustrative compute-optimal relation D ≈ 20N, C ≈ 120N². A 100-fold increase in C implies roughly tenfold increases in N and D. Longer training, different mixtures and post-training change the optimum. Source.

The public-text constraint forecast refers to one stock of information. Data repetition can stretch that stock—experiments found roughly four passes can retain much of the value of fresh tokens—but returns eventually decay. RL with trusted tests can create a different learning loop. Open-ended environments still need feedback that distinguishes a new truth from a plausible error. Data, repetition, RL.

October 2026's OpenAI releases are important evidence, with different verification levels. Some artifacts support machine checking. That does not by itself validate the formal statement's scientific relevance, a general success rate, or claims requiring experiments. Preserve attempted-problem counts and independent verification rather than selecting only successes. 6 October, 26 October.

METR's May 2026 report gives a historical reliability reference, not a measurement of the November frontier: a public software-task horizon of roughly 12 human-equivalent hours at 50% success, versus about 1.5 hours at 80% success, with wide uncertainty and limits beyond 16 hours. These are task lengths for humans, not hours an agent ran, and they do not transfer automatically to experimental science. Source.

Update upward: independent replication, novel prospective experiments, high-reliability autonomy across domains, better fixed-budget learning. Update downward or later: flat held-out project performance across several much larger generations, persistent failures of integrity, expensive feedback, or poor economics of useful output.

The visible compute calculation gives power, memory, token and manufacturing budgets. Annual facility electricity and commissioned capacity remain distinct; the calculation is not a census of every operating project.

See industrial replication and robot stock/flow. General-model robotics evidence also includes an October preprint across twelve benchmarks and eight physical configurations. A September study’s simulated locomotion used paused physics and slow model calls: task success must retain real-time and controller conditions. October study; September study.

See the visible energy portfolio, deep-well, fusion, service/material and medical calculations. For context, the FDA’s typical phase-three range is 300–3,000 patients over 1–4 years; it is descriptive rather than an immutable minimum. FDA.

Planetary-boundary uncertainty and regional ocean exposure

The visible nine-boundary table states the 2026 classification. The 2025 ocean paper reconstructed 2020 global surface aragonite saturation Ω = 2.90 ± 0.06 and a 17.3 ± 5.0% decline, overlapping uncertainty around the original 20% boundary. The current assessment uses a precautionary baseline and later reconstructions. Its breached classification does not imply a sudden chemical jump. Under the paper’s precautionary baseline, roughly 40% of surface waters exceeded a 20% regional saturation decline by 2020. That is exposure, not a percentage of organisms killed. Original study; assessment methodology.

What historical collapses can and cannot tell us

Mayapán, approximately 1400–1450: an archaeological, historical and climate study links drought with factional conflict and the loss of a political centre. It also finds that regional political and economic structures persisted. Climate affected the resources over which groups competed; institutions influenced how the stress propagated. This is a later case than the Classic Maya disruptions around 800–1000. Neither implies that Maya peoples disappeared. Original study.

Roman Britain and the late Western Empire: a 2025 study connected reconstructed drought with military and political disruption. A March 2026 critique challenges its conflict data and interpretation of historical texts; the authors defend their analysis. Climate proxies can establish a stress without proving a particular chain from harvest to famine to rebellion. War, fiscal arrangements, administrative capacity and political choices require their own evidence. “Climate caused Rome to fall” is not a settled result. Study, critique, response.

Do not select only failures. A study of sixteen regional population reconstructions finds evidence of adaptation and recovery after disturbances, with important proxy and selection limitations. Migration, institutional reorganisation and trade can preserve society when a centre fails. Resilience study.

The inference for SI: test pathways, not an empire-collapse frequency. Food or energy stress can reduce revenues; failing services can weaken legitimacy; conflict can damage supply and deepen the original shortage. Modern forecasting, medicine, trade and production improve resilience. Global dependence on shared networks can also synchronise failure. These are competing mechanisms, not a claim that modern society must repeat an ancient trajectory.

Resilience depends on the combination of exposure and response: simultaneous crop failures, spare supply, inventories, diversity of providers, trusted crisis communication and the ability to coordinate. SI can improve those buffers or be deployed in ways that undermine them. A regional collapse observation does not directly estimate P(irreversible global doom).

Climate and doom: stocks, flows and correlated failure

Jancovici's physical framing and the Shift Project's infrastructure analysis are valuable starting points. They do not substitute for current climate assessment. The Shift report's life-cycle CO₂-equivalent scenarios include manufacture and construction, unlike electricity-only CO₂ accounts. Avoid treating different boundaries as contradictory estimates. Jancovici, Shift.

Warming calculation: rounded annual CO₂ emissions 40 Gt × 10 years = 400 Gt. Multiply by the IPCC central response of 0.45°C/1,000 Gt to obtain 0.18°C, with roughly 0.11–0.25°C across its likely response range. Assumes constant net CO₂ emissions; it excludes changes in other forcings and is not a forecast of total warming. Global net-zero CO₂ approximately stabilises CO₂-induced temperature rise; sea level continues responding for centuries to millennia. Carbon budget, IPCC response, long-term impacts.

Removal calculation: one published liquid direct-air-capture design uses 5.25 GJ of natural gas and 366 kWh of electricity per tonne captured. At 40 Gt/year that is 210 EJ/year gas input plus 14,640 TWh/year electricity: approximately 6.7 TW gas-input power and 1.7 TW electrical power. This is an extreme scale exercise using a particular process, not a universal lower bound or a recommendation to offset everything. Net removal also needs low life-cycle emissions and durable storage. Process study.

The IEA's widespread-adoption scenario gives existing AI applications potential avoided emissions of 1.4 GtCO₂/year in 2035. That is order 10⁹ tonnes against today's order 4 × 10¹⁰, a few percent comparison across different dates—not measured net savings and not an upper bound on SI. Deployment and rebound determine realised benefit. Source.

Pathway	Evidence and countervailing mechanisms	Useful update
War	~12,000 nuclear weapons already exist. Better verification competes with escalation, misuse and shortened decision times. Technology amplifies existing political conflict; inevitability is not established.	Actual crisis stability, arms control, decision safeguards and harmful deployment
Disease	WHO estimates ~15 million COVID-associated excess deaths in 2020–2021. AI can help surveillance and countermeasures; misuse and unequal delivery also matter.	Verified time from outbreak to effective, widely delivered protection
Climate	Accumulated emissions and vulnerable food/water systems persist despite better forecasts. Catastrophic impact is not synonymous with extinction.	Net emissions, resilience and deployment rates, not plans alone
Resources	IEA's 25% copper pipeline gap for 2035 is a scenario about supply rates, not exhaustion of copper atoms. Substitution, recycling and new projects can change it.	Refining concentration, actual new supply, substitution and shocks
Loss of control	Strong capabilities, harmful behaviour and enabling deployment are distinct conditions. The February 2026 safety assessment reports early relevant signs, but assessed then-current systems as below levels enabling loss of control. That dated finding does not assess all November models.	Independent evaluations of autonomy, oversight and containment as capability rises

Nuclear inventory · Pandemic mortality · Minerals · AI safety assessment.

A nuclear-war food-system study models billions of deaths in severe scenarios. That is evidence about a conditional mechanism, not a probability of war or certain extinction. Food, war, climate and disease can be stages of one cascade. For overlapping pathways Dᵢ, max P(Dᵢ | S) ≤ P(∪Dᵢ | S) ≤ ΣP(Dᵢ | S); neither independence nor a simple sum is justified. Study.

Sources and verification notes

Vaclav Smil, How the World Really Works (2022). Supplied ebook PDF: chapter 3, “Understanding Our Material World,” PDF page 84 for the 2019 material quantities; chapter 7, “Understanding the Future,” PDF page 226 for the quotation about enormous scales. PDF pages refer to the supplied conversion, not printed-edition pagination. The introduction and relevant passages of chapters 3, 6 and 7 inform the physical framing; historical numbers are labelled by year. The copyrighted book itself is not redistributed.

Primary sources and scientific syntheses are listed below. Company statements establish what was reported or planned, not independent verification. Some Nature articles and IPCC pages were accessible only through their indexed official text during research; no claim of full-text review is made for those sources. The underlying research memos retain verification limits. Expanded physical sections cite their primary sources inline; the eight specialist memos and independent arithmetic review preserve the assumptions and access limits for those calculations. The full 2026 Planetary Health Check PDF exceeded the reader size limit; its institutional announcement and boundary-specific pages were inspected.

Artificial Analysis · 27 October 2026. Intelligence Index v5. Nine evaluations; 60% private/held-out. Scores depend on index version, configuration and effort.
OpenAI · 3 September 2026. GPT-6 Astra. Provider results across science, CAD, software and computer use.
Anthropic · 22 September 2026. Claude Opus 5.5. Provider results, not a universal AGI certification.
ARC Prize · September 2026; checked 1 November. Verified OpenAI GPT-6 results. Standard and provider-adapter harnesses give different scores.
SenseNova embodied study · 15 October 2026, revised 19 October. General intelligence in embodied tasks. Twelve benchmarks and eight physical robot configurations; preprint with deployment limitations.
Galbot embodied-policy evaluation · 29 September 2026. GPT-6 Astra across six embodied domains. Simulation and physical experiments need different interpretations; locomotion test pauses physics.
International Federation of Robotics · 24 September 2026. Five Million Robots now Operate in Factories Globally. Approximately five million operating industrial robots in 2025; over 600,000 annual installations.
International Federation of Robotics · 30 September 2026. Global sales of professional service robots. About 7,000 full-size humanoids sold in 2025 under its >140 cm definition; often teleoperated.
BMW · 25 June 2026. Physical AI in production at Spartanburg. Figure 02 supported one body-shop task on 30,000 vehicles over ten months; it did not build whole cars.
Schaeffler · 2 September 2026. Formed strain-wave gearboxes for humanoid robots. Supplier-reported manufacturing improvements; mass rollout planned for 2027.
Tesla and Panasonic · 31 July 2014. Gigafactory agreement. Original 35 GWh/year cell-production target; partner roles and co-located manufacturing.
Tesla · 24 January 2023. Continuing Our Investment in Nevada. Company-reported production above 37 GWh/year; further expansion was a plan.
IEA · Global EV Outlook 2026. Electric vehicle batteries. Over 4 TWh/year global nameplate manufacturing capacity at end-2025; capacity is not output.
Panasonic · corporate history, 1994 entry. Lithium-ion battery development. Industrial battery expertise predated the Nevada partnership.
Potsdam Institute for Climate Impact Research · 21 September 2026. Planetary Health Check 2026. Seven boundaries classified as breached; observation dates vary.
Planetary Health Check · checked 1 November 2026. The planetary boundaries framework. Precautionary risk zones and interacting processes, not probabilities of extinction.
NOAA PMEL · checked 1 November 2026. A primer on pH. Historical surface-ocean decline about 0.11 pH units.
Planetary Health Check · checked 1 November 2026. Ocean acidification. Global surface aragonite saturation boundary 2.86, with precautionary preindustrial baseline 3.57.
Findlay et al. · 9 June 2025. Ocean acidification boundary reassessment. Global mean, regional exposure, uncertainty and baseline choice must be distinguished.
NOAA · 11 June 2025. Ocean acidification is more pervasive than previously thought. Habitat suitability estimates are not observed mortality.
IPCC · AR6 WGI, 2021. Chapter 5: Global Carbon and other Biogeochemical Cycles and Feedbacks. Carbon removal can reverse surface acidification; deeper recovery is slower.
UNEP Ozone Secretariat · checked 1 November 2026. Facts and figures on ozone protection. 99% phaseout of controlled ozone-depleting substances; recovery with long regional lags.
Kennett et al. · 19 July 2022. Drought, civil conflict, and political collapse at Mayapán. Interacting pressures on a political centre, with regional continuity.
Norman et al. · 16 April 2025. Roman drought and conflict analysis. Climate reconstruction and historical interpretation; causal argument contested.
Foxhall Forbes et al. · 31 March 2026. Critique of the Roman drought/conflict analysis. Challenges conflict data and documentary interpretation.
Norman et al. · 2026. Reply to the Roman drought/conflict critique. The dispute is not presented as settled.
Riris et al. · 1 May 2024. Long-term resilience from past population dynamics. Sixteen regional reconstructions; proxy and selection limits remain.
Hoffmann et al. · 2022. Training Compute-Optimal Large Language Models. Empirical training scaling; not a guarantee of general intelligence.
Villalobos et al. · 2024 revision. Will we run out of data?. Conditional 2026–2032 public human-text constraint; not all forms of learning.
Muennighoff et al. · 2023; revised 2025. Scaling Data-Constrained Language Models. Repetition helps initially; diminishing returns remain.
DeepSeek-AI · 2025; revised 2026. DeepSeek-R1. Reinforcement learning with verifiable rewards.
OpenAI · 6 October 2026. Sharing AI progress in mathematics. Provider release with inspectable formal artifacts; selected results.
OpenAI · 26 October 2026. Working with academic communities to build scientific understanding. Provider research claims; independent acceptance and artifact availability vary.
METR · 19 May 2026. Frontier Risk Report, February–March 2026. Independent software task evaluation, with large uncertainty and suite limits.
IEA · 16 April 2026. Key Questions on Energy and AI. 485 TWh in 2025; 950 TWh in 2030 is a projection, all data centres.
OpenAI · 18 January 2026. A business that scales with the value of intelligence. Company-reported available capacity, not metered annual facility electricity.
OpenAI · 29 April 2026. Building the compute infrastructure for the Intelligence Age. More than 10 GW secured; distinguish pipeline from operation.
Crusoe · 27 March 2026. Abilene campus construction and energisation. Dated evidence of two energised 100 MW buildings; future campus targets.
NVIDIA · specifications checked 1 November 2026. H100 Tensor Core GPU. Peak arithmetic, not actual inference energy; distinguish sparse from dense FLOPs.
United Nations · 11 July 2024. World Population Prospects 2024. 8.2 billion in 2024; this calculation uses 8.5 billion as a rounded modelling denominator.
IEA PVPS · September 2026. Trends in Photovoltaic Applications 2026. Approximately 690 GW installed during 2025; 2.96 TW cumulative nameplate capacity.
IEA · October 2023. Electricity Grids and Secure Energy Transitions. 80 million km added or refurbished by 2040 to meet national goals; includes distribution.
IAEA · 2025. Nuclear Technology Review 2025. 377 GW(e) operational at end-2024; 2,617.5 TWh generation during 2024.
US Department of Energy · checked 1 November 2026. Geothermal Electricity Generation. US capacity just over 4 GW; 90 GW by 2050 is conditional potential.
Lawrence Livermore National Laboratory · updated through June 2026. Achieving fusion ignition. 8.6 MJ fusion yield in April 2025; target gain excludes facility electricity.
ASML · Annual Report 2025, published 2026. Annual report and financial statements. 48 EUV systems sold in 2025.
TSMC · Arizona project overview, checked 1 November 2026. TSMC Arizona. First volume production in Q4 2024; later fabs remain construction and scheduled production.
Abramson et al. · 8 May 2024. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Structure prediction is not clinical efficacy.
Xu et al. · 3 June 2025. AI-generated rentosertib: randomised phase-2a study. 71 patients, 12 weeks; safety/tolerability primary endpoint; efficacy exploratory.
FDA · checked 1 November 2026. Step 3: Clinical Research. Typical trial ranges; not immutable minimum durations.
FDA · Casgevy label, July 2026. Casgevy prescribing information. Individual cell collection, manufacture and conditioning remain delivery constraints.
WHO · 4 February 2026. Malaria vaccines: RTS,S and R21. An existing prevention technology with a multi-dose delivery programme.
Jean-Marc Jancovici · 2011, updated 2015. What is energy, actually?. Physical framing; dated numerical examples are not used as current measurements.
The Shift Project · 1 October 2025. AI, data and computing: infrastructure in a decarbonised world. Life-cycle boundaries include construction/manufacture; not the same as electricity-only emissions.
Friedlingstein et al. · 13 May 2026. Global Carbon Budget 2025. 2024 total anthropogenic CO₂: 42.4 ± 3.2 Gt; preliminary 2025: 42.2 Gt.
IPCC · 2021. AR6 Working Group I, Summary for Policymakers. Central transient response: 0.45°C per 1,000 GtCO₂; likely range 0.27–0.63°C.
IPCC · 2023. AR6 Synthesis Report, Summary for Policymakers. Sea-level rise persists on much longer timescales than temperature stabilisation.
Keith et al. · 2018. A Process for Capturing CO₂ from the Atmosphere. One process design: 5.25 GJ gas plus 366 kWh electricity per tonne. Not a universal lower bound.
IEA · April 2025. AI and climate change. Potential 1.4 GtCO₂/year avoided in 2035 in a widespread-adoption scenario; not measured savings.
IEA · April 2025. AI for energy optimisation and innovation. Existing-AI deployment scenarios; not SI performance forecasts.
IMO · 2020. Fourth IMO Greenhouse Gas Study. All-shipping 2018 CO₂ emissions of 1,056 Mt; used only as a dated scale example.
SIPRI · 8 June 2026. SIPRI Yearbook 2026 nuclear inventory. Estimated 12,187 total nuclear warheads in January 2026.
Xia et al. · 15 August 2022. Global food insecurity and famine from nuclear war. Conditional scenario modelling, not a forecast or a numerical extinction probability.
WHO · 5 May 2022. COVID-19 excess mortality, 2020–2021. 14.9 million estimated excess deaths, direct and indirect.
IEA · 16 July 2026. Global Critical Minerals Outlook 2026. Projected 2035 copper pipeline gap of 25%; not inevitable geological exhaustion.
International AI Safety Report · 3 February 2026. International AI Safety Report 2026. Capabilities, harmful propensity and enabling deployment are distinct conditions for loss of control.