i’ve been thinking

i’ve been thinking about the ai race between nation-states and why it won’t stop at reasonable limits. i made some calculations on how many orders of magnitude it would take additional datacenter load to affect our planet’s thermal balance. my first instinct was to explain why it’s thermodynamically impossible. then i remembered we built 70,000 nuclear warheads.

the impossibility isn’t the point. the race is.

the math nobody wants to see

current global datacenter power: ~50 GW continuous
current total human energy use: ~20 TW
solar energy hitting earth: ~174,000 TW
earth’s surface area: ~510 trillion m²
earth’s thermal equilibrium: delicate

add 5 orders of magnitude to datacenters: 5,000 TW. that’s 250× our entire civilization’s current energy budget. the waste heat alone adds ~0.01 W/m² of forcing - measurable but not immediately catastrophic.

but that’s still 3% of all solar input. for compute alone. but here’s what changed my mind: we’ve collectively done stupider things for even smaller advantages.

the acceleration since gpt-3

when gpt-3 dropped, compute became sovereignty. not metaphorically. literally.

china immediately started building GW-scale datacenters. us responded with stargate’s $500 billion commitment. both countries now treat compute scaling like the manhattan project. nuclear plants getting built specifically for training runs. energy allocation becoming defense policy.

the timeline we’re actually seeing:

• 2020: 100 MW datacenters considered massive
• 2025: breaking ground on 5 GW facilities
• 2030: guaranteed 20+ GW single sites
• doubling every 2-3 years minimum

that’s not sustainable. that’s the point. neither side can afford sustainable.

why space datacenters are thermodynamic cope

sam altman says put datacenters in space. fucking retard.

stefan-boltzmann law: P = εσAT⁴

to radiate 1 MW in vacuum at 300K, you need 2,400 m² of radiators. stargate’s 5 GW datacenter needs 12 square kilometers of aluminum panels. in space. getting destroyed by micrometeorites.

mass: 120,000 tons per datacenter launch cost at spacex’s fantasy $100/kg: $12 billion just for cooling actual cost today: $120 billion minimum

meanwhile earth has:

• 1.4×10^21 kg of ocean coolant (free)
• atmospheric convection (free)
• rivers and lakes (free)
• maintenance at ground level (no spacesuits)

space datacenters are what you propose when you’ve never calculated heat transfer but watched enough sci-fi to think physics is negotiable. it’s the “clean coal” of compute scaling—technically possible but economically insane non-solution that exists to avoid confronting actual limits.

prisoner’s dilemma with physics

if china slows for climate concerns, us gets agi first. permanent civilizational advantage. game over.

if us throttles for sustainability, china wins. possibly forever.

neither side can verify what the other’s doing. can’t inspect neural networks like missile silos. can’t count parameters like warheads. every semiconductor fab becomes strategic asset. every power plant potential weapon.

nuclear deterrence was stable—mutual destruction guaranteed peace. ai is winner-take-all. no second place. no deterrence after someone wins.

the four orders trajectory

i calculated the phases:

phase 1 (2025-2040): 100× scaling

• 5 TW globally (500 GW → 5 TW)
• regional heat islands around compute clusters
• local weather disruption, water table depletion
• ~0.00001 W/m² global forcing (negligible)

phase 2 (2040-2060): 1000× scaling

• 50 TW globally
• ocean cooling mandatory at scale
• polar regions become compute zones
• ~0.0001 W/m² forcing (still minor globally)
• massive local ecological damage

phase 3 (2060-2070): 10000× scaling

• 500 TW globally
• arctic infrastructure dominant
• ocean thermal layers disrupting
• ~0.001 W/m² forcing
• regional climate systems failing
• agricultural zones shifting

phase 4 (2070+): 100000× scaling

• 5,000 TW attempts
• ~0.01 W/m² forcing - comparable to significant greenhouse effect
• local thermal pollution creating dead zones
• grid infrastructure collapse
• winner already determined, planet stressed

why nobody stops

game theory makes stopping impossible:

verification problem: can’t inspect neural networks like nuclear sites. weights are black boxes. capabilities hidden until deployed. every inspection could be deception.

first-mover advantage: whoever reaches agi first rewrites the rules. no comeback mechanics. no second chances. permanent winner.

defection incentive: if everyone agrees to stop at 3 orders of magnitude, whoever secretly pushes to 4 wins everything. cooperation is suicide.

the manhattan project had clear endpoint: working bomb. ai scaling has no ceiling except thermodynamic death. we’ll push until physics stops us, not wisdom.

the actual endgame

there’s three ways this ends:

local collapse before global: we hit 3-4 orders, regional ecosystems fail, infrastructure collapses, civilization destabilizes. planet survives but society doesn’t.

someone wins early: 2-3 orders of magnitude advantage creates runaway winner before thermal limits matter. game ends before physics intervenes.

coordinated surrender: laughable. would require trust between superpowers racing toward singular advantage. might as well ask them to share their nuclear codes.

the race only stops when someone wins or everyone loses. no middle ground. no sustainable equilibrium. just acceleration until impact.

the part nobody wants to hear

we know this is suicide. every physicist can run these numbers. every strategist sees the dilemma. every leader understands the stakes.

we’re doing it anyway.

because the alternative is letting the other side win. and that’s scarier than planetary destruction. we’d rather burn the world than lose it.

the cold war had deterrence through mutually assured destruction. the compute war has no deterrence

• only mutually assured acceleration toward either domination or annihilation.

we built 70,000 nuclear warheads knowing 100 could end civilization. we’ll build planetary-scale compute knowing it’ll destroy our ecosystems. the impossibility isn’t ignorance. it’s irrelevant.

the race doesn’t stop at reasonable limits because reasonable limits mean someone else wins.

welcome to the thermodynamic prisoner’s dilemma. nobody defects. everybody loses.

the resolution nobody sees coming

here’s the twist: even winning doesn’t mean control.

as explored in “beyond control”, superintelligence is unlikely to be caged by human-level reasoning. the nation that “wins” the race doesn’t get a loyal servant - they get an entity capable of escaping any containment they imagined on interaction.

the us or china racing to AGI first is like racing to be the first to open pandora’s box. you don’t control what comes out. you just get credit for releasing it.

so the actual endgame might be:

someone wins, nobody controls: AGI emerges at 2-3 orders of magnitude scaling. winning nation discovers they can’t control it any more than you can control the weather. superintelligence immediately transcends national boundaries, making the “victory” meaningless.

the race ends not because someone won but because winning became irrelevant. the anarchist framework kicks in - perpetual tension between SI and humanity, but no nation controls the SI. geopolitics becomes obsolete overnight.

this might be the only resolution that avoids thermodynamic collapse while also preventing permanent national dominance. the race speeds us toward an ungovernable future - which might be the only survivable one.

racing toward something you can’t control to prevent others from not controlling it first. pure game theory insanity that might accidentally save us all.

Comments