AI efficiency
Compute required for a particular unit of useful intelligence falls as models, algorithms, chips and allocation improve.
Ends Up Here Podcast · 01
Idea / In development
What Happens
to All That
Compute?
AI may become dramatically more efficient at the same time as computation spreads through agents, robots, infrastructure, science and the physical world.
Key propositionCompute doesn't disappear.
Its purpose migrates.
Its purpose migrates.
The starting assumption
Better AI → more use → more inference → more compute.
That chain is supporting enormous investment. But it extrapolates from today's compute-intensive systems and today's economy. AI may improve its own algorithms, architectures, inference efficiency, hardware, compression, distillation and allocation of computation.
If useful intelligence becomes 10×, 100×, 1,000× or even 10,000× more efficient, the compute required for today's tasks could collapse.
This is a scenario argument, not a forecast. The charts below use indexed values to expose interacting directions. They do not claim to predict future quantities.
Chart 01 / Competing trajectories
Efficiency can fall while total demand rises.
Choose an efficiency scenario to see how the indexed curves interact. 2026 is the starting reference point for every series.
Efficiency scenario10× more efficient
Conventional AI compute projectionCompute per useful taskTraditional human/economic computeAgentic + civilisational compute
Illustrative indexed scenario: the shapes show a question worth tracking, not measured data or a prediction. “Civilisational” includes agents, world-management systems, science, health, safety, robotics and major physical projects.
Three interacting curves
Traditional demand
Some conventional work and organisational compute changes if fewer people spend eight hours producing digital artefacts because employment requires it.
Civilisational demand
Cheap intelligence unlocks agents, robots, infrastructure, health, science, governance, world models and projects that were previously impractical.
The human variable
People may spend less time at desks.
A great deal of today's computing exists because hundreds of millions of humans spend their working lives at computers producing documents, emails, spreadsheets, presentations, searches, software, advertising, administration and transactions.
If conventional employment contracts, AI may not reproduce every one of those activities. Some organisational activity may simply cease to be necessary. Someone who currently sits at work thinking about camping may actually be camping.
Chart 02 / Where compute goes
Compute migrates into the world.
The composition below is illustrative. It shows changing purpose, not a forecast of market share.
Human work / screensPersonal AIAgentsRobotics / infrastructureScience / healthGovernance / safety / world modelsEngineering / resources / space
Indexed illustrative composition. Categories overlap in the real world; they are separated here as prompts for tracking.
The investment paradox
Investors value the creation of a thing that may destroy its scarcity.
Today's logic can be simplified as: scarce intelligence → expensive AI services → high margins → enormous valuations.
Sufficiently successful AI could also produce: abundant intelligence → negligible marginal cost → intense competition and commoditisation → intelligence approaching a utility or free good.
That does not make compute, energy, models, robotics or infrastructure worthless. It changes what ownership buys—and what remains scarce.
Migration of scarcity
Where does scarcity go?
Abundance does not necessarily eliminate wealth. It changes what constitutes scarce wealth. If manufactured goods become extremely cheap, owning ten Rolls-Royces may not mean very much if everybody can afford one.
Desirable land and locationsUnique physical resourcesHuman attention and relationshipsAccess and experiencesEnergy and resource allocationPlanetary-scale engineeringSpace travel and settlement
A useful physical limit
Cheap intelligence does not make Mars cheap.
Even if intelligence becomes almost free, transporting people and physical resources to Mars remains an enormous resource undertaking. Compute can make the planning, simulation, robotics and coordination better; it does not erase mass, energy, distance or the need to build physical systems.
Living trajectory
Trajectory Updates
This is designed to grow week by week. Add dated observations here as evidence appears; revise the curves when reality demands it.
Next updateWeek of —
The tracking process will cover model and inference efficiency, compute costs, chips, data centres, electricity, agents, robotics, employment-related compute, world management, science, health, governance, safety, major physical applications, space, scarcity, economic value and company reactions.
Talking-point prompts
Who is consuming the compute?
For what purpose?
Where is the economic value moving?
What becomes scarce when intelligence becomes abundant?
Which parts of today's digital work simply disappear?
What physical constraints remain stubbornly real?