What changes when AI's real constraint is compute, energy, and data centers — the physical substrate?

The shift

Cognition — drafting, synthesis, code, analysis — went abundant and near-free. But the substrate that produces it did not: chips, electricity, cooling, water, land, grid interconnection, and fab capacity are physical, capital-heavy, and slow to build. The scarcity didn't disappear when intelligence got cheap; it moved down a layer, from the model to the machines and megawatts that run it.

The axioms

1. Software scales at ~zero marginal cost — one more copy, one more user, is effectively free.
2. Compute is a fungible commodity: you rent what you need, when you need it, and supply is elastic.
3. Energy is ample and not the binding input on a digital business.
4. Intelligence — model quality, research talent, algorithms — is the bottleneck; whoever has the best model wins.
5. Capability improves on a software cadence: ship, iterate, ship again in weeks.
6. The moat is the model and the data pipeline, both of which are information goods.
7. A frontier AI company is a software company and can be valued and run like one — light on physical assets, high on margin.

Each rests on an assumed abundance: that the physical substrate underneath the software is cheap, elastic, and fast to acquire.

Invalid axioms

1. The marginal cost of AI is ~zero. True for shipping a finished app; false for serving a model. Every inference burns electricity and amortizes silicon that wears out and must be replaced on a ~3–5 year cycle. Habit-trap: pricing, unit economics, and "just scale it" planning still assume software-style near-zero marginal cost, when the marginal cost is now a real, metered draw of power and depreciating hardware.
2. Compute is a fungible commodity you rent on demand. Frontier training and large-scale inference now run into hard supply limits — leading-edge accelerators, advanced packaging (CoWoS), and high-bandwidth memory are allocation-constrained, and the data-center capacity to house them is booked years out. Habit-trap: strategy still treats compute as a spot input you buy when needed, when securing it now looks more like a multi-year capital procurement and site-development problem.
3. Intelligence is the bottleneck; the best model wins. Model quality matters, but the binding constraint on who can train and serve at frontier scale has shifted to who can secure power, chips, and buildings. Habit-trap: companies still compete and recruit as if algorithmic talent alone were decisive, while the actual gate is access to megawatts and accelerators.
4. Capability improves on a software cadence. The software layer still iterates in weeks; the substrate does not. A new fab is a multi-year, tens-of-billions build; a gigawatt of new power and grid interconnection is a multi-year queue. Habit-trap: roadmaps and market expectations assume the whole stack moves at software speed, when the rate-limiting layer now moves at infrastructure speed.
5. A frontier AI company is a light-asset software company. At the frontier it increasingly resembles a capital-intensive industrial operator — closer to a utility or a chipmaker in cost structure than to a SaaS business, with balance sheets dominated by data centers, power contracts, and hardware. Habit-trap: valuing and financing these companies on software multiples and margins misreads where the money actually goes.

Unchanged axioms

1. Physical build-out is slow, capital-heavy, and needs the trades. Data centers, substations, transmission lines, and fabs are poured, wired, and commissioned by electricians, pipefitters, steelworkers, and heavy construction — headcount and lead times measured in years. No abundance of cognition strings a transmission line or pours a foundation. This is the single largest reason AI's physical layer lags its software layer.
2. Energy is a hard limit, not a line item. A data center's draw is bounded by what a region can actually generate and deliver. Power availability, grid interconnection timelines, and local generation capacity gate where and how fast compute can be built — a physical ceiling that money alone doesn't lift on any short timescale.
3. Leading-edge chips are geopolitically concentrated and non-substitutable. Frontier accelerators depend on a fragile chain: advanced fabrication concentrated in Taiwan and a handful of others, EUV lithography from effectively one vendor (ASML), and specialized packaging and memory with few suppliers. This concentration is a physical and political fact that abundant intelligence does not dissolve; it is a genuine chokepoint and single point of failure.
4. Someone accountable must own the physical assets and their risks. Power-purchase agreements, water rights, grid-impact commitments, environmental permits, and community relationships require a liable entity that stakes capital over decades. A model can plan a data center; it cannot sign a 15-year PPA or be answerable when a build overruns or a community objects.
5. Multi-year capital allocation is a judgment bet, not a data problem. Deciding to commit tens of billions to fabs, power, and buildings against uncertain future demand is a wager on the trajectory of the technology itself. Better forecasts sharpen the inputs; they don't remove the bet or the scarce willingness to stake capital on it.

New axioms

1. Compute becomes the strategic-scarce resource, and whoever owns it captures the surplus. When intelligence is cheap but the substrate to run it is scarce, economic power concentrates in whoever controls chips, power, and data-center capacity — chipmakers, hyperscalers, and increasingly the states and utilities that gate them. The open problem: a small number of substrate-owners may capture most of the value AI creates, with everyone else renting access to a controlled resource.
2. AI's energy demand collides with climate and grid commitments. Compute build-out is adding load fast enough to strain regional grids, revive fossil and delay retirements, and put upward pressure on prices for other ratepayers. The problem to solve: reconciling exponential compute demand with decarbonization targets and the interests of communities who share the grid but not the upside.
3. Grid strain and siting become the real bottleneck on AI progress. Interconnection queues, transformer and switchgear lead times, water for cooling, and local permitting now gate frontier scale more than model research does. The open problem is coordinating power, land, water, and grid capacity at a pace the technology layer expects but the physical layer has never delivered.
4. Chip chokepoints turn AI capability into a geopolitical dependency. With leading-edge fabrication and lithography concentrated in a few jurisdictions, AI capacity is now hostage to export controls, trade policy, and the security of specific regions. The problem: capability is exposed to disruptions — political or physical — entirely outside the control of the companies depending on it.
5. The build-out creates real, durable demand for the trades and energy sector. The constraint on scaling AI is now electricians, HVAC and cooling specialists, high-voltage and construction labor, and new generation — a physical-economy demand pulse that didn't exist at this scale before. The open problem: this labor and generation capacity is itself scarce and slow to grow, and it competes with electrification and reshoring for the same workers and megawatts.

Where it breaks

"The marginal cost of AI is ~zero / it scales like software" (invalid) collides with "compute is the new strategic-scarce resource and energy is a hard limit" (new): the industry still prices, promises, and plans capability on software economics while the binding cost has become metered power and allocation-constrained silicon — so demand is being written checks the physical layer can't cash on the expected timeline.

Separately, "capability improves on a software cadence" (invalid) collides with "grid strain, siting, and chip chokepoints are the real bottleneck" (new): expectations and capital are set to software-speed iteration, but the rate-limiting layer moves at fab-and-transmission speed — a mismatch that shows up as data centers waiting years for power and accelerators, and roadmaps that assume a cadence the substrate has never sustained.

Calibration note (mid-2026): the STILL HOLDS and NEW buckets here hinge on fast-moving numbers — data-center power draw, interconnection queue lengths, accelerator and packaging supply, and fab timelines. Directionally these constraints are binding today, but specific figures and the pace of relief (new generation coming online, added packaging/HBM capacity, efficiency gains per token) are moving fast and should be checked against current data before any figure-dependent decision. The structural claim — that scarcity moved from cognition to substrate — is more stable than any single number under it.

ai-axioms.vercel.app/axioms/society/compute-energy-substrate-ai-axioms