What actually limits AI?

AI is manufactured, like steel. Three ingredients — chips, electricity, money — go into a building; out comes a model. Every debate about AI's future is secretly about whether we can keep getting more of those three.

Quick vocabulary: FLOPFLOP: One floating-point operation. Training compute is measured in total FLOP — GPT-4 took about 2.1×10²⁵. measures training math. TrainingTraining: The one-time, months-long run that teaches a model. Building the engine. builds the model; InferenceInference: Running the trained model to answer queries — now about two-thirds of all AI compute. answers your questions. A gigawatt (GW)gigawatt (GW): A billion watts — one large nuclear reactor, or ~750,000 homes. is how AI infrastructure is planned; a tokentoken: A chunk of text, ~¾ of a word. Intelligence is sold by the million tokens. is how intelligence is sold.

The scale: GPT-4 trained on 28 MW — a small town. Grok 4, two years later: ~250 MW. The runs planned for 2028–2030 need 2–5 GW sustained for most of a year — a steel industry's worth of demand, arriving all at once, in a dozen places.

Start with inventory: both supply chains reduced to their physical inputs, each with a verdict. The famous "rare" metals mostly earn a shrug — grams, at the periphery. The red flags are boring: turbine blade castings, transformer steel, uranium enrichment.

Blue is the chip chain, orange the energy chain. They converge in the data center; out comes the only product — tokens, split between training and inference.

The widths tell the story: an H100 holds a few hundred dollars of metal and sells for $28,000. The scarcity lives in the middle — fab, memory, packaging, turbine, and grid capacity. The bottleneck is never the stuff; it's the machines that transform the stuff.

The rare-metals fear checks out on paper — China refines ~99% of gallium and ~91% of rare earths, and keeps toggling export bans. But those materials touch the racks around the GPU (power supplies, cooling fans), not the GPU. If China squeezes, prices jump, engineers redesign, and no training run stops.

The real chip chokepoints: TSMC's 3nm lines, HBMHBM: High-bandwidth memory — DRAM stacked onto the AI chip. ~40% of an accelerator's cost., and CoWoSCoWoS: TSMC's packaging step that fuses GPU and memory into one unit — the narrowest funnel in the chip chain. packaging — all sold out through 2026. Yet the funnel keeps widening: from 1.4M H100-equivalents shipped in 2023 to an installed base of 15M+ by 2026, compounding 3.3×/yr. Scarcity raises prices; it hasn't bent the curve.

The honest chip risk is a map, not a mineral: >90% of advanced logic comes from one island. Bloomberg prices a Taiwan war at $10.6T in year one; TSMC Arizona covers ~30% of leading-edge by 2030, a node behind. Every other risk is a delay. That one is a cliff.

Here's the actual wall. US data centers used 4.4% of electricity in 2023; every forecaster sees that doubling or tripling by 2030. America has plenty of gas, uranium, and sun. What it can't buy before 2029 is the equipment that turns them into electrons:

GPUs arrive in months. A gas turbine slot: GE Vernova's backlog is 116 GW against ~20 GW/yr of output — sold out into 2029–30, limited by single-crystal blade casting. A big transformer: 2.5–4 years, on a 30% steel deficit. A grid connection: 2,290 GW in queue, median wait over 4 years.

So every gigawatt campus plays the same move: become the grid. xAI bolted 35 turbines to a Memphis factory; Microsoft is restarting Three Mile Island; Meta and Amazon bought whole nuclear plants' output for 20 years. Meanwhile the cost lands on everyone else — the PJM capacity auction rose 11× in two years, mostly attributed to data centers.

Two growth rates carry this whole essay: compute per run grows 4–5×/yr; power only 2.2×. The gap is chip efficiency — each generation does ~2× the math per watt — absorbing half of every scale-up. Gigawatts must supply the rest: Epoch projects 1–2 GW per run by 2028, 4–16 GW by 2030, with a 2030-scale run consuming ~Ireland's annual electricity.

Money is the loose constraint: $78M (GPT-4) → ~$490M (Grok 4) → $1B+ by 2027. Energy is just 2–6% of a run's cost — the problem isn't paying the bill, it's that the electricity physically isn't there.

The 2026 buildout bill: ~$725B of big-tech capex, up from $251B in 2024 — 34% of revenue, double the telecom bubble's peak intensity.

Three different money machines. Hyperscalers pay from cash flow — they can do this forever. Labs pay with equity: OpenAI's $122B round at $852B (against $115B of projected losses); Anthropic's $65B at $965B. Middlemen pay with debt: Oracle $108B+, CoreWeave $35B — some literally collateralized by GPUs.

And yes — the IPO instinct is correct. VC funds write $100M checks; a gigawatt campus costs $35–50B. So 2025–26 invented two exits: AI project finance (Meta's Hyperion: a $27B private-credit deal, pension money owning the campus, lease payments servicing the bonds) and going public — SpaceX, with xAI inside, raised ~$75B in June 2026, the largest IPO ever; Anthropic reportedly filed; OpenAI's restructure cleared its ~2027 path. Public markets are the only pool deep enough for this debt.

The catch: all of it stands on maybe $50–150B of actual AI revenue — a gap Allianz says is already wider than the 2001 telecom bubble's. Bulls point at lab revenue compounding (OpenAI ~$25B, Anthropic ~$47B run-rates vs ~$5B combined in 2024). Bears watch Oracle's credit spreads, already at 2008 levels. If revenue stalls, the leveraged middlemen crack first.

The price of intelligence is two curves. Any fixed capability collapses ~10×/yr — GPT-4-level answers went from $37.50 to ~8¢ per million tokens, a 450× drop. The frontier has cost roughly the same for three years: labs spend every efficiency gain on smarter, not cheaper.

Three engines drive the collapse: algorithms (compute for fixed capability halves every ~8 months), hardware (2–3× per watt per generation), competition (open weights at ~85% of frontier for a tenth the price). Energy per answer falls too — Google's median Gemini prompt: 0.24 Wh, down 33× in a year.

The hypothesis: capability S-curves when energy runs out, the industry pivots to intelligence density — more smarts per joule — then a later capacity wave unlocks the next jump. Directionally right, and the numbers are checkable.

Two clocks. The efficiency clock is fast: at an 8-month halving, algorithms alone deliver a 10×-scale-up's worth of capability every ~2.2 years, no new power needed. The capacity clock is slow — 3–5 year equipment lead times — but 2024–25's orders (116 GW of turbines, dozens of campuses, SMR fleets) land in 2028–2031. A 2027-ish energy wall means growth at ~4×/yr instead of ~20×: a flattening shoulder, not a plateau, until the capacity wave arrives. The five-year story is really a three-to-five-year story.

One asymmetry: an energy cap hurts training far more than inference. Serving efficiency improves ~33×/yr — so even energy-starved, the intelligence we already have keeps getting radically cheaper to distribute. The S-curve flattens the frontier, not the deployment.

Space data centers: real engineering, premature economics. Starcloud flew an H100 in orbit (its radiator couldn't run it at full power); Google launches TPU satellites in 2027. But chips in vacuum shed heat by radiation only — a 40 MW cluster needs ~120,000 m² of radiators — and a 1 GW orbital station pencils at ~$42B, 3× terrestrial, unrepairable. Works if Starship hits ~$100–250/kg; Google's own feasibility year is ~2035. Watch launch prices, ignore demos.

Nearer-term: local politics ($130B of projects blocked or delayed in Q1 2026 alone; moratoria floated in 14 states), people (~439,000 missing construction workers; electricians are Microsoft's #1 obstacle), water (solvable — closed-loop cooling cuts use >100×), and export politics (Gulf gigawatts traded for US chip access; rules rewritten twice in 18 months).

Winning regardless: the toll collectors — Nvidia (~71% margins on $216B), TSMC (+40%/yr), and the energy trio GE Vernova ($176B backlog), Vertiv, Eaton. They're paid whether or not AGI arrives.

Winning if the curve holds: funded labs (Anthropic growing into its valuation; OpenAI's pre-paid war chest), hyperscalers who own scarce gigawatts, and power producers selling the one input nobody can negotiate down.

Squeezed: the leveraged middle — Oracle ($108B debt, one huge negative-margin customer), CoreWeave ($640M/quarter interest on depreciating collateral), patron-less labs bidding against sovereign wealth for turbine slots — and ratepayers, whose backlash is itself a supply-chain risk.

Ranked by bite: 1) power equipment and interconnection (binding now–2029). 2) labor and local opposition (compounding). 3) HBM and packaging (tight, widening). 4) capital (abundant, fragile if revenue stumbles). 5) raw materials (price risk only — except Western uranium enrichment). Above it all: Taiwan, the one input failure that resets every other answer.

The S-curve is half-here: single-site power runs out around 2027–28 and the pivot to density is underway. But efficiency alone delivers a free 10× every ~2.2 years, and the 2029–31 capacity wave is already ordered — so expect a shoulder, not a ceiling. The bottleneck decides who builds the next frontier model, not whether you can afford the last one.

Every number traces to a public source. Where companies keep secrets (training costs, current FLOP), we use Epoch AI's estimation methodology; projections are labeled. Costs are full training-run estimates; lab revenues are annualized run-rates. Sankey widths are approximate shares.