Every part that must work for one electron to become one AI token
Eighteen steps from the fuels upstream of the power plant to the token arriving at a human — or, increasingly, at another machine. Each step documents what is hardest today (2026) and what becomes hardest as AI scales toward 2031–2036, with disagreements between credible sources presented rather than averaged away.
Only a fraction of primary energy ends up switching transistors. Generation losses, ~5% transmission, conversion overhead (historically 10–50%, now <10% in the best facilities), cooling, idle silicon (MFU <50% — the chain's single largest waste), and computation spent on anything but the useful token. Improving any step's efficiency is equivalent to building power plants.
Model architectures change in months; chips in ~2 years; datacenters in 2–4; grid connections in 5–10; power plants and transmission in up to 15; mines in 10–20. AI demand grows on the fastest clock while its foundations move on the slowest. Nearly every bottleneck in this document is, at root, this mismatch — and nearly every durable business lives in the gearbox between the clocks.
Measured, not guessed (Step 16): Google's fleet median ≈ 0.24 Wh/prompt; an efficient 120B model ≈ 0.11–0.20 Wh/query; a reasoning marathon ≈ 15–16 Wh — a 100× spread dominated by one variable: how many reasoning tokens the workload burns. The metric is a distribution, not a scalar.
Demand growth — is the step's market outrunning incumbents? · Supply-base maturity — young and fragmented, or entrenched? · Capital intensity — stated as a fact to finance, never a feasibility verdict · Timing fit — will the problem exist when a product ships in 2–4 years? · Defensibility & buyer power — what stops copying; do the natural customers in-house this layer? · Buyer structure — five competing qualifiers beat one monopolist. No step is ruled out in advance (Crusoe, Cerebras and Commonwealth Fusion all frontally attacked "impossible" layers); severity of a bottleneck and quality of an opportunity are different axes.
The chain, step by step
Full deep-dive content per step: what happens there, the verified numbers, where credible sources disagree, what is hardest today vs. 2031–2036, and the opportunity scan. Search filters titles and full text.
The consolidation — ~50 opportunities, 14 clusters, 3 tiers
Two rankings side by side. Market merit scores purely on the six lens dimensions. Team-weighted adds fit as a seventh dimension, on a thin and explicitly stated basis: a team of very talented hardware engineers (disciplines unspecified), presumed Netherlands/EU. Toggle to see Tier 1 reorder.
Scoreboard & prediction ledger
Severity of each step as a constraint on total token production — revised after all deep dives — plus the eight dated, resolvable disagreements that will sharpen this document as reality reports back.
Severity shown on a 1–5 scale (sequential ramp: darker = more binding). Clock-speed bar = how slowly the step's assets change (longer bar = slower clock, log scale from months to 20 years). Original first-pass judgments held up well; upgrades landed on construction labor (§4.2), memory scarcity (§9.1), packaging's long-run role (§10.4) and inference economics (§16.2). Caveat: severity assumes demand holds — ledger rows 1 and 6 are the load-bearing assumptions. Pattern: today's worst bottlenecks sit at the two ends of the clock-speed spectrum — the grid (slowest) and packaging/HBM (most supply-constrained) — while 2031–2036 bottlenecks concentrate in energy, the package, memory and inference economics.
Synthesis & context
The original question — what gets hardest as AI scales over 5–10 years — answered once, as a sequence of three phases; plus the cross-cutting flows, the outside forces, the waste side, and the document's own history.
The constraint is time
Everything the chain needs exists but is queued: turbines sold out to 2030, interconnection to 2031+, 2nm and CoWoS allocated, 2027 memory pre-sold. The premium accrues to whatever shortens calendars — behind-the-meter power, prefab construction, flexible interconnection, faster qualification.
Scarcity flips to selection
Capacity ordered in 2024–26 lands (24 GW/yr turbines, tripled transformer output, 130K CoWoS wafers/month, doubled HBM) just as training-compute growth decelerates from 10× to 3–4×/yr. If demand skeptics are even half right, several links whipsaw from shortage to glut. Winners hold reliability, test, instrumentation and integration positions — layers that earn in both directions.
Physical and political at once
The package becomes the new Moore's Law — power, cooling, memory and optics fused into one thermally-impossible object, exactly where the Netherlands holds three nuclei. The memory wall is the deepest architectural constraint; inference the dominant, user-adjacent grid load; energy-per-token disclosure and allocation governance become regulated matters. The two great unhedged exposures persist: Taiwan, and a software breakthrough that voids the capacity arithmetic in a quarter.
Three of the disciplines that fuse into the 2030 AI package have their global centers of gravity within ~50 km of each other:
The chain's value migration (power §5.3, cooling §6.3, litho's own hedge §7.2, the HBM base die §9.2, optics §12) converges on the package — the most Dutch-flavored artifact in the whole chain. This extends the Europa 2031 stack-map story: the crown jewel argument now reaches one layer up from ASML.