A - Substrate Indicators#
Contents
- §1 A1 - Frontier training run cost
- §2 A2 - Revenue against capex
- §3 A3 - The financing mix
- §4 A4 - Interconnection and power
- §5 A5 - Accelerator efficiency and depreciation
- §6 A6 - Leading-edge capacity outside Taiwan
- §7 A7 - Real rates and AI capex cost of capital
- §8 Reading the family as a system
- §8.1 A6 and the gray zone
Baselines as of mid-2026. Cadence: quarterly.
These test whether the ceilings in Part I bind on the schedule the document assumes. All four are relatively hard to fake, because each is downstream of a capital commitment or a physical object.
A1 - Frontier training run cost#
| Baseline | ~$1B largest frontier run (2025) - third-party estimate (Epoch AI); no run this size is publicly disclosed, and the r25 Ground pass could not close that gap |
| Trajectory assumed | ~$10B by 2027–28, ~$100B by 2029–30 |
| Trigger - slower | No run above $5B by end-2028 |
| Trigger - faster | A credible >$20B single run before 2028 |
| Revises | Compute; the whole capex-wall argument |
The slow trigger is the more informative one. If frontier spend plateaus while capability keeps improving, test-time compute and algorithmic efficiency substituted for scale - which retires the capital ceiling as a binding constraint and moves the entire analysis toward the optimistic case without any correction occurring.
A2 - Revenue against capex#
| Baseline | ~$60B AI revenue run-rate (2025); model-layer leaders alone ~$55B annualized by spring 2026 (Anthropic ~$30B Apr-2026, OpenAI ~$25B Feb-2026, company figures via press); big-four hyperscaler capex guided ~$700–725B for 2026, up ~77% y/y (company guidance, aggregated mid-2026) |
| Trajectory assumed | 45–55% CAGR to $400–700B/yr by 2030 |
| Trigger - correction | Two consecutive years below ~40% CAGR |
| Trigger - vindication | Revenue above ~$200B/yr by 2028 |
| Revises | Compute, Capital, Part V row 6 |
Read the composition, not just the total. Seat-priced assistant revenue and outcome-priced agent revenue have different ceilings; the second is what the $400–700B case requires, and it is the one that would show up as unusually high revenue per customer rather than unusually many customers.
A3 - The financing mix#
| Baseline | Migrating from operating cash flow toward private credit, SPVs, and vendor financing |
| Trigger - fragility | Majority of new AI infrastructure funded off-balance-sheet |
| Trigger - discipline | A large operator publicly cutting capex guidance on ROI grounds |
| Revises | Capital - determines whether a downturn is a pause or a credit event |
This is the least-watched indicator with the highest information content. A spending slowdown funded by cash flow is a decision; the same slowdown funded by debt is an accident. The second propagates; the first does not.
A4 - Interconnection and power#
| Baseline | 5–7 yr queues; ~523GW CAISO active (LBNL Queued Up 2026 edition, Jun-2026); ~380GW MISO; 3–4 yrs ERCOT for >75MW |
| Secondary | Wholesale prices at datacenter-adjacent nodes up ~267%; 30–50% of 2026 capacity slipping to 2028 |
| Trigger - loosening | Any major market halving typical wait, or federal transmission-siting preemption enacted |
| Trigger - tightening | State-level large-load moratoria in two or more of TX, VA, GA, OH |
| Revises | Energy, Uncertainty 2, Geopolitics |
Track the behind-the-meter share alongside the queue. If operators route around the grid faster than the grid reforms, the queue stops being the binding constraint and the political fight relocates from siting to ratepayer allocation. Those are different fights with different timelines.
The queue's composition moved in 2025 in exactly the direction the datacenter story predicts: total active US queue volume fell ~10% y/y while active natural-gas capacity rose ~86% to ~253GW and solar/wind/storage all shrank (LBNL Queued Up 2026 edition, Jun-2026). A shrinking queue is not loosening if the shrinkage is withdrawals; read it with the gas share, which is the firm-power signature of large-load demand.
A5 - Accelerator efficiency and depreciation#
| Baseline | ~1.35×/yr performance per dollar; 3–6 yr assumed useful life |
| Trigger - bullish | Operators extending depreciation schedules past six years |
| Trigger - bearish | Write-downs of prior-generation inventory |
| Revises | Capital - capital intensity of the whole build |
Depreciation-schedule changes are disclosed in filings, arrive before the underlying reality is common knowledge, and are one of the few places where an operator has to state what it actually believes about hardware longevity. The caveat: schedules also move for earnings-management reasons, so a lengthening that coincides with margin pressure elsewhere in the filing is a weaker signal than one that does not.
The indicator has started firing, in both directions at once: Amazon shortened a subset of server and networking life from six to five years effective Jan-2025 (10-K, Feb-2025, ~$0.7B operating-income impact), while Meta extended most of its fleet to ~5.5 years in the same period. Divergence between operators is itself a reading - the uniform six-year consensus of 2023–24 is gone, and whichever schedule the next filing cycle converges toward is the honest one.
A6 - Leading-edge capacity outside Taiwan#
| Baseline | Vast majority of leading-edge logic wafers still Taiwan-concentrated; announced multi-region fabs years from volume at node |
| Trigger - assumption softens | Non-TW share of advanced wafers crosses a material threshold (double-digit % of leading-edge) with yields, not ribbon-cuttings |
| Trigger - assumption fails | Blockade, kinetic conflict, or multi-quarter interruption of TW leading-edge output |
| Revises | Uncertainty 4, Bipolar, every compute-gated timeline |
This is the geographic assumption the corpus rests on, stated at ~90% continuity through 2032. A6 does not predict geopolitics; it tracks whether the consequence of disruption is shrinking (substitution) and whether disruption has occurred (binary).
A7 - Real rates and AI capex cost of capital#
| Baseline | Real rates elevated vs 2010s; AI build financed under mid-2020s WACC assumptions |
| Trigger - financial governor binds | Sustained rise in real yields coinciding with AI-related return narratives, raising hurdle rates on multi-year campus builds |
| Trigger - demographics win | Real rates stay low or fall despite visible AI deployment and productivity chatter |
| Revises | Rates, Capital, Uncertainty 1 financial governor |
The document's claim: success can raise the neutral rate and make the next round of infrastructure more expensive to fund - a negative feedback on takeoff that is not technical. A7 is how you see it in markets rather than in papers.
Reading the family as a system#
Single indicators mislead in ways pairs do not. A1-slow plus A2-vindication is not a mixed reading - it is the specific signature of efficiency substituting for scale, and it retires the capex wall rather than confirming half of it. A2-correction plus A3-fragility is the credit-event path; A2-correction plus A3-discipline is the orderly-pause path - the same demand fact with opposite macro consequences, which is why A3 must always be read alongside A2 rather than after it. A4-tightening plus A7-rising is the double governor, physical and financial constraints binding together, and it is the strongest confirmation the constrained case can receive from this family.
Also worth naming plainly: most of these series are produced by parties with a stake in the reading. Disclosed training-run costs are strategic communications; revenue definitions can be widened to fold bundled products into "AI revenue"; financing structures are designed partly to keep commitments off the statements A3 reads. The defense is triangulation against physical proxies that are expensive to fake - power contracted, wafers ordered, steel in the ground - and a standing discount on any number whose publisher benefits from its size. Where a self-reported figure and a physical proxy disagree, the proxy wins.
A6 and the gray zone#
A6 (non-Taiwan advanced wafer share) retires consequence of Taiwan disruption as capacity diversifies. It does not retire probability - and the gray-zone table on Uncertainty 4 can fire while A6 is still low (repricing without substitution). Read A6 for shield decay / consequence shrink; read freight insurance and license denials for soft failure of the consumed assumption. Do not wait for A6 to move before scoring gray-zone stress.
Related: Part I - Substrate · Capital · Energy · Part IX - Macro