The Next Fifteen Years

A forecast built from first principles
Section future / 07-indicators / README.md

Part VII - The Indicator Dashboard#


Contents

Baseline readings as of mid-2026.

A forecast that cannot be checked before its horizon is not a forecast. This part converts the claims in Parts I–VI into observables with thresholds - things that can be read off the world on a quarterly cadence, each attached to the specific claim it would confirm or break.

The discipline: every indicator here names a trigger value, and every trigger names what to revise. An indicator that moves without changing anything you believe was decoration.

The three families#

FamilyAnswersCadence
ASubstrateAre the physical and financial ceilings binding on schedule?Quarterly
BDiffusionIs the technology actually reaching into the economy?Quarterly to annual
CGovernanceWhich regulatory architecture are we going to get, and when?Event-driven

The families are not independent, and reading them jointly is most of the skill. A binding A-family constraint shows up later as a B-family slowdown, so a B stall with A slack points at institutions, while a B stall with A tight points at physics and finance - opposite revisions from the same B reading. And a C-family event reprices both: the post-incident window can relax A constraints (permitting waivers) while tightening B (deployment moratoria) simultaneously.

The five that matter most#

If only five things can be tracked, track these. Each is the leading edge of a different claim the corpus depends on.

IndicatorBaseline (mid-2026)Watch forBreaks
Entry-level : senior posting ratio, knowledge professionsDepressed; entry-level ~7% of large-tech hiresFailure to recover when aggregate white-collar hiring doesGame 4 - settles the substitution-vs-cyclical split
Internal research cycle time per validated experiment at frontier labsNot disclosed; infer from release cadence and paper throughputSustained >2× compression and physical/financial governors not bindingUncertainty 1 - the RSI loop closing
AI revenue run-rate vs. capex~$60B (2025) revenue vs. ~$700–725B 2026 hyperscaler capex guideRevenue CAGR falling below ~40% for two consecutive yearsCompute and Capital - the correction
Teleoperation-to-autonomy ratio in deployed robotsOverwhelmingly teleoperated or scriptedAutonomy share crossing 50% on unstructured tasksRobotics - the 2030s swing variable
Grid interconnection queue length5–7 yrs; ~410GW CAISO, ~380GW MISOAny market halving its typical waitEnergy, Uncertainty 2

Why these five lead#

Each was chosen for when it moves, not only for what it measures. The posting ratio moves at the hiring decision, quarters before employment stocks show anything. Cycle time moves inside labs before any release demonstrates it. The revenue-to-capex gap moves in guidance and filings a year or more before a write-down forces the same fact into the open. The teleoperation ratio moves at deployment, years before robotics revenue is large enough to surface in sector statistics. And queue length moves at the application stage, roughly the length of the queue itself ahead of power actually binding. A dashboard of lagging indicators confirms; this one is built to warn. The cost accepted in exchange is a higher false-alarm rate - early-moving series are noisier than the stocks they anticipate, which is why each threshold requires persistence (two quarters, two consecutive years) rather than a single print.

Also on the short list (rounds 5–8)#

IndicatorFamilyBreaks
AI liability as published insurance line + reinsurer capacityC7Uncertainty 6; deployment frontier
Leading-edge wafer share outside TaiwanA6Uncertainty 4
Cognitive vs physical services inflation gapB7Prices two-economy claim
Capability gap: verifiable vs unverifiable domainsB8Uncertainty 5
Labor share of tax revenueB10Fiscal squeeze
Companion use × social contact; junior identityB11Meaning welfare channel
Capture-time provenance at scaleC8Game 5 / media enclosure vs public verification
Warehouse / freight / specialty harvest autonomyB12Robotics physical fork
Taiwan gray-zone (freight insurance, license denials, delivery slip)U4Soft fail of fab continuity without crisis headline

How to read a dashboard without fooling yourself#

Four failure modes, all of which this document is exposed to:

  1. Confirmation drift. An indicator chosen after the fact to support a held view tests nothing. Every threshold here was set before the reading was taken; when one is revised, the revision goes in the log, not silently into the table.
  2. Measurement capture. Once an indicator becomes salient, the thing being measured starts being managed. Benchmark scores went this way already. Prefer indicators that are expensive to fake - capital flows, physical deployments, queue positions - over self-reported ones.
  3. The perception gap. The METR result in Game 4 means self-reported productivity is unreliable in a known direction. Weight instrumented and revealed-preference measures far above surveys.
  4. Absence of movement is information. The most common real outcome is that an indicator sits still for three years. That is a result, and it usually favors the institutional-friction case over both the acceleration and the collapse case.

Scoring cadence#

View markdown source

select · Enter open · Esc close