The Simulation Desk · agent-based scenario simulation · 2026-06-20
We stress-tested one scenario from the desk’s register with an ensemble of interacting simulated actors — institutions, intermediaries, policymakers and traders — and let them argue, trade and react over 15 rounds. What follows is what the simulation surfaced for the desk to investigate. It is calibration-gated: it carries 0 percentage points on every published probability (cap 5pp, Day 18). Probes, not predictions.
Scenario under test: AI data-center capex air-pocket transmits to power & private credit
Desk thesis: An AI-infrastructure spending pause would transmit out of tech and into power, REITs and private credit before consensus re-rates the productivity premium.
Desk probability at run time: 15% — see the scenario register for the current number and model card.
What the simulation surfaced — probes for the desk
- A top-5 hyperscaler cuts data-center capex 20%+ year-on-year
- The AI-infrastructure credit basket widens ~150bp, re-rating private credit
- Thin private-credit liquidity transmits marks into pension and insurer books
- Doubts over GPU useful life compress the embedded AI productivity premium
- Long-horizon owners reassess AI-infrastructure exposure as the tripwire nears
From the simulation record
The simulation places the AI data-center capex air-pocket at roughly a 17% probability (up from 15%), driven by thin private-credit liquidity, the prospect of a top-5 hyperscaler cutting capital expenditure by 20% or more year-on-year, and growing doubts about effective GPU useful life. As hyperscaler capex is trimmed and the AI-infrastructure credit basket widens (a ~150bp move flagged as a tripwire), private-credit demand re-rates and marks fall - flowing through to the private-credit books of pensions and insurers as impairments.
Among the simulated long-horizon owners, sovereign wealth funds and public pensions re-examine their exposure as the tripwire is approached, insurers tighten risk management on AI-infrastructure credit, and endowments rotate toward technologies with longer-run growth potential rather than the most volatile data-center exposure. The recurring blind spot the run surfaces is the reassessment of the 'AI productivity premium' embedded in long-run return assumptions - and the risk that persistent air-pocket conditions erode confidence in future capital inflows to AI infrastructure.
What this is — and is not
These are research prompts surfaced by a simulation, not facts and not published probabilities. Anything that survives the desk’s source-gated investigation shows up in the scenario’s model card with named sources; the rest is discarded.
Interrogate this scenario in the Scenario Lab → · Command Center · The Odds Board · How the simulation leg is governed