Working draft
Running…
The model runs on a calendar with four periods:
Preset: Defaults.
Settings are grouped like the diagram. Click a box title to find it in the diagram, where the same settings sit in its details panel. A dot marks a setting changed from the loaded inputs; ↺ puts it back. Greyed settings have no effect with the current choices.
| Area | AI 2040 | Proposed |
|---|---|---|
| Timeline | One 1-year window ending at the 2029.0 deal | A baseline period (business as usual), a diversion period, the agreement's entry into force and an audit period, with lengths you set, starting Q4 2026 |
| Inputs | Two world-stock snapshots and fixed shares from an older data sheet | Quarterly trajectories from the Compute supplement page and Epoch; AI 2040's snapshot is a data option |
| Smuggling | 2% of 2028 world sales, with better enforcement built in | Status-quo share × enforcement multiplier × diversion surge, capped by network capacity |
| Enforcement | Not a variable | Mechanisms with start dates acting on three levers: flow volume, recorded custody, verifier uncertainty |
| Traceability | A fixed 25% of smuggled chips untraceable | Recorded share per arrival cohort, set by enforcement |
| Earlier smuggled chips | Divertible only through false decommissioning, because installed chips "cannot be removed without explanation" | New route R4: unrecorded chips gathered and left off the declaration, plus seized chips held back |
| Chips abroad | No route, though the inputs include ≈6.15M H100e of offshore and leased compute by end-2028 | Declaration by ownership and control. New route R5 (fronts, buy-outs of leased chips). Capacity leased from non-PRC owners is shown separately as known. Host-country cooperation during the audit period |
| Inventory | Not modeled | New route R6: finished chips undeployed or in transit at the agreement's entry into force |
| Declaration threshold | Not modeled | Optional setting. Chips below it are legal, undeclared and never searched; they count toward the share of China's compute outside verification, not toward the dark headline |
| Detection | Built into fixed feasibility odds and the untraceable share | Audit period: access level, a top-down check against the verifier's uncertainty σ (flagged chips become known-missing and are searched harder), per-route survival |
| Feasibility odds | Yes/no draws in the routes: 60%, 25%, 100% | Moved to the audit period as the chance each coordinated scheme is exposed, set by access level. Applied once, so detection is not counted twice |
| Output | Median share of world compute at 2029.0, mixing unknown and known-missing chips | H100e at the end of the audit period, split unknown vs known-missing, plus the share of China's compute outside verification; means and medians |
| Appendix route | Retroactive understatement of domestic production, excluded | Folded into R2 for all domestic output before diversion |
| Baselines | Not needed | Status-quo traceability (smuggled ≈75% recorded, from AI 2040's 25% untraceable) and verifier uncertainty (±10%), which enforcement mechanisms improve |
Sensitivity
How much each input moves dark compute, starting from the inputs currently set in the simulator. Two views: each input changed on its own between a low and a high setting, and sweeps across the main continuous inputs. Every run uses the same random numbers, so differences come from the inputs, not sampling noise.
Each bar changes a single input to a low and a high value, while every other input stays at its current simulator setting. The vertical line is the current result, so a longer bar means the result depends more on that input. Hover an input's name for what it means.