Agentic Racing

Theory paper; calibration only; no calendar stop date

Race Calibration Dashboard

The dashboard is a macro-style calibration layer for the model, not an empirical estimate and not a forecast. It first shows how intense the AI race has been in recent public aggregate data, then translates the model into threshold sensitivities: how fast the remaining task-frontier-value-and-scaling term must grow for racing to remain privately justified.

The recent race calibrates the model's starting environment. It does not imply that racing should continue indefinitely. The forward object is a threshold in \(d_{\varepsilon t}=\log\mathcal R_{\varepsilon t}\), not a calendar year.

Primitive	Calibration number	Role in the model
Recent technical push	Epoch reports frontier training compute near \(5\times\)/year and pre-training efficiency near \(3\times\)/year	Raises inherited effective compute and can keep the race above the cutoff.
Recent rent pressure	Epoch reports frontier training-cost growth around \(3.5\times\)/year	Gives the denominator growth that persistent value must beat.
Infrastructure demand	NVIDIA data-center revenue rises sharply from fiscal 2024 to the fiscal 2027 run rate	Shows that the common-input margin is macro-scale.
Lead half-life	Use months of private lead persistence as a sensitivity primitive	Converts headline capability into monetizable private task-frontier option value.
Scaling curvature	Use fixed-protocol \(q'(C_t)\) and slowdown \(z_t\)	Determines how much capability the next effective-compute step buys.

1. Recent Race Intensity

Normalized public proxies for the recent AI race. These are calibration moments, not structural estimates.

Effective compute trend Training cost trend NVIDIA DC revenue

Sources: Epoch AI Trends; NVIDIA FY2024; NVIDIA FY2025; NVIDIA FY2026; NVIDIA Q1 FY2027.

2. Threshold Sensitivity

Required task-frontier-value-and-scaling growth after crediting \(3\times\) efficiency growth.

Base rent pressure Rent shock

Source: model identity \(G_A=G_R/(G_\ell G_e)\), with \(G_e=3\), base \(G_R=3.5\), and rent-shock \(G_R=7\). Lead half-life determines \(G_\ell=2^{-12/h}\).

Loading calibration moments...

Reading

The first graph says the recent race has been intense on observable aggregate margins. That is a calibration fact the theory should rationalize, not an estimate of welfare or a prediction that the race must continue.

The second graph gives the paper's sharper object. Shorter lead half-lives require much faster growth in task-frontier value and fixed-protocol scaling slope to offset rent growth. This is the threshold logic of the theory.

Source Ledger

Graph	How it works	Correct citation
1. Recent Race Intensity	Indexes public aggregate proxies to the start of the displayed window. Effective compute combines Epoch's frontier training-compute and efficiency trends. Training cost uses Epoch's frontier training-cost trend. NVIDIA data-center revenue is a supplier-demand proxy.	Epoch AI (2026), Trends in Artificial Intelligence; NVIDIA fiscal releases for FY2024, FY2025, FY2026, and Q1 FY2027.
2. Threshold Sensitivity	Computes the growth in remaining task-frontier value and scaling slope required to keep the race the same distance from the cutoff under alternative lead half-lives.	Derived from the paper's race-index identity and calibrated with Epoch's \(3.5\times\) training-cost trend and \(3\times\) efficiency trend.
Calibration status	The website graphs are not standard errors, confidence intervals, or structural estimates. They are reproducible calibration artifacts.	`verification/race_intensity_calibration.py`.