AI Lab · Methodology

How we model AI scale

Every scale figure in the XGame AI Lab derives from the single model on this page. Nothing here is a report of operational traffic — it is a published set of assumptions that anyone can recompute.

Token Equivalent

Normalized estimate of model input/output and context-processing workload.

A single unit for comparing heterogeneous processing — generation, retrieval, context assembly — across architectures. It is a modeling unit, not a billing meter: no vendor invoice corresponds to it.

Agent Capacity

Maximum modeled concurrent or persistent identities under a specified architecture scenario.

A ceiling declared by the architecture study, comparable to rated capacity in infrastructure engineering. It is not a count of agents currently running.

Simulation

Architecture scenarios used to evaluate scale and economics; values may not represent production activity.

Scenarios exist to compare deployment scales and stress-test economics. Where a page shows a scenario figure next to an operational one, the scenario figure is always labeled.

The model

Four parameters produce every derived figure:

decisions/mo   = A × D × 30
token-eq/mo    = decisions × K
agent-events/mo= decisions × E

Parameter	Symbol	Value (Global)	Notes
Agent capacity	A	2,400,000	Architecture ceiling; varies per scenario
Decisions / agent / day	D	40	Shared across scenarios
Token-eq / decision	K	4,000	Shared across scenarios
Sub-events / decision	E	13	Perception, memory reads/writes, planning, action
Days / month	—	30	Fixed

Deployment scenarios

Three scales of the same architecture. Site-wide displays default to the Global scenario and say so.

Metric	Prototype	Regional	Global
Agents	25K	400K	2.4M
Decisions / mo	30M	480M	2.88B
Agent events / mo	390M	6.2B	37.4B
Token-eq / mo	120B	1.92T	11.52T

Disclosure

Scenario outputs on this site are modeled values for research and communication. They are not measurements of production systems.