Spacebus
FIELD BRIEF № 003
SUBJECT AI INFRA · TRAINING ⟷ INFERENCE
STATUS --:--:-- UTC
INFRASTRUCTURE TELEMETRY // WHERE COMPUTE LIVES // 2026
The AI Infrastructure Divide:
Training vs. Inference
AI Training
The Power-Rich Remote Factory
AI Inference
The Latency-Critical Urban Edge
FIG. 01 — PHYSICAL TOPOLOGY / FLOW MAP
SCALE N.T.S.
ElectricityGrid power → training cluster
Water / coolantClosed-loop + evaporative
DataInference responses → users
01
Two workloads, two geographies
Remote / Rural areas
Primary Location
Urban / Metro-adjacent
Power abundance & scale
Primary Driver
Proximity to users
Not latency-sensitive · weeks
Latency Sensitivity
Highly latency-sensitive · ms
02
What each side optimizes for
TrainingPower-rich remote factory
A1
Centralized in power-rich rural hubs
Training favors remote locations like Indiana or Wyoming where abundant, cheap power is available.
A2
High throughput, low latency sensitivity
Training processes massive datasets over weeks; millisecond delays do not impact model quality.
Timescale · weeksA3
Scale-first infrastructure
Prioritizes multi-megawatt capacity and high-density liquid cooling over proximity to end users.
Multi-MW · liquid-cooledInferenceLatency-critical urban edge
B1
Distributed in metro-adjacent clusters
Infrastructure sits in and around urban hubs like Silicon Valley to stay physically close to users.
B2
Latency is a factor of distance
Physical distance creates network hops and lag, which degrades real-time user experiences.
Distance = hops + lagB3
Real-time responsiveness required
Applications like autonomous vehicles and chatbots require millisecond response times to be effective.
Response · milliseconds
AI Infrastructure Field Brief — Training vs. Inference, 2026. Illustrative topology; not to scale.