Topic
AI Infrastructure
GPU infrastructure inherited the Kubernetes control plane without inheriting any of the assumptions that make it work. Accelerators are indivisible in ways CPUs are not, ten to fifty times more expensive per unit of idle time, supply constrained on a timescale of quarters, and attached to workloads whose runtime characteristics differ so sharply between training and serving that treating them as one platform problem guarantees you will do both badly.
What gets covered here: fractional GPU strategies and their honest limits — time-slicing, MPS, MIG, and when each is the wrong answer. Multi-tenant fairness when a single job can hold a $30,000 device hostage for six days. The architectural split between training clusters, which are throughput machines tolerant of preemption, and inference fleets, which are latency machines that are not. Plus the market layer around all of this, including how AI-era naming and premium domain valuation actually behave once you look at comparable sales rather than registrar marketing.
Articles in AI Infrastructure
-
The .ai Domain Market Priced In a Decade in Eighteen Months
How .ai domain pricing actually behaves: registration volume, comparable sales data, the two-year renewal structure, and a framework for valuing a premium technical domain.
-
Training Clusters and Inference Fleets Are Not the Same System
The architectural split between training and inference infrastructure: interconnect requirements, scheduling models, checkpointing, autoscaling behaviour, and the cost structures that follow from each.
-
GPU Sharing in Multi-Tenant Clusters: Four Mechanisms, Three of Them Traps
A technical comparison of GPU sharing on Kubernetes — time-slicing, MPS, MIG, and dedicated allocation — with the fairness, isolation, and utilisation tradeoffs of each.