What 20 kW racks change about cooling design
Air has limits. How containment, in-row cooling and floor layout shift when rack density quadruples.
July 2026 · 9 min read Read more →Blog
Notes from the data hall
Engineering notes, product deep-dives and tutorials from the people who run the facility.
Featured · Engineering
Topology-aware scheduling: where 20% NCCL gains actually come from
Everyone quotes the headline number; few explain the mechanism. We walk through how fabric topology files reach Slurm, what "closest nodes" means on a fat-tree, and when the gains disappear — with measurements from our own pods.
Read the article
Engineering
InfiniBand vs Ethernet for LLM training: what we measured
All-reduce latency, tail behavior and cost per GPU-hour on both fabrics — numbers from production pods.
June 2026 · 12 min read Read more →Checkpointing strategies for week-long training runs
How often to checkpoint, where to put them, and how to resume in minutes when a node dies at 3 AM.
May 2026 · 8 min read Read more →Tutorials
Your first training job on Infraeon: a 15-minute quickstart
From credentials to a running Slurm job — images, storage mounts and the three mistakes everyone makes first.
July 2026 · 10 min read Read more →RAG on Token Factory: from production logs to grounded answers
Embedding endpoints, managed vector storage and a working retrieval loop in an afternoon.
June 2026 · 14 min read Read more →Serving a fine-tuned model on a dedicated endpoint
Package your checkpoint, pick a GPU, set rate limits — and keep the per-token bill predictable.
May 2026 · 11 min read Read more →Product & company
How per-token pricing actually works on Token Factory
Input vs output tokens, shared vs dedicated tiers, and why most teams overestimate their serving costs.
July 2026 · 6 min read Read more →Why we built Infraeon in Bangkok
Latency maps, cable landings, energy and talent — the four reasons this city wins for AI infrastructure in SEA.
April 2026 · 7 min read Read more →Start your journey today
Tell us about your workload — an engineer, not a sales bot, will get back to you within one business day.