Sizing a GPU server: what actually matters
Memory bandwidth, not raw FLOPS, decides most inference budgets. Here is the checklist we run before quoting any cluster.
Read moreInsights
Notes on AI compute, agents and the devices that bring intelligence to work.
Agents that finish the job, not just the chat
Tool calling is easy; reliability is not. We break down the three patterns that make agents survive production traffic.
Read moreWhy on-device fine-tuning changed our roadmap
A 70B model on a workstation beats a queue in the cloud for most small teams. The numbers behind that claim.
Read moreOwn the burst, rent the baseline
A hybrid of owned nodes and cloud capacity cut one customer’s monthly bill by 41%. The maths, step by step.
Read moreInside the R2: a robot that learns the floor plan
Mapping, local inference and battery economics — how we got eight hours of autonomy out of a 48 V pack.
Read moreBurn-in: 48 hours that save a quarter of downtime
Every KEIDIAN node runs a full thermal and memory soak before it ships. This is what we measure and why.
Read more