Tiered Memory Architecture for Long-Context LLM Inference
A tiered memory hierarchy solves long-context LLM serving bottlenecks.
Celeste Oforiwaa
Staff Writer, Procurement & Strategy
Celeste Oforiwaa
Celeste cut her teeth covering enterprise software licensing disputes for a trade publication before pivoting to the broader question of how engineering organizations decide what to own versus what to rent. She brings a policy and contracting background that distinguishes her coverage of vendor negotiations and long-horizon infrastructure bets.
4 stories
Paged Attention and Memory Efficiency in Production LLM Serving
How PagedAttention solved the memory crisis that made LLM serving uneconomical.
Celeste Oforiwaa
Checkpoint Storage Cost During LLM Training Runs
Optimizer state, not model weights, drives checkpoint costs during training.
Celeste Oforiwaa
CapEx vs. OpEx Models for AI Infrastructure at Startups vs. Enterprises
Cloud pricing absorbs GPU obsolescence risk that CapEx buyers must face alone.
Celeste Oforiwaa