Accelerator utilization rates below 70% are not a GPU problem — they are a storage problem. Google Cloud’s Hyperdisk ML and Rapid Storage are built specifically to fix that, and the numbers are not subtle.
TurboQuant: Kind of a Big Deal.
Google Research just published a way to cut AI serving costs by 50% with zero accuracy loss. The interesting part is what happens to the ISVs who figure this out first.
Google Built a TPU for the Age of Inference. Meet Ironwood.
TPU Ironwood is Google’s 7th-generation custom AI chip, and unlike its predecessors, it was built for inference first. Here’s what that means and why it matters.
LLM Traffic Is Weird. Your Infrastructure Needs to Know That.
Standard load balancers treat LLM inference like any other HTTP traffic. That is expensive and slow. GKE Inference Gateway knows the difference.