Home Today’s Briefing Deep Dives In Play Trending For You Discover My Stack Saved Off Radar Search

Loading today’s stories

Today

AI-assisted reporting, grounded in a per-article claim ledger. Perspective Coverage and Reality Strip separate source viewpoints from evidence, adoption, hype, incentives, and confidence.

Today

Front Page
Today’s Briefing
Newsletters
Trending
For You
Discover
My Stack
Saved
Off Radar
Search

Sections

Security
Build
Invest
Product
Leadership
Science

Company

About
Mission
How this is made
Evidence Index
The desks
The sources we read
Corrections
Press and brand
The daily edition

Trending now

Software Supply Chain Security
Frontier Model Releases
Agentic AI Security
Agent Context Management
Vulnerability Disclosure
Antitrust remedies
End-to-End Driving Models
Detection Engineering
Prompt injection
Publisher Monetization

How this is made

Today For You My Stack Off Radar Search

Topic

GPU cold-start latency

The startup tax between a pending inference pod and its first served token, and where the seconds go.

Current clusters

build1 distinct publisher

Model size picks which of six GPU cold-start bottlenecks you pay for

An instrumented run from pod creation to first response found eight minutes spread over six phases, with kernel recompilation eating a 64 GB model's startup and an S3 download pattern eating a 203 GB model's.

Publishers:thenewstack.io

Reality

Evidence56
Adoption: Insufficient
Hype gap+28
Incentives76

Confidence

GPU cold-start latency — Clarity Today