Particle.news
Download on the App Store

AI Market Shifts From Training to Inference, Lifting ASIC Makers and Rattling Big Chip Winners

The move toward always-on model serving favors custom ASICs like Broadcom's chips and is driving investors to reposition into cloud owners such as Amazon.

Overview

  • Industry coverage shows the AI market is moving from training large models to continuous inference, a change that makes cost‑efficient, task‑specific chips more attractive than general‑purpose GPUs for large-scale deployments.
  • Broadcom reported its AI chip sales rose about 65% to $20 billion in fiscal 2025 and told reporters it expects far larger AI revenue as customers scale inference workloads.
  • Nvidia remains a dominant supplier for training with very strong recent results, including a surge in data‑center sales, but its pricing power could weaken as customers seek cheaper inference alternatives.
  • Quarterly regulatory filings reveal many billionaire fund managers have pared stakes in Nvidia and Alphabet and concentrated holdings in Amazon as their preferred public AI exposure.
  • The change could push cloud providers to buy or develop more custom silicon to cut inference costs, a shift that would lower running costs for AI services and alter where AI spending flows in data centers.