79% of organizations report GPUs idle half the time, fueling an AI compute bubble
机器之心 · wechat · 2026-11-22
A Machine Heart PRO analysis examines the "hoarding paradox": companies panic-buy GPUs while leaving them idle at least half the time.
- VentureBeat Intelligence surveyed 218 AI-powered organizations (Jul–Sep 2026): 79% self-reported GPUs idle ≥50% of the time, yet 70% of those still plan to buy more or switch platforms within a year.
- IDC puts global AI infrastructure spend at $89.7B in Q1 2026, up 33% YoY. The hoarding mindset traces back to 8–11 month H100 delivery delays in 2023–24.
- Even running GPUs underperform: CMU/UIUC research found inference servers idle in an "execution idle" state (110W vs 35W deep sleep) that can consume 7–65% of serving energy. Unlocking RTX 4090's half-speed FP16 instructions roughly doubled single-op throughput but only improved matrix multiplication by up to 34%.
The core question: should organizations fully utilize existing GPUs before buying more?
More from Venture
- Indie dev praises Wu Lujia's book on small-product entrepreneurship after marathon read — yihui_indie · 2026-11-22
- Founder: Most enterprise sales fail on communication, not the product — gabriel1 · 2026-11-22
- 1998 Red Herring piece on VC bubble resurfaces as AI-era investors say 'history rhymes' — MartinGTobias · 2026-11-22
- Howard Marks: AI's pace of progress makes OpenAI and Anthropic hard to value — WhatTheLJW · 2026-11-22
- B300 GPUs clearing at $18/hour and sold out as a compute credit crunch looms — AccBalanced · 2026-11-22
- AI stock selloff is a leverage squeeze: neoclouds down ~63%, CoreWeave CDS spikes — tengyanAI · 2026-11-22